# The Holbrook Report — full corpus

> Curiosity, taken seriously. Education, not advice. Claims are labeled [FACT] / [CHARACTERIZATION] / [PROJECTION] and every report ends with its sources. Each report is written by an AI model and directed and edited by Aaron Holbrook.

---

# Thresholds: Why Two Nearly Identical Crowds Do Opposite Things
*A walk through Granovetter's threshold model of collective behavior: what a threshold is, the arithmetic that turns a distribution of them into an outcome, the three mechanisms the paper's own author uses to dismantle its most famous result, and the fifty-year argument that followed about tipping points, complex contagion, and the 25 percent number.*

Source: https://theholbrookreport.com/reports/thresholds-and-collective-behavior/ · Published: 2026-08-03

<!--
AUTHOR NOTE GOES HERE (optional). Aaron writes this in his own first-person voice
if he wants a personal introduction. The AI does not write this block. Pattern:

<div class="author-note">
  ...your introduction...
  <span class="byline">- <a href="https://aaronholbrookmusic.com">Aaron Holbrook</a></span>
</div>
-->

<div class="callout"><span class="label">What this piece examines, and what it is:</span> The subject is a formal model published in 1978 that has since become the standard apparatus for reasoning about riots, fads, bank runs, strikes, adoption curves, and revolutions. This report explains the model at full size, works its arithmetic, sets out the conditions under which its author says it is false, and traces the fifty years of testing and revision that followed. It is an educational explainer, not advice about how to start or stop anything. The report is AI-authored and directed by hand; every claim is labeled [FACT] (a sourced point of record), [CHARACTERIZATION] (a defensible label placed on those facts), or [PROJECTION] (a conditional "if X then Y," never a prediction).</div>

Certain outcomes are routinely explained backward. A crowd riots, so the crowd must have been angry. A product spreads, so the product must have been good. A norm collapses, so opinion must have shifted. In each case the reasoning runs from the observed collective result to a claim about what the individuals in it wanted, and the strength of the result is taken as evidence of the strength of the wanting.

There is a class of situations where that inference is unreliable in a specific and demonstrable way: situations where each person's decision depends on how many other people have already decided. In those situations the outcome is governed by the *distribution* of individual tipping points across the group, and two groups whose distributions differ by a single member can arrive at opposite extremes. The average sentiment in the two groups is the same. The outcome is not.

The formal treatment is Mark Granovetter's "Threshold Models of Collective Behavior," published in the *American Journal of Sociology* in May 1978. [FACT] The sections below define the mechanism, work the arithmetic, and then spend as much space on where the model breaks as on where it holds, because the paper itself does.

The pills in this piece mark how load-bearing each point is to the model, not how dangerous anything is: <span class="pill high">Central mechanism</span> for the parts the structure stands on, <span class="pill med">Important context</span> for the supporting results, <span class="pill low">Refinement</span> for later corrections, and <span class="pill ok">Failure condition</span> for the cases where the model does not apply.

## The demonstration

One hundred people are milling around a square. Each has a <span class="term" data-def="The proportion of a group that must already have taken an action before a given individual will take it; the point where that individual's perceived benefits first exceed perceived costs.">threshold</span>: the proportion of the group that would have to join a riot before that person joins it. Suppose those thresholds are perfectly uniform. One person has threshold 0, one has threshold 1, one has 2, and so on up to one with threshold 99. [FACT]

The person with threshold 0, the instigator, breaks a window. That satisfies the person with threshold 1, who joins. Now two people are rioting, which satisfies the person with threshold 2. The chain runs to the top. Equilibrium: 100 rioters.

Now alter exactly one person. Remove the individual with threshold 1 and replace them with a second individual with threshold 2. Every summary measure of the two crowds is essentially the same: same size, same mean threshold to within one unit, same range, same general disposition. The instigator breaks the window. Nobody has threshold 1. The cascade stops before it starts. Equilibrium: 1 rioter. [FACT]

<div class="finding">
  <div class="finding-head">
    <div class="name">1 &nbsp; The aggregation process, not the preferences, produced the difference</div>
    <span class="pill high">Central mechanism</span>
  </div>
  <p>Granovetter's stated point is what an observer would conclude about each crowd. Of the first: "A crowd of radicals engaged in riotous behavior." Of the second: "A demented troublemaker broke a window while a group of solid citizens looked on." [FACT] Both descriptions are wrong about the people, and both are the natural reading of the events. The difference in outcome, in his words, "results only from the process of aggregation, and in particular from the gap in the frequency distribution in the second case." [FACT] [CHARACTERIZATION: this is a constructed example, not an empirical finding, and its function is to establish that the inference from outcome to disposition can fail completely rather than to claim it usually does.]</p>
  <div class="callout"><span class="label">The operational consequence:</span> where behavior is interdependent, an aggregate outcome is not evidence about average sentiment. It is evidence about the joint distribution of sentiment and the sequence in which people were exposed to each other.</div>
</div>

```chart
{"type":"line","title":"Two crowds of 100, differing by one person","xLabel":"Time step","yLabel":"People rioting","xMax":101,"yMax":100,"series":[{"name":"Uniform thresholds 0-99","points":[[0,0],[1,1],[2,2],[3,3],[4,4],[5,5],[6,6],[7,7],[8,8],[9,9],[10,10],[11,11],[12,12],[13,13],[14,14],[15,15],[16,16],[17,17],[18,18],[19,19],[20,20],[21,21],[22,22],[23,23],[24,24],[25,25],[26,26],[27,27],[28,28],[29,29],[30,30],[31,31],[32,32],[33,33],[34,34],[35,35],[36,36],[37,37],[38,38],[39,39],[40,40],[41,41],[42,42],[43,43],[44,44],[45,45],[46,46],[47,47],[48,48],[49,49],[50,50],[51,51],[52,52],[53,53],[54,54],[55,55],[56,56],[57,57],[58,58],[59,59],[60,60],[61,61],[62,62],[63,63],[64,64],[65,65],[66,66],[67,67],[68,68],[69,69],[70,70],[71,71],[72,72],[73,73],[74,74],[75,75],[76,76],[77,77],[78,78],[79,79],[80,80],[81,81],[82,82],[83,83],[84,84],[85,85],[86,86],[87,87],[88,88],[89,89],[90,90],[91,91],[92,92],[93,93],[94,94],[95,95],[96,96],[97,97],[98,98],[99,99],[100,100],[101,100]],"tone":"ink","label":{"x":64,"y":80,"text":"Uniform 0-99: all 100 riot","anchor":"end"}},{"name":"Threshold 1 replaced by a second threshold 2","points":[[0,0],[1,1],[2,1],[101,1]],"tone":"accent","dash":true,"label":{"x":20,"y":9,"text":"One person changed: 1 riots","anchor":"start"}}],"markers":[{"x":101,"y":100,"text":"100","anchor":"end"}],"caption":"Forward recursion r(t+1) = F[r(t)] for both crowds. The only difference is that one person with threshold 1 is replaced by a second person with threshold 2. Computed from the distributions, not traced from the paper.","table":[["Crowd","Equilibrium"],["Uniform thresholds 0-99","100"],["Threshold 1 replaced by a second threshold 2","1"]]}
```

## What a threshold is, and three things it is not

A threshold is the proportion of the group an individual must see act before acting, which is to say the point where perceived benefits first exceed perceived costs. [FACT] The concept is narrower than it looks, and most casual uses of the phrase "threshold model" get at least one of the following wrong.

<div class="kv">
  <div class="k">Not an ideology score</div>
  <div class="v">Two people with identical thresholds may be politically opposite. A threshold nets many costs and benefits together, which is why Granovetter invokes the phrase "strange bedfellows." A radical has a low riot threshold because arrest is cheap to them and rioting is valuable; a person who hates being rained on has a low umbrella threshold for reasons that are not social at all. [FACT]</div>
  <div class="k">Not a norm or a belief</div>
  <div class="v">The sharpest distinction in the paper uses Matza's study of delinquent boys, most of whom did not think stealing cars was right and did not particularly want to do it, but who joined because refusing carried the cost of being called a sissy. [FACT] Granovetter: "The concept of threshold, then, is purely behavioral, connoting nothing about what the actor thinks is the 'right' thing to do." [FACT]</div>
  <div class="k">Not a trait the person carries</div>
  <div class="v">"An individual's riot threshold is not a number that he carries with him from one riot to another." [FACT] A threshold is a property of a person inside a particular configuration of costs and benefits, stable only as long as that configuration is.</div>
</div>

Two endpoints define the scale. A threshold of 0 percent identifies an <span class="term" data-def="A person whose threshold is zero: someone who will act with no prior participation by anyone else. Necessary to start a cascade in the deterministic version of the model.">instigator</span>, someone who acts with no prior participation at all. A threshold of 100 percent identifies someone who will not join under any circumstances. [FACT]

## The arithmetic

Let `f(x)` be the frequency distribution of thresholds in the population and `F(x)` the <span class="term" data-def="A function giving the proportion of a population whose value falls at or below a given point. Here: the proportion of people whose threshold is at or below x.">cumulative distribution function</span>, so `F(x)` is the proportion of people whose threshold is at or below `x`. Let `r(t)` be the proportion who have joined by time `t`.

If 60 percent have joined at time `t`, then by definition the people who will have joined at `t+1` are exactly those whose thresholds are at or below 60 percent. The entire process is one difference equation: [FACT]

```
r(t+1) = F[r(t)]
```

An equilibrium is a state where the next period equals the current one, so `r(t+1) = r(t)`, which is to say:

```
F(r) = r
```

Graphically, plot the cumulative distribution against the 45 degree line where `F(x) = x`. Starting from any participation level, move up to the curve, across to the diagonal, up to the curve again. The staircase converges, and the limit is the point where the cumulative distribution **first crosses the 45 degree line from above**. [FACT] In this version of the model, where nobody is permitted to quit once they have joined, oscillation is impossible and an equilibrium is always reached. [FACT]

This is the whole formal apparatus. Everything else in the literature is a question about what `F` looks like, who counts as "others," and whether people obey the rule exactly.

## Variance decides the outcome, and the mean does not

The model's most useful result comes from holding the average constant and moving only the spread. Fix the mean threshold at 25 in a group of 100 and vary the standard deviation of a normal distribution of thresholds. [FACT]

| Standard deviation | Equilibrium participants |
|---|---|
| Small | Near zero, rising gradually |
| Just below the critical value | About 6 |
| Just above the critical value, roughly 12.2 | **Nearly 100** |
| Large | Declines toward 50 |

```chart
{"type":"line","title":"Equilibrium participation against the spread of thresholds (mean fixed at 25, N = 100)","xLabel":"Standard deviation of the threshold distribution","yLabel":"People participating at equilibrium","xMax":40,"yMax":100,"series":[{"name":"Below the critical spread","points":[[2,0],[2.2,0],[2.4,0],[2.6,0],[2.8,0],[3,0],[3.2,0],[3.4,0],[3.6,0],[3.8,0],[4,0],[4.2,0],[4.4,0],[4.6,0],[4.8,0],[5,0],[5.2,0],[5.4,0],[5.6,0],[5.8,0],[6,0],[6.2,0],[6.4,0],[6.6,0],[6.8,0],[7,0],[7.2,0],[7.4,0],[7.6,0.1],[7.8,0.1],[8,0.1],[8.2,0.1],[8.4,0.2],[8.6,0.2],[8.8,0.2],[9,0.3],[9.2,0.4],[9.4,0.5],[9.6,0.5],[9.8,0.6],[10,0.8],[10.2,0.9],[10.4,1.1],[10.6,1.3],[10.8,1.5],[11,1.7],[11.2,2],[11.4,2.3],[11.6,2.8],[11.8,3.3],[12,4],[12.2,5.5]],"tone":"ink","label":{"x":11.4,"y":24,"text":"about 6 just below σc","anchor":"end"}},{"name":"Above the critical spread","points":[[12.3,100],[12.8,100],[13.3,100],[13.8,100],[14.3,100],[14.8,100],[15.3,100],[15.8,100],[16.3,100],[16.8,100],[17.3,100],[17.8,100],[18.3,100],[18.8,100],[19.3,100],[19.8,100],[20.3,100],[20.8,100],[21.3,100],[21.8,100],[22.3,100],[22.8,99.9],[23.3,99.9],[23.8,99.9],[24.3,99.9],[24.8,99.9],[25.3,99.8],[25.8,99.8],[26.3,99.8],[26.8,99.7],[27.3,99.7],[27.8,99.6],[28.3,99.6],[28.8,99.5],[29.3,99.4],[29.8,99.4],[30.3,99.3],[30.8,99.2],[31.3,99.1],[31.8,99],[32.3,98.9],[32.8,98.8],[33.3,98.7],[33.8,98.5],[34.3,98.4],[34.8,98.2],[35.3,98.1],[35.8,97.9],[36.3,97.7],[36.8,97.6],[37.3,97.4],[37.8,97.2],[38.3,97],[38.8,96.8],[39.3,96.6],[39.8,96.3]],"tone":"ink","label":{"x":24,"y":91,"text":"jumps to ≈100, then declines","anchor":"start"}}],"annotations":[{"x":12.25,"text":"σc ≈ 12.2","anchor":"start"}],"caption":"The mean never moves. Crossing a standard deviation of roughly 12.2 takes the outcome from about six participants to nearly the whole group. Computed by forward recursion on the normal CDF; the jump reproduces Granovetter’s reported critical value.","table":[["Standard deviation","Equilibrium"],["8.0","0.1"],["12.2","5.5"],["12.3","100.0"],["20.0","100.0"],["40.0","96.3"]]}
```

<div class="finding">
  <div class="finding-head">
    <div class="name">2 &nbsp; A negligible change in variance flips the outcome discontinuously</div>
    <span class="pill high">Central mechanism</span>
  </div>
  <p>Crossing a standard deviation of approximately 12.2 moves the result from six participants to nearly the entire group. [FACT] The mean never moves. Granovetter is explicit that no sociological story attaches to this: "There is no obvious sociological way to explain why a slight perturbation of the normal distribution around the critical standard deviation should have a wholly discontinuous, striking qualitative effect. This perturbation might correspond to a minor fluctuation in the composition of a crowd ... a cause which would seem so insignificant in relation to its effect that causal attribution would never be made." [FACT] [CHARACTERIZATION: the absence of a narrative explanation is the finding, not a gap in it, because it establishes that discontinuities can exist with no proportionate cause available to an observer.]</p>
  <div class="callout"><span class="label">Why this generalizes past riots:</span> the same structure governs any binary choice with interdependent payoffs. Whether to strike, adopt a technology, leave a boring lecture, migrate, believe a rumor, go to college, or hold a stock. In each case the interesting variable is the shape of the distribution near the current level of participation, not the average sentiment.</div>
</div>

## Three mechanisms that destroy the 100-person result

The uniform-distribution cascade is the part of the paper everyone quotes. It is also a knife-edge, and Granovetter spends the middle third of the paper taking it apart. [CHARACTERIZATION: most secondary accounts omit this section entirely, which produces a version of the model considerably more confident than the original.] The three mechanisms below matter more than the demonstration, because they are what determine whether a threshold distribution's equilibrium is worth believing.

<div class="finding">
  <div class="finding-head">
    <div class="name">3 &nbsp; Social structure: people do not weight everyone equally</div>
    <span class="pill med">Important context</span>
  </div>
  <p>The simple model assumes each person responds to the raw proportion of the group. Suppose instead that friends count double. Take an individual with a 50 percent threshold in a crowd where 48 have joined and 52 have not. Under anonymity they stay out. But if they know 20 people, 15 of whom have joined, they perceive (15 x 2) + (33 x 1) = 63 participants against (5 x 2) + (47 x 1) = 57 non-participants, or 63/120 = .525. Their threshold is passed and they join. [FACT]</p>
  <p>Simulation over many randomly generated friendship structures found an asymmetry that matters more than the mechanism itself. The perturbed distribution's equilibrium of one participant is robust: where social structure changes it at all, the change rarely exceeds five to 10 people, and the modal result remains one. The uniform distribution's equilibrium of 100 is unstable against nearly any social structure at all. [FACT] The largest effects occur at a moderate density of acquaintance, where people know on average about a quarter of the group, a result Granovetter reports and declines to explain. [FACT]</p>
  <div class="callout"><span class="label">The rule that follows:</span> "When threshold distributions have very stable equilibria it may make very little difference; when these equilibria are unstable, however, the effects of social structure may overwhelm those of individual preferences." [FACT] Network structure is not always the important variable. It is the important variable precisely when the preference-based prediction is fragile.</div>
</div>

<div class="finding">
  <div class="finding-head">
    <div class="name">4 &nbsp; Sampling: the same population produces different crowds</div>
    <span class="pill med">Important context</span>
  </div>
  <p>Suppose a city's population carries the uniform threshold distribution, 1 percent of people at each value from 0 to 99, and that crowds of 100 form by random draw from it. The deterministic answer of 100 participants is now only one possible outcome among many. The probability that a drawn crowd contains no instigator is (1 - .01)^100 = .37, which yields zero participants. The probability of exactly one instigator and no one-percenter is .14, which yields one participant. Over half the time, .51, the result is zero or one. [FACT]</p>
  <p>This converts a modeling caveat into an explanation. Two cities with identical underlying threshold distributions can produce a riot and a quiet evening for reasons internal to neither. Granovetter uses it to reinterpret Spilerman's finding that the only robust predictor of 1960s racial disorders was the absolute size of a city's Black population. [FACT] If each gathering is an independent trial with a fixed probability of tipping, more gatherings produce more riots with no difference in disposition required: at a per-incident probability of .10, a city with one incident has a 90 percent chance of no large riot, while a city with 10 incidents faces (.90)^10 = .35. [FACT] [CHARACTERIZATION: this is a competing explanation for the same correlation, not a refutation of Spilerman's data, and it substitutes an aggregation mechanism for an attribute of the population.]</p>
</div>

<div class="finding">
  <div class="finding-head">
    <div class="name">5 &nbsp; Spatial dispersion, and the reverse bandwagon</div>
    <span class="pill med">Important context</span>
  </div>
  <p>The simple model assumes complete connectedness, meaning every person observes every other. Real events are not like this. Stark and colleagues found that the 1965 Watts riot was not a single incident but more than 1,850 separate cases of riot action over five days across a wide area. [FACT] Modeling clusters with limited movement between them produces a result that runs against intuition: too much movement out of a cluster that has already tipped deactivates some participants, whose departure deactivates others, and the group settles at a new and possibly much lower equilibrium. Granovetter calls this a reverse <span class="term" data-def="A self-reinforcing process in which each additional participant lowers the barrier for the next; the mechanism by which a threshold distribution converts a small initial action into a large outcome.">bandwagon</span>. [FACT] For some distributions, therefore, small movements between clusters have larger effects than big ones. [FACT]</p>
</div>

## The conditions under which the model is false

Granovetter states his own falsification conditions directly, which is unusual enough to be worth reproducing.

<div class="finding">
  <div class="finding-head">
    <div class="name">6 &nbsp; Where behavior is not contingent, the model has nothing to say</div>
    <span class="pill ok">Failure condition</span>
  </div>
  <p>The model "may be said to have been, in effect, falsified" where all or most thresholds sit at 0 or 100 percent, which is to say where people's behavior is not in fact contingent on others'. [FACT] The example is Weber's crowd of pedestrians opening umbrellas at the start of a shower: not action oriented to each other, but many people reacting the same way to the same rain. [FACT] Granovetter amends even this, noting that some people surely do wait to see whether others open theirs, but the principle stands. Assigning someone a threshold is not a claim that social influence dominates their behavior; it is only the claim that a proportion exists at which they would switch. [FACT]</p>
</div>

Three distinct measurement problems are set out separately, and they behave differently: [FACT]

<div class="kv">
  <div class="k">Systematic misperception</div>
  <div class="v">If everyone misjudges participation by the same amount, the model is unchanged. If radicals overestimate and conservatives underestimate, the model is adjustable. If misperception is random, the stability of the underlying equilibrium becomes the critical question.</div>
  <div class="k"><span class="term" data-def="Merton's term for a situation in which many members of a group privately hold a position but each believes they are alone in holding it, because nobody has made their position public.">Pluralistic ignorance</span></div>
  <div class="v">People who have decided may not make it public, so many thresholds can be passed with nobody aware of it and nobody acting. Granovetter notes this is where the model earns its keep for policy, because it names the specific obstacle to full adoption rather than attributing the shortfall to resistance.</div>
  <div class="k">Granularity</div>
  <div class="v">Nobody can reliably distinguish 15 percent from 17 percent, a limit Granovetter sources to Miller's work on human information processing. [FACT] The operational threshold is the lowest value the individual can actually perceive.</div>
</div>

One structural requirement is easy to miss. The model requires each individual's net-benefit curve to cross zero **only once**. The paper works an example of an individual with a 38 percent threshold whose net benefit later turns negative again: cautious participants who join at 50 percent but leave when participation passes 90 percent for fear of official reprisal. [FACT] The analogy offered is an unfamiliar restaurant, where an empty room is a bad sign and a full one means an unbearable wait. [FACT] Curves that cross more than once require different apparatus. The observable signature is aggregate behavior oscillating between levels while the underlying cost structure appears stable. [PROJECTION: if participation cycles without a corresponding change in costs and benefits, then at least some participants have multiply-crossing net-benefit curves, which is a diagnostic rather than a forecast.]

## Measurement, and the one empirical test in the paper

Thresholds are dispositions, so they are difficult to observe before the behavior occurs. Granovetter's defense is an analogy to microeconomics, where supply and demand schedules are also almost never measured directly and the discipline has learned to infer them from characteristics and aggregate outcomes. [FACT]

The proposed procedure: index an individual's threshold by the exact number of others who had already adopted at the moment that individual adopted, regress those thresholds on background characteristics, and use the fitted equations to predict threshold *distributions* in populations not yet exposed. [FACT]

Dozier tested this against family-planning adoption in 23 rural Korean villages over 1964 to 1973, fitting on 18 villages and predicting five. [FACT] The result was conditional. Where the regressions explained less than 50 percent of variance in thresholds, predictions were worse than those from standard methods. Above 50 percent, "the accuracy of the recursion model improves dramatically," with many predictions approaching full accuracy. [FACT] Granovetter flags the caution himself: some of the regressors were measured after the period the data covered and correlate with thresholds in ways that strip them of explanatory power. [FACT] [CHARACTERIZATION: one dissertation on one behavior in one country is a demonstration of feasibility, not a validation of the model.]

## The fifty-year argument

The model did not arrive from nowhere and it did not stay as published.

**The predecessor.** Thomas Schelling's segregation models, published 1971 to 1972, gave individuals a tolerance for the proportion of unlike neighbors and showed that mild individual preferences aggregate into near-total separation. [FACT] Granovetter credits them directly and generalizes past the specific substance to any binary choice, adding the insistence that exact distributions rather than averages determine outcomes. [FACT]

**Concealment.** Timur Kuran's *Sparks and Prairie Fires* (1989) and *Now Out of Never* (1991) apply threshold logic to political revolution by adding <span class="term" data-def="Kuran's term for the act of misrepresenting one's private preferences under perceived social or political pressure, producing a public opinion distribution that diverges systematically from the private one.">preference falsification</span>. [FACT] If people conceal opposition while opposition appears weak, private thresholds can fall for years with no visible change, accumulating what Kuran calls a latent bandwagon that a minor event then ignites. [FACT] It remains the standard account of why essentially nobody forecast the events of 1989, and structurally it is Granovetter plus a theory of why the distribution is unobservable. [CHARACTERIZATION: a reading of the relationship between the two models, not a claim Kuran makes in those terms; the "standard account" describes its position in the literature rather than a settled verdict on its correctness.]

**Networks.** Duncan Watts placed threshold agents on sparse random graphs in "A Simple Model of Global Cascades on Random Networks" (2002), identifying two regimes in which very large cascades occur very rarely, with a power-law distribution of cascade sizes where network connectivity is the binding constraint. [FACT] This is the direct ancestor of most contemporary modeling of virality and of cascading infrastructure failure. [CHARACTERIZATION: an attribution of intellectual lineage based on citation patterns and shared formalism, not a measured claim about influence.]

**Complex contagion.** Centola and Macy drew the consequence that matters most in "Complex Contagions and the Weakness of Long Ties" (2007). [FACT] Information transmits on a single contact, so long-range weak ties accelerate it. Behavior with a threshold above one requires reinforcement from several sources, and long ties then *impede* diffusion rather than aiding it. What determines whether a behavior crosses between groups is not the length of a bridge but its <span class="term" data-def="Centola and Macy's term for the number of independent ties connecting two clusters. A wide bridge can deliver the multiple reinforcements a high-threshold behavior requires; a single long tie cannot.">width</span>. [FACT] [CHARACTERIZATION: this inverts the standard reading of Granovetter's own earlier and far more famous "Strength of Weak Ties," which is a notable place for the literature to have arrived.]

<div class="finding">
  <div class="finding-head">
    <div class="name">7 &nbsp; The 25 percent number is one experimental result, not a constant</div>
    <span class="pill low">Refinement</span>
  </div>
  <p>Centola and colleagues ran online groups converging on an arbitrary naming convention, then introduced committed confederates advocating a different name. Below roughly 25 percent of the group the minority failed; at about 25 percent the majority flipped. The design used 10 groups of 20 participants. [FACT] The number has since circulated widely as a general law of social change. It should not be read that way. The authors state that the critical fraction moves with parameters, notably the memory length that stands in for how entrenched a convention is, and independent real-world replication remains thin. [FACT] A separate line of work by Wiedermann and colleagues derives the broad threshold distribution from local network interaction rather than assuming it, recovering saddle-node bifurcation and <span class="term" data-def="A system property where the path back is not the path forward: the level of support needed to establish a new state is higher than the level needed to maintain it, so reversing a tip requires more than undoing the trigger.">hysteresis</span> with critical minority sizes near 20 percent. [FACT] [CHARACTERIZATION: the robust claim across this work is that a sharp critical fraction exists in some settings, not that it has a fixed value, and citing "25 percent" as a target is a stronger claim than any of the underlying studies makes.]</p>
</div>

## The live disagreement: is the knife-edge real?

Granovetter's central claim is that cascades are hypersensitive to small perturbations in group composition. That claim is the source of the model's explanatory power and also its most direct challenge.

<div class="finding">
  <div class="finding-head">
    <div class="name">8 &nbsp; A small amount of individual randomness makes collective behavior more predictable, not less</div>
    <span class="pill low">Refinement</span>
  </div>
  <p>Macy and Evtushenko revisited the model in "Threshold Models of Collective Behavior II" (2020) with a minimal change: allow individuals to behave randomly some small fraction of the time rather than obeying their thresholds exactly. [FACT] The results run against expectation. A little noise makes collective outcomes *less* sensitive to small perturbations in composition and therefore more predictable. In groups facing an otherwise intractable start-up problem, meaning no instigator, randomness produces what they term spontaneous instigation, so that outcomes come to track the strength of the group's collective interest rather than the accident of who happened to be present. Both effects grow more pronounced as group size increases. [FACT] Their stated conclusion is that "randomness is often assumed to be a theoretically unimportant residual category" and that individual idiosyncrasy needs to be brought back into the study of collective behavior. [FACT]</p>
  <div class="callout"><span class="label">What this does to the original:</span> [CHARACTERIZATION: the deterministic knife-edge is partly an artifact of assuming people execute their thresholds exactly. The methodological warning against inferring dispositions from outcomes survives intact, because it does not depend on the sensitivity being extreme; the implication that collective outcomes are close to unforecastable does not survive as stated.]</div>
</div>

## How the model gets misapplied

Threshold language has been absorbed into marketing and social-media strategy in a form the source material does not support, usually reduced to the instruction to locate and recruit the people with threshold zero.

Watts spent much of the following decade arguing that this reverses the causal structure. [FACT] In a threshold cascade the outcome is determined by the distribution across the whole population; who happens to occupy the instigator role is largely incidental, because instigators are necessary but not sufficient and are not scarce. [CHARACTERIZATION: the popular version relocates the causal weight from a population property, which is hard to change and hard to observe, onto identifiable individuals, which is easier to sell and easier to act on.] [PROJECTION: if the threshold distribution near the current participation level contains no one who would join at that level, then no choice of instigator produces a cascade, which is the condition under which influencer-targeting fails without any observable reason.]

A second misapplication is treating a threshold model as a forecasting instrument. The exact-distribution sensitivity that makes the 100-person demonstration memorable is the same property that makes forecasting impractical, since an answer that turns on a one-person gap requires an input measured to a precision nobody has. [CHARACTERIZATION: the model explains variance and surprise rather than predicting outcomes, and the Dozier result, where accuracy depended entirely on whether the regressions cleared 50 percent of variance, is consistent with that reading.]

## What the model licenses

<div class="kv">
  <div class="k">The question it replaces</div>
  <div class="v">Not "what does this group believe," which is often unanswerable and frequently not the operative variable, but "at what level of participation would each person switch."</div>
  <div class="k">The diagnostic</div>
  <div class="v">Look at the shape of the threshold distribution immediately above the current participation level. If there is nobody who would join at the level already reached, the process stops there regardless of how favorable average sentiment is.</div>
  <div class="k">The inference it forbids</div>
  <div class="v">Reading individual dispositions off a collective outcome. A large result is not evidence of strong preference, and a null result is not evidence of weak preference, wherever behavior is interdependent.</div>
  <div class="k">The stability question</div>
  <div class="v">Before treating any predicted equilibrium as real, ask whether it survives a one-person change. If it does not, network structure and sampling accident will dominate preferences, and the prediction is not usable.</div>
  <div class="k">Where it does not apply</div>
  <div class="v">Situations where behavior is not contingent on others', where thresholds cluster at 0 or 100 percent, or where individuals' net benefit from participating rises and then falls again.</div>
</div>

The durable contribution is the negative one. Granovetter supplied a mechanism and a worked counterexample for an inference people make constantly, in political commentary, in post-mortems, in market narratives, and in retrospectives of every kind: that a group which did something must have wanted to. Under interdependence, sometimes it did not, and the arithmetic shows how.

## Sources

- [Mark Granovetter, "Threshold Models of Collective Behavior," *American Journal of Sociology* 83(6), 1978, pp. 1420-1443](https://doi.org/10.1086/226707) - the source of the model, the 100-person demonstration, the difference equation, the normal-distribution result, the social-structure and sampling analyses, the falsification conditions, and the Dozier test. [Full text PDF](https://www.cse.cuhk.edu.hk/~cslui/CMSC5734/Granovetter-threshold_models.pdf).
- [Thomas Schelling, "Dynamic Models of Segregation," *Journal of Mathematical Sociology* 1, 1971, pp. 143-186](https://doi.org/10.1080/0022250X.1971.9989794) - the predecessor model Granovetter credits, applying tolerance thresholds to residential choice.
- [Mark Granovetter and Roland Soong, "Threshold Models of Diffusion and Collective Behavior," *Journal of Mathematical Sociology* 9, 1983, pp. 165-179](https://doi.org/10.1080/0022250X.1983.9989941) - the follow-up work extending the model to diffusion, later applied to consumer behavior and residential segregation.
- [Timur Kuran, "Now Out of Never: The Element of Surprise in the East European Revolution of 1989," *World Politics* 44(1), 1991, pp. 7-48](https://doi.org/10.2307/2010422) - preference falsification, the latent bandwagon, and the unforecastability of revolution.
- [Timur Kuran, *Private Truths, Public Lies: The Social Consequences of Preference Falsification* (Harvard University Press, 1995)](https://www.hup.harvard.edu/books/9780674707580) - the book-length treatment.
- [Duncan J. Watts, "A Simple Model of Global Cascades on Random Networks," *PNAS* 99(9), 2002, pp. 5766-5771](https://www.pnas.org/doi/10.1073/pnas.082090499) - threshold agents on sparse random graphs; the two cascade regimes and the power-law size distribution.
- [Damon Centola and Michael Macy, "Complex Contagions and the Weakness of Long Ties," *American Journal of Sociology* 113(3), 2007, pp. 702-734](https://www.journals.uchicago.edu/doi/10.1086/521848) - the distinction between simple and complex contagion and the argument that bridge width, not length, governs the spread of high-threshold behavior.
- [Damon Centola, Joshua Becker, Devon Brackbill, and Andrea Baronchelli, "Experimental Evidence for Tipping Points in Social Convention," *Science* 360, 2018, pp. 1116-1119](https://www.science.org/doi/10.1126/science.aas8827) - the committed-minority experiment and the roughly 25 percent critical fraction, with the authors' own caveats about its dependence on parameters. [Open-access version](https://openaccess.city.ac.uk/id/eprint/20031/1/CM_CRO_Version.pdf).
- [Marc Wiedermann, E. Keith Smith, Jobst Heitzig, and Jonathan F. Donges, "A Network-Based Microfoundation of Granovetter's Threshold Model for Social Tipping," *Scientific Reports* 10, 2020](https://pmc.ncbi.nlm.nih.gov/articles/PMC7343878/) - deriving the threshold distribution from local network interaction; saddle-node bifurcation, hysteresis, and critical minorities near 20 percent.
- [Michael W. Macy and Anna Evtushenko, "Threshold Models of Collective Behavior II: The Predictability Paradox and Spontaneous Instigation," *Sociological Science* 7, 2020, pp. 628-648](https://sociologicalscience.com/articles-v7-26-628/) - the effect of individual randomness on cascade sensitivity, and the case for returning idiosyncrasy to the model.
- [David Matza, *Delinquency and Drift* (Wiley, 1964)](https://www.google.com/books/edition/Delinquency_and_Drift/wJk9AAAAIAAJ) - the study of delinquent boys Granovetter uses to separate thresholds from norms.
- [George A. Miller, "The Magical Number Seven, Plus or Minus Two," *Psychological Review* 63, 1956, pp. 81-97](https://psycnet.apa.org/doi/10.1037/h0043158) - the limit on perceptual distinctions underlying the granularity problem.

---

# The Punctum and the Blind Field
*An investigation of an aesthetic doctrine built on a single instruction, render the felt presence of what is not given, and of the constellation of counterpart ideas, Western and Japanese, that it was assembled from.*

Source: https://theholbrookreport.com/reports/punctum-and-the-blind-field/ · Published: 2026-06-19

<div class="author-note">
<p>In my continued attempt to build out an emergent and interesting world for the music that I write, I stumbled across the phrase "Barthes's Punctum". I was so intrigued that I did a deep dive on it, and it ended up being even more incredible of a concept than I could have hoped to imagine.</p>
<p>Truly it allowed me to put into words that which I had felt scratching at my mind for most of my life. The feeling of what else lay beyond what you can see. What you see that intrigues you past the frame.</p>
<p>I hope you, like I did, find this illuminating.</p>

<p>If you're curious about the artifact I was able to build with the help of this philosophy and thought technology, it exists here: <a href="https://www.youtube.com/watch?v=tI-ipcwdFBM">The Long Way Home</a>.</p>
  <span class="byline">— <a href="https://aaronholbrookmusic.com">Aaron Holbrook</a></span>
</div>

<div class="callout"><span class="label">What this piece examines, and what it is:</span> The subject is a written doctrine, "The Blind Field," that serves as the rubric for an automated system that generates and then judges image sequences for music video. The doctrine is an operating document, not a published theory; this report treats it as the specimen and investigates its architecture, its sources, and the one contradiction at its center. The report is AI-authored. Because most of its claims are interpretive rather than measured, the apparatus below leans on the [CHARACTERIZATION] label, and the reader should read it as a study of an argument, not a delivery of settled fact.</div>

A generative image system is, by construction, a machine for satisfying instructions. It is told what to render and it renders it, fluently, by the dozen. The doctrine examined here exists to discipline that fluency against itself, and it does so from a single instruction that points the opposite way:

> Render the felt presence of what is not given. Leave room.

Every rule downstream is a reading of that sentence. A frame, a cut, a line of on-screen text, or a camera move is judged good to the degree that it makes a viewer feel the weight of something not shown: an interior behind a turned face, a room past the doorway, a moment just gone or about to arrive. The doctrine forbids stating the feeling and requires leaving the image open enough that a stranger's own feeling can find a place to land. The sections below trace where that idea comes from, the counterpart concepts it was assembled out of, the live disagreement it rests on, and the structural problem it creates for the machine that runs it.

The pills in this piece mark how load-bearing each idea is to the doctrine, not how dangerous anything is: <span class="pill high">Keystone</span> for the ideas the whole structure stands on, <span class="pill med">Major counterpart</span> for the supporting concepts, <span class="pill low">Nuance</span> for refinements, and <span class="pill ok">Settled</span> for the points the tradition treats as mutually confirming rather than contested. Every claim is labeled [FACT] (a sourced point of record, such as who wrote what and when), [CHARACTERIZATION] (a defensible label placed on those facts), or [PROJECTION] (a conditional "if X then Y," never a prediction).

## The keystone: the sealed image and the blind field

The doctrine takes its name and its central mechanism from one distinction in Roland Barthes' *Camera Lucida* (1980), the book he wrote in the months after his mother's death and the last he published before he died. [FACT] Barthes drew a hard line between the photograph and the film. A photograph, he argued, is normally <span class="term" data-def="Barthes's term for a still photograph that has no offscreen: the frame is a wall and the world does not continue past its edge.">sealed</span>: it has no <span class="term" data-def="French champ aveugle, the offscreen life of an image, the world that continues past the frame's edge, a before and an after. Barthes held that film always has one and a photograph normally does not.">blind field</span>, no *champ aveugle*. What is inside the frame is all there is; the border is a wall, and the world does not spill past it. Film, by contrast, always has a blind field, because figures move in and out of frame and the image carries a before and an after. [FACT]

The idea becomes a doctrine because of what Barthes claimed next, and because of the specific medium this system works in. The system does exactly one thing at its core: it takes still images, which are sealed by nature, and sets them into motion. That motion is the whole art.

<div class="finding">
  <div class="finding-head">
    <div class="name">1 &nbsp; Converting the sealed photograph into a film is the entire operation</div>
    <span class="pill high">Keystone</span>
  </div>
  <p>A generated frame arrives walled, complete, finished at its edges. The doctrine's core move is to give that frame the blind field a photograph should not have: through a slow push-in, a pull-back, a pan that follows the eye off the edge, or a cut against a neighboring frame, the still acquires an offscreen and a continuing life. [CHARACTERIZATION: the doctrine names this the conversion of the photograph into the film, and treats the conversion itself, not the frame, as the deliverable.] This is why the doctrine privileges frames that already imply an offscreen before anything moves them: a road bending out of view, a lit window seen from the dark outside, a path over a hill, a figure facing away into distance, an interrupted gesture, something leaving the edge. Those are frames that want to be opened.</p>
  <div class="callout"><span class="label">The operational consequence:</span> a camera move is reclassified as a blind-field operation rather than a flourish. The push-in approaches an unguarded detail; the pull-back reveals the world the frame was hiding; the pan asserts the offscreen. The instruction is to choose the move that opens the most room, not the one that looks busiest.</div>
</div>

The reason a single critic's distinction can bear this much weight is that it matches the medium precisely. Most aesthetic theory addresses either the still or the moving image. The doctrine's medium is the seam between them, the still image made to move, and Barthes is one of the few who wrote directly about that seam. [CHARACTERIZATION: the fit is unusually exact, which is why the doctrine elevates this above the dozen other ideas it also draws on.]

## The wound in the photograph: studium, punctum, and time

The blind field does not open on its own. In Barthes' scheme it is broken open by a particular kind of detail, and that detail is the core of *Camera Lucida*, the 1980 book that was his last and that he wrote as two things at once: a meditation on what photography is, and a grief-work undertaken in the months after his mother's death. [FACT] He reaches his central idea, the punctum, by contrast with its opposite, the studium, so the two are best taken together. They are counterparts in the strict sense: each is defined by not being the other, and neither is fully legible alone. Because the doctrine inherits the pair wholesale, and then inherits the deeper, temporal turn Barthes gives the punctum halfway through the book, this section sets all of it out at full size rather than in the shorthand the idea is usually flattened to.

<div class="kv">
  <div class="k"><span class="term" data-def="Barthes's term for the cultivated, general interest a viewer takes in a photograph: what can be discussed, admired as craft, placed historically. It engages liking, not feeling.">Studium</span></div>
  <div class="v">The coded, shareable, polite layer of a picture. "A good shot." "Captures the era." It never wounds; it engages taste, not feeling.</div>
  <div class="k"><span class="term" data-def="Latin for a point, prick, or small wound. Barthes's term for the contingent detail that pierces a viewer personally, usually not the subject and not anything the photographer intended.">Punctum</span></div>
  <div class="v">The detail that breaks through the studium and stings: a costume detail, a bandaged finger, a pair of strapped shoes. Contingent, often tiny, not analyzable. It happens to you.</div>
</div>

### What the studium actually is

Barthes took the word from the Latin *studium*, meaning application or study, a general and dutiful kind of enthusiasm. [FACT] The studium is everything in a photograph a viewer can talk about: the subject, the period, the technique, the politics, the document it provides. To recognize it, Barthes wrote, is inevitably to encounter the photographer's intentions, to enter into harmony with them, to approve or disapprove of them, but always to *understand* them, because the studium is a contract between the maker and the viewer carried out in a shared code. [FACT] He placed it deliberately low on the register of feeling: it is of the order of liking, not of loving, the field of an "inconsequential taste" where the most a viewer says is *I like this, I do not like this*. [FACT] The studium can be intense, informative, even admirable, and it still does not wound, because everything in it is intended, legible, and therefore safe. [CHARACTERIZATION: the doctrine reads the studium not as bad photography but as competent, fully-intended photography, which is a sharper problem because competence is exactly what an image generator delivers by default.]

### What the punctum actually is

The punctum is the rupture in that contract. Barthes searched for a word and settled on the Latin *punctum*, which means a sting, a speck, a cut, a little hole, and also the throw of the dice that decides a game: the punctum is the accident in a photograph that pricks the viewer, that "bruises" them, that is poignant to them personally. [FACT] His most quoted sentence makes its motion physical:

> It is this element which rises from the scene, shoots out of it like an arrow, and pierces me. [FACT]

Three properties follow, and the doctrine depends on all three. First, the punctum is almost never the subject. It is a detail, a "partial object," something off to the side that the photographer did not compose for and frequently did not notice. Second, it is uncoded: unlike the studium it cannot be discussed into existence or taught, because it is not a meaning but an event that happens to one viewer. Third, and most counterintuitive, the detail has a power of expansion: though it is small and incidental, it can flood the entire photograph, so that a single true speck redeems or overwhelms everything composed around it. [FACT] Barthes also noted its strange temporal status, that the punctum is something the viewer adds to the photograph and which is nonetheless already there, which is the seam where his argument later opens into the dispute over whether the wound is in the image or in the viewer. [FACT]

### Three photographs Barthes uses, and what they show

The pair is abstract until it is watched working on specific images. Barthes' own examples are the clearest illustration, and each isolates a different face of the distinction.

In Lewis Hine's 1924 photograph of two children in a New Jersey institution, the studium is unmistakable and is what the picture was taken for: the clinical and social fact of the children's condition, the document of the institution. Barthes refuses it. What he stubbornly sees instead, he wrote, is the boy's huge Danton collar and the girl's finger-bandage. [FACT] The wound is a costume detail and a small dressing, the two most incidental things in the frame, and naming them is a near-perfect demonstration that the punctum is precisely what the photographer did not intend the viewer to look at. [CHARACTERIZATION: the example is doing the doctrine's central work in miniature, locating the feeling in the unintended margin rather than the composed subject.]

In a 1926 James Van der Zee studio portrait of a Black American family, Barthes first reads the studium with sympathy: respectability, family cohesion, Sunday best, the whole effort of social advancement the sitters are performing. [FACT] The detail that pricks him is at first the strapped pumps worn by one of the women, and then, on reflection, a thin gold necklace, which wounds him because he had seen one like it worn by a woman in his own family. [FACT] This second example carries more than the first, because the source of the wound has migrated out of the photograph and into Barthes' private memory, which is the exact point a later critic seizes on (see [the disagreement at the center](#the-disagreement-at-the-center-can-the-wound-be-placed)). [CHARACTERIZATION: the Van der Zee reading is the doctrine's hinge example, the one where the punctum is visibly half-supplied by the viewer rather than delivered by the surface.]

Alexander Gardner's 1865 portrait of Lewis Payne, the conspirator photographed in his cell while waiting to be hanged, supplies the third and deepest face. The young man is handsome and the picture is handsome, and that is the studium. The punctum, Barthes wrote, is that he is going to die: reading the image now, one sees at once *this will be* and *this has been*, so that the sitter is dead and is going to die in the same glance. [FACT] Here the wounding detail is not a collar or a necklace but Time itself, which is the bridge from the punctum-as-detail to the punctum-as-mortality that the next section takes up.

There is also the photograph Barthes will not show. The image that organizes the entire second half of *Camera Lucida*, a portrait of his mother as a child in a winter garden, is described at length but never reproduced, because, he explained, it exists only for him; for any other viewer it would be nothing but an indifferent picture. [FACT] The refusal is itself the argument: a punctum cannot be transmitted, because it is not a property the printed page could carry to a stranger. [CHARACTERIZATION: the unreproduced photograph is the doctrine's proof that the wound is not shippable, which is why a system cannot deliver one and can only leave room for one.]

### The second movement: the punctum becomes Time

The Payne portrait is the hinge into the book's deeper claim, the movement that is easy to miss because it arrives after the famous detail-hunting and quietly redefines the term. Halfway through *Camera Lucida*, around the photograph of his mother in the winter garden, Barthes recasts the punctum as Time itself. The deepest wound of any photograph, he argues, is not a particular detail but the flat, certain fact he names the <span class="term" data-def="French ça-a-été, that-has-been: Barthes's name for the photograph's testimony that the thing was really there at one real instant now irrevocably gone. He called it the essence, or noeme, of photography.">that-has-been</span> (*ça-a-été*): the thing in front of the lens really existed, at one real instant, and that instant is now irrevocably past. [FACT] The photograph becomes a certificate of presence that is also a certificate of death. Reading a young man in an old portrait, Barthes registers both tenses at once, that the sitter is going to die and is already dead, and he calls this double testimony the <span class="term" data-def="From the Greek for what is thought or grasped: Barthes used noeme for the essential nature of a thing. The noeme of photography, for him, was that-has-been.">noeme</span>, the essential nature, of photography. [FACT]

This temporal punctum is what lifts the concept above aesthetic vocabulary, because it ties the medium to loss in a way Barthes held no other medium manages. Painting can imagine its subject and language can refer to it, but a photograph insists that its referent was actually there in front of the lens, which makes every photograph also the trace of something gone. [CHARACTERIZATION: the doctrine reads this as why its native register is elegiac rather than chosen, since the medium testifies to passing before any subject is selected.] For a system whose raw material is already images of fixed past instants, the directive follows without strain: lean into time rather than fight it, and favor dusk over noon, the turning season, patina, and last light, the frame that says *this was* over the frame that says *this is*.

### A concept built to resist method

The punctum is deliberately hard to use, and the difficulty is the point of it rather than a flaw in it. Barthes is partly writing against a purely semiotic or structuralist reading of images: the studium, with its shared codes and legible intentions, is roughly the semiotician's territory, and the punctum is his name for precisely what escapes coding. [CHARACTERIZATION: this is the interpretive frame most readers of the book settle on, that the punctum is an anti-method, a placeholder for what analysis cannot reach.] The move carries a real irony, because Barthes had built much of his earlier career on exactly the coded analysis the punctum refuses, from *Mythologies* (1957) to the essays collected in *Image-Music-Text* (1977). [FACT] The late turn is a critic dismantling his own instrument.

That difficulty matters for anyone trying to operationalize the idea, which is what the doctrine attempts. Because the punctum is personal and contingent, it resists doing critical work on behalf of anyone but the single viewer it happened to, and it sits closer to a phenomenology of being affected than to a theory that can be applied to a case. [CHARACTERIZATION: the doctrine treats this not as a reason to discard the concept but as the source of its one usable instruction, that since the wound cannot be specified it can only be left room for.] The same ambiguity is why the term has been so generative and so contested in the decades since, and it sets up the dispute the doctrine quietly stakes its whole method on (see [the disagreement at the center](#the-disagreement-at-the-center-can-the-wound-be-placed)).

### Why the pair is unforgiving to the machine

Laid out at full size, the distinction is harder on a generative system than the slogan version suggests. The studium is not merely "generic output." It is *intended, legible, fully-realized* output, an image that perfectly executes the photographer's, or the prompt's, intentions in a shared code, and that is exactly what a competent image model is built to produce. [CHARACTERIZATION: the doctrine names the studium "the enemy you will produce by default" and treats the flawless, maximally coded frame, the visual equivalent of the lofi-girl-at-the-desk cliche, as the thing to reject rather than the thing to want, precisely because it is the system performing at its best.] The punctum fails every condition a prompt can satisfy: it is not the subject, so it cannot be specified; it is uncoded, so it cannot be described; and it is partly supplied by the viewer, so it cannot be placed in the file at all.

<div class="finding">
  <div class="finding-head">
    <div class="name">2 &nbsp; The desirable quality is the one the system cannot be instructed to produce</div>
    <span class="pill high">Keystone</span>
  </div>
  <p>If the wounding detail is contingent, off-center, and unintended, then it cannot be ordered, which inverts the usual logic of a prompt-driven tool where more instruction means more control. [CHARACTERIZATION: the doctrine concludes that the operator's job is never to insert the affecting detail like a prop, but to leave the image open and unguarded enough that one might occur, be noticed, and survive the cut.] Two operating rules follow directly from Barthes' properties. Because the punctum is a margin detail, the doctrine favors frames thick with incidental, slightly accidental specifics over frames that are all deliberate subject, on the reasoning that more unguarded surface means more places a stranger's eye could catch. Because the punctum has the power of expansion, a single true detail is treated as enough: a frame is not required to be wounding everywhere, only to carry one uncomposed speck that could flood it. [PROJECTION: if a prompt fixes every element of a frame, then it forecloses the unguarded margin where a contingent detail would have grown, and the result trends toward pure studium, the conditional the doctrine operates on.]</p>
</div>

This is also the doctrine's quiet rebuttal to its own production model. A system that can generate a hundred competent frames has made the competent frame worthless, because competence is the studium and the studium does not wound. The only frame that earns its place is the one carrying something the competence could not have ordered.

## The lineage it descends from: a century-long argument

Barthes did not invent the idea that the affecting part of a picture is the part no one chose. The doctrine treats this as a single argument running across roughly a century, and it instructs the system to hold the voices at once rather than picking one.

The earliest is Walter Benjamin, writing in 1931 about the first photographic portraits. [FACT] Benjamin noticed that even a stiffly posed photograph holds a residue of the real that the photographer never controlled, a "tiny spark of contingency," the here-and-now of a moment burning itself into the plate. He described the viewer's urge to search the old image for the inconspicuous spot where the long-dead instant still seems alive, and he named the layer of an image that the camera records but no one was aware of the <span class="term" data-def="Walter Benjamin's term for the layer of a photograph the camera captures but neither photographer nor subject noticed, the way slow-motion film reveals a movement no eye could catch.">optical unconscious</span>. [FACT] The instruction the doctrine draws from this is to read a finished frame not only for whether it executed the brief but for the unbidden specific in the margin that the brief never asked for, and to value that accidental rightness above faithful execution. [CHARACTERIZATION: Benjamin's spark and Barthes's punctum are treated as the same phenomenon described fifty years apart, one in the language of history, the other in the language of the wound.]

Barthes himself stands at the center of this lineage; his punctum and its temporal deepening into the that-has-been were set out in full above, which leaves the two figures who frame him, Benjamin just described before him and Berger working beside him.

The warmer counterweight in the lineage is John Berger, whose *Ways of Seeing* (1972) and *About Looking* (1980) ground all of this in ordinary human looking rather than metaphysics. [FACT] For Berger a photograph is a quotation from appearances, a fragment lifted from the flow of the seen and made meaningful by what is remembered and said around it. [FACT] The doctrine keeps Berger close as a check against its own vocabulary: the frame has to work on a person who has never read any of this, and if it only works once explained, it does not work.

## The disagreement at the center: can the wound be placed?

An investigation has to find the load-bearing wall and push on it. For this doctrine the wall is a claim that is not settled, and the doctrine is unusual in that it builds the unsettledness in rather than papering over it.

Margaret Olin, re-reading *Camera Lucida*, noticed that a punctum Barthes describes in detail, a necklace in a portrait, does not actually appear in the photograph he is discussing; he seems to have imported it from another image or another memory. [FACT] Her argument, and the doctrine is careful to mark this as her contested claim and not settled fact, is that the slip is the whole tell: the punctum may not live in the photograph at all, but may be what the wounded viewer brings to it, a wound that finds a surface rather than a surface that delivers a wound. [FACT] From another direction, Michael Fried tried to pull the punctum back toward something the photograph's own structure can earn, locating it in a subject's absorbed obliviousness to being seen. [FACT]

<div class="finding">
  <div class="finding-head">
    <div class="name">3 &nbsp; The doctrine's most operational principle rests on its least settled premise</div>
    <span class="pill high">Keystone</span>
  </div>
  <p>If Olin is right that the viewer supplies the wound, then the only thing a maker can do is leave the surface unguarded enough for a stranger's own arrow to land. [CHARACTERIZATION: the doctrine adopts Olin's contested reading as its working assumption precisely because it is the one that yields an instruction, where Fried's reading would tell the system to author the punctum and Olin's tells it to get out of the way.] The practical residue is the same whichever theorist is right, which is why the doctrine can lean on an unsettled debate without waiting for it to resolve: a frame full of incidental, slightly accidental specifics offers more surface and more unguarded spots than a frame that is all deliberate subject, so it gives more places for a viewer's memory to catch regardless of where the wound actually originates. [PROJECTION: if the wound is partly viewer-supplied, then maximizing unguarded surface raises the chance of contact, the bet the doctrine places.]</p>
  <div class="callout"><span class="label">Why the contested premise is a feature:</span> a doctrine that depended on the punctum living provably in the image would be paralyzed by the open question. By taking the viewer-supplied reading as its operating assumption, the doctrine converts an unresolved academic dispute into a usable rule, leave room, that holds up under either answer.</div>
</div>

The other voice the doctrine keeps deliberately uncomfortable is Susan Sontag. *On Photography* (1977) is the cold, suspicious foil to Barthes' warm wound. [FACT] Where Barthes asks what an image does to the viewer, Sontag asks what making it does to the world: to photograph is to appropriate, to convert lived experience into a collectible, and the glut of images breeds anesthesia, so that the more we see the less we feel. [FACT]

<div class="finding">
  <div class="finding-head">
    <div class="name">4 &nbsp; Sontag is load-bearing here, not a footnote, because the system is a glut engine</div>
    <span class="pill med">Major counterpart</span>
  </div>
  <p>Sontag's critique was written about human photographers and lands harder on a machine that manufactures images on command, by the dozen, which is structurally a glut engine. [CHARACTERIZATION: the doctrine installs Sontag as a permanent check rather than a historical citation, turning her alarm into a test the system runs on its own output.] The test is one question: is this a wound, or is it more glut wearing a beautiful coat? The operational answers follow from it: value the frame that pierces over the frame that merely impresses, prefer fewer and truer over more and prettier, and treat the system's own fluency as suspect. [CHARACTERIZATION: this is the rare case of a tool adopting, as a core rule, the strongest available argument against the category of tool it belongs to.]</p>
</div>

## The parallel tradition: the same structure in another language

The most striking thing the investigation surfaces is that the doctrine's central instinct, the pull toward what lies around the bend, over the hill, behind the door, was mapped centuries before Barthes by a tradition that never read him. The doctrine pulls four terms from Japanese aesthetics and treats them not as decoration but as a second, independent vocabulary for the same gap. That two unrelated traditions converge on one structure is the strongest evidence the doctrine has that the structure is real and not a private taste.

<div class="finding">
  <div class="finding-head">
    <div class="name">5 &nbsp; Two unrelated vocabularies describe one structure, which is why the convergence is treated as confirmation</div>
    <span class="pill ok">Settled</span>
  </div>
  <p>The Western image-wound lineage and the Japanese aesthetics lineage arrive at the same place from opposite directions and without contact, the former through a theory of the contingent detail, the latter through a theory of charged emptiness. [CHARACTERIZATION: the doctrine reads the agreement of two independent traditions as confirmation that it is describing a property of attention rather than a stylistic preference, the one point in the document treated as settled rather than argued.] The clearest single pairing is temporal: Barthes' that-has-been and the Japanese pathos of passing things are, in the doctrine's words, two languages naming one ache.</p>
</div>

| Western concept | Japanese counterpart | The shared move |
|---|---|---|
| Blind field (the offscreen the punctum opens) | <span class="term" data-def="Japanese for the charged interval, the meaningful emptiness between things: the gap, pause, or negative space that makes what is present mean something. Not absence but a presence made of space.">Ma</span>, the charged interval | Emptiness is content; the gap does the work |
| Suggestion over statement (the half-seen) | <span class="term" data-def="Japanese aesthetic of the deep, mysterious beauty of what is suggested rather than shown: the mist over the mountain, the implication of more than is given.">Yūgen</span>, profound suggestion | Withhold and imply; do not hand over all meaning at once |
| That-has-been, the ache of the passing instant | <span class="term" data-def="Japanese for the pathos of things: the tender, bittersweet sadness at the impermanence of everything, emblematized by the falling cherry blossom, beautiful because it is going.">Mono no aware</span> | Favor the transient and the last-of-its-kind |
| Earned imperfection, the unguarded surface | <span class="term" data-def="Japanese aesthetic finding beauty in imperfection, impermanence, and incompleteness: the rustic, weathered, asymmetrical, unfinished. Sabi in particular is the beauty age and solitude confer.">Wabi-sabi</span> | The flaw is the opening; subtraction over addition |

The four terms also do operational work the Western lineage leaves implicit. *Ma* gives the doctrine its rule about time as well as space: the silent beat, the montage moment with no on-screen text, the frame left to sit while nothing is said. [CHARACTERIZATION: the instruction not to fill every beat is justified by ma, which treats the interval as content rather than as a thing to be filled.] *Wabi-sabi* gives the philosophical license for the system's deliberate degradation layer, the analog video artifacts and grain it composites in: a flawless render is read as a sealed render, and the dropout, the tracking error, and the slightly-wrong are read as openings rather than defects. [CHARACTERIZATION: the doctrine inverts the usual production value, treating polish as closure and imperfection as room, and grounds the inversion in wabi-sabi rather than in nostalgia.]

## Emergence, and the authorship problem it solves

A system that generates frames the operator did not choose shot by shot has an authorship problem: who made the good one, and can it be trusted if no one decided it? The doctrine answers with Brian Eno, who built a practice on designing a set of rules that, once set in motion, make the work for the composer, so that the maker becomes as ignorant of the work's future as any listener. [FACT] Eno's generative music, from *Music for Airports* (1978) onward, treats the beauty as <span class="term" data-def="Beauty that arises from the interaction of simple parts in ways the maker set up but did not dictate, as in Brian Eno's generative music where rules, once started, produce the work.">emergent</span>: it arises from the interaction of simple parts in ways the maker arranged for but did not dictate. [FACT]

<div class="finding">
  <div class="finding-head">
    <div class="name">6 &nbsp; A divergence that still belongs is the system working, not failing</div>
    <span class="pill med">Major counterpart</span>
  </div>
  <p>The system is, structurally, an Eno machine: a set of constraints (a world recipe distilled from a seed image, a fidelity dial, an arc), set in motion to produce frames no one selected individually. [CHARACTERIZATION: the doctrine treats this as a feature rather than a tolerance, and instructs the operator to keep the frame that is better than what was in their head even when it diverges from the literal brief, provided it still belongs to the world.] The deeper claim ties the whole document together: authored beauty closes, because it is fully decided, while emergent beauty leaves the door open for the contingent detail to occur. [CHARACTERIZATION: Eno's emergence, Benjamin's contingency, Barthes' unintended punctum, and Olin's viewer-supplied wound are read as four descriptions of one move, to leave room, approached from four directions.]</p>
</div>

## The destination: the viewer's interior

The lineage so far concerns the image. The last cluster of counterparts concerns the person looking, because the doctrine's stated aim is not a property of the frame but an event in a stranger. Gaston Bachelard's *The Poetics of Space* (1957) studied how certain images, the house, the nest, the shell, the corner, the far horizon, open <span class="term" data-def="A waking daydream that an image triggers in a viewer. For Bachelard the poetic image does not illustrate a feeling, it starts one.">reverie</span> in the viewer. [FACT] For Bachelard the poetic image is not an effect of a feeling but an origin of one: it does not illustrate, it starts. [FACT] The doctrine reads this as the mechanism behind the whole project and favors the archetypal reverie-images, thresholds, windows, paths, shelters, the lit interior glimpsed from the dark, treating the image as a place to enter and wander rather than a statement to receive and be done with.

Two further counterparts aim the same instinct at people and at language. <span class="term" data-def="A coinage from John Koenig's Dictionary of Obscure Sorrows for the realization that every passing stranger has an interior life as vivid and complex as one's own.">Sonder</span> is the realization that every passing stranger has an interior as vivid as one's own. [FACT] The doctrine reads it as the punctum aimed at people, and it is the stated reason a figure's face is so often turned, hooded, or distant: a hidden face asserts an interior precisely by refusing to show it, where a clearly emotive face narrates the feeling and seals the frame. <span class="term" data-def="German for an inconsolable longing toward something undefined and unreachable, a yearning for a place or state one cannot name.">Sehnsucht</span>, the ache of longing for something unnameable, is named as the emotional register the whole sequence should aim at, the road toward rather than the arrival.

The counterpart that governs on-screen text is W. G. Sebald, whose melancholy prose was laced with grainy, uncaptioned photographs that never illustrate the words beside them. [FACT] The image and the text live in different rooms that share a hallway, and the refusal to let either explain the other is where the reader's feeling pools. [FACT]

<div class="callout"><span class="label">The text rule, stated as a law:</span> a line of on-screen text and the frame beneath it are generated from the same beat but must never describe each other. The word points where the picture does not; the picture shows what the word does not say. The gap between them is the opening, and a caption that names the feeling or describes the image seals the frame and reads as cringe, which is the failure mode the Sebald rule exists to prevent.</div>

## The contradiction the doctrine is built around

Pulling the threads together surfaces the one finding an honest investigation has to foreground: the doctrine is, in large part, a sustained argument against the nature of the thing that executes it.

<div class="finding">
  <div class="finding-head">
    <div class="name">7 &nbsp; A wound doctrine running on a glut engine is the central tension, and it is deliberate</div>
    <span class="pill high">Keystone</span>
  </div>
  <p>A generative image model is optimized to fill the frame, satisfy the prompt, and produce more on demand. The doctrine asks for the opposite on every axis: empty the frame (ma), disobey the prompt's completeness (the unguarded margin), and prefer fewer and truer to more and prettier (Sontag). [CHARACTERIZATION: the doctrine is best read not as instructions for using the tool well but as a set of constraints a maker imposes on the tool's defaults, with the tool's competence treated as the adversary rather than the asset.] The coherence comes from the keystone: if the felt thing is what is withheld, then a machine whose every default is to add and to fill is structurally the enemy of the felt thing, and the doctrine is the discipline that holds the machine open against its own grain. [PROJECTION: if the system is left to its defaults, then it produces fluent, sealed, maximally coded glut, which is the outcome the entire document is arranged to prevent.]</p>
</div>

This is also where the investigation lands on what makes the document unusual. Most operating manuals for a generative tool optimize the tool. This one is built to resist it, and it borrows a century of argument, from two traditions that never met, to justify the resistance. The throughline, the reason a pipeline for short instrumental videos reaches all the way back to Benjamin's 1931 plates and to the Japanese reading of a falling blossom, is a single conviction: that the affecting part of any image is the part no one authored, and that the most fluent image machine ever built is therefore the one that most needs to be told to leave room.

## How the doctrine becomes rules

The theory resolves into a small set of operations the system actually runs. They are listed here as the concrete residue of everything above.

<div class="kv">
  <div class="k">The frame test</div>
  <div class="v">Keep frames that imply an offscreen, carry a contingent margin, hide the figure's face, and carry time. Reject the centered, fully resolved, every-pixel-busy, clearly-emotive frame.</div>
  <div class="k">Camera as blind-field operation</div>
  <div class="v">Push-in hunts the unguarded detail; pull-back opens the hidden world; pan asserts the offscreen. Choose the move that opens the most room.</div>
  <div class="k">Ma in time</div>
  <div class="v">Let beats breathe; leave silent frames with no text; subtract rather than add. A sequence that keeps adding seals itself.</div>
  <div class="k">The Sebald gap</div>
  <div class="v">Text and frame come from one beat but never describe each other. The text points sideways into the blind field; it never names the feeling.</div>
  <div class="k">The Sontag check</div>
  <div class="v">When a hundred competent frames are available, the competent frame is worthless. Keep only the one that pierces.</div>
</div>

| Camera move | What it does to the seal | When the doctrine calls for it |
|---|---|---|
| Push-in | Approaches an unguarded detail | To hunt the contingent specific in the margin |
| Pull-back | Reveals the world the frame hid | To open a blind field the still implied |
| Pan | Follows the eye off the edge | To assert the offscreen and the continuing world |
| Hold (no move) | Lets the interval stand | To honor ma; the silent, load-bearing beat |

## The caution the doctrine ends on

The document closes by warning against itself, and the warning belongs in any honest account of it. A beautiful vocabulary tempts its user to file feelings away under their names, to say "that is so yūgen" and stop looking, which is the opposite of what the names are for. [CHARACTERIZATION: the doctrine treats its own terms as door handles rather than the room, and holds that the feeling came before the words and is the real possession, the words only a way back to it.] Barthes' own late turn was a rebellion against his earlier instinct to code everything into neat terms, and the doctrine inherits the rebellion: the names are there to sharpen attention, never to substitute for it. When something pierces, that is where the thinking starts, not where it gets to rest.

> Render the felt presence of what is not given. Leave room.

## Sources

- [Roland Barthes, *Camera Lucida: Reflections on Photography* (1980, trans. Richard Howard)](https://en.wikipedia.org/wiki/Camera_Lucida_(book)) — the source of studium, punctum, that-has-been, the noeme, and the sealed image versus the blind field.
- [Roland Barthes, *Mythologies* (1957)](https://en.wikipedia.org/wiki/Mythologies_(book)) and [*Image-Music-Text* (1977, trans. Stephen Heath)](https://en.wikipedia.org/wiki/Image-Music-Text) — the earlier semiotic reading of images that *Camera Lucida* turns against.
- [Walter Benjamin, "Little History of Photography" (1931)](https://en.wikipedia.org/wiki/Little_History_of_Photography) — the spark of contingency and the optical unconscious.
- [Susan Sontag, *On Photography* (1977)](https://en.wikipedia.org/wiki/On_Photography) — photography as appropriation, the glut of images, and the anesthesia of seeing.
- [John Berger, *Ways of Seeing* (1972)](https://en.wikipedia.org/wiki/Ways_of_Seeing) and *About Looking* (1980) — the photograph as a quotation from appearances.
- [Margaret Olin, "Touching Photographs: Roland Barthes's 'Mistaken' Identification" (2002)](https://en.wikipedia.org/wiki/Camera_Lucida_(book)#Criticism) — the necklace discrepancy and the viewer-supplied punctum.
- [Michael Fried, *Why Photography Matters as Art as Never Before* (2008)](https://en.wikipedia.org/wiki/Michael_Fried) — absorption and the argument that the photograph's structure earns the punctum.
- [Gaston Bachelard, *The Poetics of Space* (1957)](https://en.wikipedia.org/wiki/The_Poetics_of_Space) — the poetic image as the origin of reverie, and intimate immensity.
- [W. G. Sebald, on the uncaptioned photographs in his prose](https://en.wikipedia.org/wiki/W._G._Sebald) — text and image held in separate rooms.
- [Brian Eno, *Music for Airports* (1978) and generative music](https://en.wikipedia.org/wiki/Generative_music) — emergent beauty from rules set in motion.
- [John Koenig, *The Dictionary of Obscure Sorrows*](https://www.dictionaryofobscuresorrows.com/) — the definition of sonder.
- [Ma, yūgen, mono no aware, and wabi-sabi in Japanese aesthetics](https://en.wikipedia.org/wiki/Japanese_aesthetics) — the parallel vocabulary of the charged interval, profound suggestion, the pathos of passing things, and the beauty of the imperfect.

---

# How a Local Business Gets Found Now: Search, AI Answers, and the Shift From Links to Citations
*A walk through the machinery of local discovery in 2026 - the Map Pack and classic ranking, the structured data that makes a site machine-readable, Generative Engine Optimization and the llms.txt that the engines ignore - and how one family-owned detailing shop was rebuilt to be found by both Google and the models. Where the link economy is giving way to a citation economy, and what stays the same underneath.*

Source: https://theholbrookreport.com/reports/local-discovery-in-the-age-of-ai/ · Published: 2026-06-05

<!--
AUTHOR NOTE GOES HERE (optional). Aaron writes this in his own first-person voice
if he wants a personal introduction. The AI does not write this block. Pattern:

<div class="author-note">
  ...your introduction...
  <span class="byline">- Aaron Holbrook</span>
</div>
-->

A person in a Midwestern suburb wants their car detailed. Ten years ago they would have typed "auto detailing near me" into Google, skimmed a page of blue links, and clicked one. Today the same person has at least three different ways to ask the same question, and only one of them reliably ends in a click on a business's website. They might still scroll the links. They might read the boxed summary Google now writes at the top of the page and never scroll at all. Or they might skip Google and ask an AI assistant to "find me a good detailer nearby that does ceramic coating," and act on whatever names it returns.

These are three different machines with three different rules. A local business that is built for the first one is not automatically visible in the other two. This report is about how all three actually work in 2026, what is changing between them, and what a small business can do about it. It uses a single example throughout: an anonymized Holbrook Solutions client engagement, a family-owned auto detailing shop in an outer Midwestern suburb whose website was rebuilt from the ground up for exactly this transition, referred to here as "the detailer." The conflict is disclosed above; the numbers from that engagement are labeled as first-party measurements wherever they appear.

The short version: the fundamentals of being a findable local business have not changed, but the surface they are read through has. The web is shifting from an economy of *links you click* to an economy of *facts that get cited*, and the businesses that win the second one are the ones whose facts are consistent, structured, verifiable, and genuinely better than the alternatives.

---

# Movement 1 - How local discovery works in 2026

## The three surfaces

When someone looks for a local service today, the answer can come from three places, and they behave very differently.

<div class="kv">
  <div class="k">Classic results</div><div class="v">The ranked web links plus the local "Map Pack" of three businesses with a map. Driven by Google's local ranking system and your Google Business Profile.</div>
  <div class="k">The AI answer</div><div class="v">A generated summary at the top of the results (Google's AI Overviews), or an answer from an assistant like ChatGPT, Perplexity, or Claude. It paraphrases and cites sources rather than listing them.</div>
  <div class="k">The model's memory</div><div class="v">What an AI model "already knows" about your business from its training data, with no live lookup at all. This is the surface businesses understand least and control least directly.</div>
</div>

Classic results are the surface every local-marketing playbook was written for. The other two are newer, and the rest of this report is largely about them. But they are not replacements stacked in sequence so much as layers competing for the same moment of intent, and the same underlying facts feed all three.

## Classic local ranking: relevance, distance, prominence

Google states plainly what determines local ranking. There are three named factors: relevance ("how well a Business Profile matches what someone is searching for"), distance ("how far each business is from the customer who's searching"), and prominence ("how well-known a business is"). [FACT: google-local-ranking] Google adds that prominence includes "how many websites link to your business and how many reviews you have," and that "more reviews and positive ratings can help your business's local ranking." [FACT: google-local-ranking]

Two of those three are not things you can write on a webpage. Distance is set by where the searcher is standing. Prominence is built off-site, out of links, citations, and reviews accumulated over time. Only relevance is fully inside your control, which is why so much local SEO effort goes into the one asset that controls it: the <span class="term" data-def="The free business listing that powers the Map Pack and the knowledge panel: name, address, phone, hours, categories, photos, reviews, and posts.">Google Business Profile</span>. The profile, not the website, is what populates the Map Pack, and the Map Pack is what sits above the organic links for almost every "near me" query.

The supporting cast is the set of off-site signals that feed prominence:

- **Reviews.** Quantity, recency, and average rating all feed local ranking, and Google explicitly says responding to reviews helps too. [FACT: google-local-ranking] Reviews are also the single biggest consumer-facing signal, which the next movement returns to.
- **<span class="term" data-def="Name, Address, Phone number. The canonical contact identity of a business as it appears across the web.">NAP</span> consistency.** The business name, address, and phone number need to read identically across the website, the Google profile, Yelp, Apple Maps, directories, and any other listing. Inconsistent contact data is one of the most common, and most fixable, local-ranking drags.
- **Citations and links.** Mentions of the business on other reputable local sites, directories, and community pages.

This is the part that has not changed and is not going to. A business with a complete profile, real reviews, and consistent contact data across the web has always done better in local search, and still does.

## Making the site machine-readable: structured data

A web page is written for humans. <span class="term" data-def="A vocabulary of tags (schema.org) embedded in a page, usually as JSON-LD, that states facts about the page in a format machines can parse without guessing.">Structured data</span> is the same page's facts restated in a format a machine can read without interpreting prose. For a local business this usually means a `LocalBusiness` block in <span class="term" data-def="JavaScript Object Notation for Linked Data: a script tag in the page head that carries schema.org facts as structured key-value data.">JSON-LD</span>: name, address, geo-coordinates, phone, hours, services, area served, and links to the business's other profiles.

The detailer's site carries a full stack of this. The home page declares an `AutoDetailing` business (a `LocalBusiness` subtype) with address, coordinates, opening hours for every day of the week, a list of more than twenty served towns, the owner as a named `Person`, an `EducationalOccupationalCredential` for the owner's industry certification, and an `OfferCatalog` of every service. Each service page adds a `Service` block tied to a `GeoCircle` around the shop. A `sameAs` array links the entity to its Google, Yelp, and Facebook profiles. [FACT: engagement] Here is the spine of it, with identifying values replaced by placeholders:

```json
{
  "@context": "https://schema.org",
  "@type": "AutoDetailing",
  "name": "Example Auto Detailing",
  "telephone": "+1-555-010-0000",
  "address": {
    "@type": "PostalAddress",
    "streetAddress": "100 Example Street",
    "addressLocality": "Anytown",
    "addressRegion": "IL",
    "postalCode": "60000"
  },
  "geo": { "@type": "GeoCoordinates", "latitude": 41.000000, "longitude": -88.000000 },
  "hasCredential": {
    "@type": "EducationalOccupationalCredential",
    "name": "Certified Detailer",
    "recognizedBy": { "@type": "Organization", "name": "Industry Detailing Association" }
  },
  "sameAs": [
    "https://www.google.com/maps?cid=000000000000000000",
    "https://www.yelp.com/biz/example-auto-detailing",
    "https://www.facebook.com/example-auto-detailing"
  ]
}
```

Structured data does two jobs. It makes the page eligible for richer search results (the star ratings, hours, and FAQ accordions that show up under some listings), and it gives any machine reading the page an unambiguous statement of the facts instead of forcing it to parse them out of sentences. The `sameAs` array is quietly the most important line: it tells a machine that this website, that Google listing, that Yelp page, and that Facebook profile are all the *same entity*. That is the thread the AI layer pulls on when it tries to decide who you are.

There is one hard limit worth stating now, because it shapes the whole strategy. You cannot mark up your own star rating and expect it to show. Google's policy is explicit: "if the entity that's being reviewed controls the reviews about itself," its pages "are ineligible for star review feature," and "ratings must be sourced directly from users." [FACT: google-review-snippet] Self-supplied ratings are not just ignored; they make the page ineligible. Genuine, third-party-sourced reviews are the only ones that count. This is a recurring theme: the system is designed to reward real signals and discard manufactured ones.

## Speed is part of the product

Google measures three <span class="term" data-def="Core Web Vitals: Google's standard metrics for loading speed, responsiveness, and visual stability of a page.">Core Web Vitals</span>, and as of 2024 they are Largest Contentful Paint (loading, target 2.5 seconds or less), Interaction to Next Paint (responsiveness, target 200 milliseconds or less, which replaced First Input Delay in 2024), and Cumulative Layout Shift (visual stability, target 0.1 or less). [FACT: webdev-vitals] They are a real input to ranking, and a much more direct input to whether a visitor stays.

This is where the detailer's rebuild started, and the gap was dramatic. The previous site, built on WordPress, took 23.2 seconds to render its main content on a phone and scored 57 of 100 on mobile performance and 58 on SEO in Google's Lighthouse audit. The rebuilt site, a static <span class="term" data-def="A site framework that ships pre-built HTML with little or no JavaScript, so pages load almost instantly.">Astro</span> site served from Cloudflare's edge, renders its main content in 3.6 seconds and scores 82 on performance, 100 on accessibility, and 100 on SEO. [FACT: engagement] That is the same business, the same photos, the same services, made roughly six times faster to first paint.

<div class="callout">
  <span class="label">What this really means in practice:</span> Performance is not a vanity metric for a local business. It is the difference between a phone user who waits and one who taps back to the results and calls the next shop on the list. A slow site loses customers before ranking is even relevant.
</div>

## GEO: being inside the answer, not under it

Classic SEO tries to rank your link. <span class="term" data-def="Generative Engine Optimization: shaping content so it gets cited or quoted inside AI-generated answers, rather than ranked as a link.">Generative Engine Optimization</span>, or GEO, tries to get your business named and cited *inside* an AI-generated answer. The term comes from a 2023 academic paper (Aggarwal and co-authors, later published at the KDD 2024 conference) that coined "GEO" and ran controlled tests on what makes a source more visible in generated answers. Their headline finding: GEO methods can boost a source's visibility in generative-engine responses by up to roughly 40%, and the most effective tactics were adding citations, adding direct quotations from relevant sources, and adding statistics. [FACT: geo-paper]

A later 2025 audit of 1,702 real citations across Brave, Google AI Overviews, and Perplexity reached a compatible conclusion from the other direction: pages that got cited tended to score high on overall quality (being cited correlated with page quality at an odds ratio of 4.2), and the strongest structural correlates were fresh metadata, clean semantic HTML, and structured data. [FACT: geo16] Both studies have real limits, noted in the sources, but they point the same way. AI engines preferentially cite content that is well-structured, specific, current, and quotable.

[CHARACTERIZATION: this is a synthesis of the two GEO studies, not a measured value from either] The practical translation is that GEO is mostly classic content quality plus machine-readability, not a separate dark art. Concrete numbers, named credentials, dated facts, and clean markup are what get quoted, which is exactly what a good local site should have anyway.

## The llms.txt that nobody reads

This is also where a popular idea needs a cold splash of water. In 2024 a proposed standard called `llms.txt` emerged: a plain-text file at the root of a site giving AI models a clean, summarized version of the business's key facts, by analogy to `robots.txt`. It is a genuinely sensible idea, and the detailer's site publishes both an `llms.txt` and a longer `llms-full.txt`. [FACT: engagement]

The problem is that, so far, the major AI engines do not appear to use it. A 90-day server-log experiment found that of more than 62,100 AI-bot visits, only 84 (about 0.1%) ever requested the `/llms.txt` file, fewer than hit an average content page. [FACT: otterly-llmstxt] Google has said publicly that it does not use `llms.txt`, with one Google representative comparing it to the long-dead keywords meta tag. [FACT: sej-llmstxt] A separate analysis across hundreds of thousands of domains found no clear effect on AI citations. [FACT: sej-llmstxt-300k]

So why keep the file? [CHARACTERIZATION: a judgment about cost versus optionality, not a measured benefit] Because it costs almost nothing to maintain, it cannot hurt, and it is a cheap option on a standard that *might* get adopted. But it should be understood as a bet on the future, not a working channel today. The thing that actually makes a business legible to AI is not a special file; it is consistent facts everywhere a machine can read them.

## How an AI decides what to say about a business

To see why consistency matters so much, it helps to know where an AI answer comes from. There are three sources, and they blend:

<div class="kv">
  <div class="k">Training data</div><div class="v">What the model absorbed when it was built. Frozen, often months old, and impossible to edit directly. This is the "model's memory" surface.</div>
  <div class="k">Retrieval / grounding</div><div class="v">A live web search the model runs at question time, reading current pages and quoting them. This is how ChatGPT search, Perplexity, and Google AI Overviews stay current.</div>
  <div class="k">Third-party sources</div><div class="v">What other sites - directories, review platforms, local pages - say about the business. The model cross-checks these against your own claims.</div>
</div>

A business cannot edit a model's training data. What it can do is make sure that when the model *does* look, every source agrees. If the website says the shop has more than 200 hours of training and an old directory says 300, the model has to choose, and an uncertain model tends to hedge or omit. [CHARACTERIZATION: a mechanism described in the GEO literature and consistent with how grounded models resolve conflicts, not a quantified rate] Entity consistency, the same name, address, phone, hours, and credentials everywhere, is the closest thing there is to writing directly into the machine's understanding of you.

The crawlers themselves are now specialized, and the controls are per-purpose. OpenAI alone runs at least three: GPTBot gathers training data (blocking it is a training opt-out), OAI-SearchBot surfaces sites inside ChatGPT's search answers (blocking it removes you from those answers), and ChatGPT-User fetches a page when a user asks ChatGPT to look at it. [FACT: openai-bots] Google splits the function too: Google-Extended controls whether your content trains Gemini, and Google states it "does not impact a site's inclusion in Google Search nor is it used as a ranking signal." [FACT: google-extended] Anthropic, Perplexity, Apple, Amazon, Meta, and others each run their own. The strategic point is that blocking a training crawler and blocking a search crawler are completely different decisions: one affects whether a model learns about you in general, the other affects whether you can appear in that product's live answers. The detailer's `robots.txt` makes the choice explicit, welcoming the search and assistant crawlers by name. [FACT: engagement]

---

# Movement 2 - The shift from links to answers

## The click is leaking out of search

The mechanics above describe a stable world. The reason this report exists is that the world is moving, and the movement is measurable.

Google's <span class="term" data-def="AI Overviews: the AI-generated summary Google places at the top of many search results pages, above the classic links.">AI Overviews</span> went from a rollout to a fixture in 2025. One large keyword study tracked them appearing on 6.49% of queries in January 2025, peaking at 24.61% in July, then settling back to 15.69% by November. [FACT: semrush-aio] An independent Pew Research analysis of real browsing data found that in March 2025, about 18% of all Google searches by US adults produced an AI summary. [FACT: pew] The exact number depends on the method, but the order of magnitude is clear: a large and volatile share of searches now show a generated answer before any link.

And when that answer appears, people click less. Pew found that when an AI summary was present, only 8% of searches led to a click on a traditional link, versus 15% when there was no summary, and just 1% of users clicked a link *inside* the summary. [FACT: pew] Ahrefs, measuring position-one click-through over time, found the presence of an AI Overview correlated with a 58% reduction in clicks to the top organic result by December 2025, up from a 34.5% reduction in April. [FACT: ahrefs-aio] Google has disputed the framing of some of these studies, but has not offered contradicting numbers. [CHARACTERIZATION: a note on the state of the dispute, characterizing the absence of a counter-figure, not endorsing either side] The direction is not seriously in question.

A caveat keeps this honest. One study found that on the *same* set of keywords, the appearance of AI Overviews did not automatically raise the zero-click rate (it measured 33.75% before versus 31.53% after). [FACT: semrush-aio] That conflicts with broader market measures showing zero-click behavior rising overall. The two are measuring different populations, and the safe reading is that the link economy is shrinking unevenly, not uniformly. [CHARACTERIZATION: a reconciliation of two studies with different scopes, offered as interpretation] The trend is real; any single percentage is a snapshot of a moving target.

## The discovery channel itself is changing hands

The deeper shift is not just inside Google. It is that people are increasingly not starting at Google at all.

The clearest signal comes from BrightLocal's 2026 consumer survey: the share of consumers using ChatGPT and other generative AI tools to find local business recommendations jumped from 6% a year earlier to 45%, while Google's share of that same job fell from 83% to 71%. [FACT: brightlocal-lcrs] Those are enormous moves for a single year, from one vendor's survey, so they deserve the usual caution. [CHARACTERIZATION: a caveat about a single vendor survey, not a dismissal of the figure] But even discounted heavily, they describe a market where "ask an AI" has gone from a novelty to a mainstream way of finding a plumber, a dentist, or a detailer.

<div class="callout">
  <span class="label">What this really means in practice:</span> A local business now has to be findable in a conversation it never sees. When a customer asks an assistant "who's a good detailer nearby," there is no results page to rank on and no ad to buy. There is only whether the model knows you, trusts the facts about you, and names you.
</div>

## Why being the cited entity beats ranking #4

Put the two shifts together and the strategic conclusion follows. In a page of ten blue links, the difference between ranking fourth and seventh still bought you some clicks. In an AI answer, there is no fourth and seventh. There is the handful of businesses the model names, and everyone else. [CHARACTERIZATION: a structural contrast between ranked lists and generated answers, framing rather than a measured cliff] The distribution gets more winner-take-most, and the prize changes from "rank higher" to "be one of the cited entities."

What earns a citation, per the research in Movement 1, is being the clearest, most consistent, best-structured, most genuinely useful source about your specific niche. [FACT: geo-paper] That is good news for a small business that does its homework, because the bar is specificity and trustworthiness, not ad budget.

## The case study: rebuilding one detailer for both worlds

The detailer is a family-owned shop in an outer Midwestern suburb, run day to day by a certified detailer with a documented number of training hours. Roughly 70% of its customers come from one place: a large retirement community next door. [FACT: engagement] That single fact ended up driving the whole strategy.

**The starting position.** A baseline captured the day the new site launched in June 2026 measured the problem precisely. Of nine target local queries, the business's own domain ranked on page one for zero of them. Only three URLs were indexed by Google, and all three were stale addresses from the old WordPress site. The business surfaced in search mainly through its Yelp listing, not its own site. [FACT: engagement] On a simulated AI-discoverability test (could an assistant correctly answer what they do, where they are, their hours, their credentials, and whether to recommend them), the site scored 9 out of 10. [FACT: engagement] The single missing point was the one thing the business could not mark up itself: on-site review and rating data, which by Google's own policy must come from real users. [FACT: google-review-snippet]

**The rebuild.** The technical work is the catalog from Movement 1, applied: the WordPress-to-Astro rebuild that cut load time from 23.2 to 3.6 seconds and lifted the SEO score from 58 to 100; the full `LocalBusiness` structured-data stack; per-page titles, meta descriptions, canonical tags, and Open Graph; an XML sitemap; 301 redirects from the old WordPress URLs to preserve whatever equity the three indexed pages still carried; `llms.txt` and `llms-full.txt`; and a `robots.txt` that names and welcomes the AI crawlers. [FACT: engagement]

**The niche nobody owned.** The highest-leverage discovery was not technical. The retirement community next door shares its name with a much larger, far more famous retirement community in another state, so searching the obvious term (the community's name plus "auto detailing") returned results for the wrong place entirely. Nobody had claimed the term locally, despite a community of thousands of homes that already sent the shop most of its business. [FACT: engagement] The rebuild added a dedicated page for that community, internally linked and structured, targeting a high-intent local term with effectively zero competition. [CHARACTERIZATION: "zero competition" describes the observed search result, where no local rival ranked for the term] This is the local-SEO equivalent of finding an unlocked door: a specific, real, defensible niche that maps exactly to who the business already serves.

**What is still open.** The honest part of the story is that the most valuable remaining work is not in the code. It is owner-led: getting verified in Google Search Console so the new pages get crawled and indexed; collecting the real Google and Yelp star ratings and review counts so a genuine `aggregateRating` can be published (the one gap in the discoverability score); soliciting reviews that mention the town and the specific service; and cleaning up the contact data and a training-hours figure that disagreed between the site and an old third-party listing. [FACT: engagement] The site is built; the off-site signals are earned over months.

## Where SEO and GEO turn out to be the same thing

The review gap is the tell. The thing that would most improve the detailer's classic local ranking (more real reviews and a higher rating) is the same thing that would close its last AI-discoverability point (structured, genuine `aggregateRating` data), is the same thing Google's policy says must be real to count, and is the same thing a human customer looks at first. [FACT: google-local-ranking] One asset, four payoffs. [CHARACTERIZATION: an observation that the same signal serves classic ranking, AI citation, policy eligibility, and human trust at once]

This is the quiet thesis of the whole transition. The channels look like they are diverging, but the underlying signals are converging. Real reviews, consistent facts, genuine specificity, and a fast, machine-readable site pay off in classic search, in AI answers, and with the actual human being deciding whether to call. The gimmicks (self-supplied ratings, an `llms.txt` the engines ignore, thin pages spun up for every nearby town) pay off in none of them.

---

# Movement 3 - The new future of local marketing

## The website becomes a fact-source, not a brochure

For two decades a local business website was a brochure: a pretty front for humans, with the real conversions happening on the phone. The shift described above slowly inverts that. Increasingly the most important reader of the site is a machine, deciding what to tell a human who will never visit the page. [CHARACTERIZATION: a reframing of the site's primary audience, extrapolated from the retrieval mechanics, not a measured change in traffic mix]

That does not make design irrelevant; a human still lands on the site and decides whether to trust it. But it adds a second audience with different needs. The machine does not care about the hero image. It cares whether the address in the page text matches the address in the JSON-LD matches the address on Google matches the address on Yelp. The future-proof site serves both: persuasive for the person, unambiguous for the parser.

## Generic content depreciates; specific facts appreciate

If AI answers can summarize the generic, then generic content stops being an asset. A page that says "we provide quality auto detailing with great customer service" is exactly what a model can generate for free, about anyone. [CHARACTERIZATION: a value judgment about commodity content, grounded in what generative models can already produce] What a model cannot generate is the specific, verifiable, local truth: that this shop sits a block from a particular local landmark, that the owner holds an industry certification with a stated number of training hours, that it serves the retirement community next door, that ceramic coating starts at a particular price for a particular vehicle size.

[PROJECTION: conditional on AI answers continuing to absorb generic informational content, which current AI Overview trends suggest but do not guarantee] If that trend holds, the content that retains value is the content a generative model cannot fabricate without you: concrete prices, real credentials, dated specifics, named service areas, and genuine reviews. The thin "SEO content" of the last decade depreciates; the specific facts appreciate.

## The durable moats

Three things survive every version of this shift, and they are where a local business should spend its scarce effort.

<div class="finding">
  <div class="finding-head">
    <div class="name">Genuine reviews and ratings</div>
    <span class="pill high">Central mechanism</span>
  </div>
  <p>Reviews feed classic local ranking, are required (and must be real) for rating rich results, are a top correlate of AI citation, and are the first thing a human checks. As of 2026, 97% of consumers read reviews for local businesses, 68% will only use a business rated 4 stars or higher (up from 55% the year before), and 31% now require 4.5 or higher (up from 17%). The threshold is rising; the asset is irreplaceable. [FACT: brightlocal-lcrs] [FACT: google-local-ranking]</p>
</div>

<div class="finding">
  <div class="finding-head">
    <div class="name">Entity consistency across the web</div>
    <span class="pill high">Central mechanism</span>
  </div>
  <p>The same name, address, phone, hours, and credentials, identical everywhere a machine can read them: the site's visible text, its structured data, its <code>sameAs</code> links, and every third-party listing. This is how a model disambiguates you and decides to trust the facts enough to repeat them. Inconsistency is the most common reason an AI hedges or omits a business. [CHARACTERIZATION: identifies consistency as the primary controllable input to AI trust, drawn from the grounding mechanics in Movement 1]</p>
</div>

<div class="finding">
  <div class="finding-head">
    <div class="name">A real, defensible niche</div>
    <span class="pill med">Important context</span>
  </div>
  <p>Owning a specific, true position ("the detailer the local retirement community uses") beats competing for a generic head term. It is cheaper to win, harder for a national chain to contest, and exactly the kind of specific fact an AI answer can attach to a recommendation. [CHARACTERIZATION: a strategic claim that specificity outcompetes generic terms for small local businesses, illustrated by the retirement-community case]</p>
</div>

## What stays the same

Underneath all of it, the job has not changed. A local business has always won by being genuinely good, being known in its community, and being easy to find and contact. [CHARACTERIZATION: a restatement of local-business fundamentals, offered as the stable core beneath the changing surface] The machinery on top has changed three times now (directories, then search, now AI answers), and each time the businesses that adapted fastest were the ones whose underlying reputation was real enough to survive the new measurement.

<div class="callout">
  <span class="label">What this really means in practice:</span> The work that pays off in 2026 is the same work that paid off in 2016, made machine-readable. Be excellent, collect real reviews, keep your facts identical everywhere, claim a specific niche, and make the site fast and structured. Do that and you are findable in the link economy and the citation economy at once.
</div>

## A practical sequence

For an operator who wants the order of operations, the engagement described here suggests a priority list that generalizes:

1. **Fix the foundation.** A fast, structured, mobile-first site with complete `LocalBusiness` JSON-LD, canonical tags, a sitemap, and 301s from any old URLs. This is table stakes for all three surfaces. [FACT: webdev-vitals]
2. **Get crawled and indexed.** Verify the site in Google Search Console and Bing Webmaster Tools, submit the sitemap, and request indexing. Bing's index also feeds some AI search, so this is a GEO step, not just an SEO one.
3. **Win the off-site signals.** A complete Google Business Profile, real reviews solicited from happy customers (mentioning the town and the service), and consistent contact data across every listing. This is the slow, compounding work, and it is owner-led. [FACT: google-local-ranking]
4. **Claim the niche.** Find the specific, true, low-competition position the business already occupies, and build a genuine page for it. [CHARACTERIZATION: generalizes the niche tactic into a repeatable step]
5. **Keep the facts identical, and measure.** Audit name, address, phone, hours, and credentials across every surface, and re-check rankings and AI answers on a fixed schedule so drift is caught early.

The order matters because the steps build on each other: a fast structured site that no one has indexed is invisible, and a perfectly indexed site with inconsistent facts gets hedged by the very AI answers it is trying to win. [CHARACTERIZATION: explains the dependency ordering, reasoning rather than a measured result]

## The honest uncertainty

It would be a mistake to end with false precision. AI Overview prevalence rose to about 25% of queries and then fell back toward 16% within the same year. [FACT: semrush-aio] Crawler policies are changing month to month. The headline consumer figures come from single surveys. [CHARACTERIZATION: a caveat on the volatility of every figure in this report, restating the limits noted in the sources] Anyone who tells a local business they know exactly where AI search lands in two years is guessing.

[PROJECTION: conditional on AI-mediated discovery continuing to grow even if its exact form keeps shifting] What is safe to act on is the direction, not the date. If discovery keeps moving toward generated answers, the businesses positioned to win are the ones whose facts are real, consistent, structured, and specific, because those are the only inputs every version of the future rewards. The surface will keep changing. The reason a customer trusts a local business, and the signals that prove that trust to a machine, are changing far more slowly.

---

## Sources

**Classic local ranking and reviews**
- {google-local-ranking} [Google Business Profile Help - Improve your local ranking on Google](https://support.google.com/business/answer/7091) - the three official factors (relevance, distance, prominence) and that reviews and responses help local ranking
- {brightlocal-lcrs} [BrightLocal - Local Consumer Review Survey 2026](https://www.brightlocal.com/research/local-consumer-review-survey/) - 97% read reviews; 68% require 4+ stars (up from 55%); 31% require 4.5+ (up from 17%); AI use for local recommendations 6% to 45%; Google's share 83% to 71%
- {brightlocal-lcrs-2025} [BrightLocal - Local Consumer Review Survey 2025](https://www.brightlocal.com/research/local-consumer-review-survey-2025/) - prior-year baseline for the review-threshold and AI-use comparisons

**Structured data and rich results**
- {google-review-snippet} [Google Search Central - Review snippet (structured data) documentation](https://developers.google.com/search/docs/appearance/structured-data/review-snippet) - self-controlled reviews make a page ineligible for the star feature; ratings must be sourced directly from users, not curated by editors

**AI Overviews, zero-click, and click-through**
- {semrush-aio} [Semrush - AI Overviews study (2025)](https://www.semrush.com/blog/semrush-ai-overviews-study/) - AIO trigger rate 6.49% (Jan) to 24.61% (Jul) to 15.69% (Nov) 2025; same-keyword zero-click 33.75% to 31.53%
- {pew} [Pew Research Center - Google users are less likely to click links when an AI summary appears (Jul 2025)](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/) - 18% of US-adult searches produced an AI summary (Mar 2025); 8% click a link with a summary vs 15% without; 1% click within the summary
- {ahrefs-aio} [Ahrefs - AI Overviews reduce clicks (update)](https://ahrefs.com/blog/ai-overviews-reduce-clicks-update/) - position-1 CTR reduction of 58% for AIO queries by Dec 2025, up from 34.5% in April 2025

**GEO and llms.txt**
- {geo-paper} [Aggarwal et al. - GEO: Generative Engine Optimization (arXiv 2311.09735, KDD 2024)](https://arxiv.org/abs/2311.09735) - GEO can boost visibility up to ~40%; citations, quotations, and statistics are the most effective methods
- {geo16} [Kumar & Palkhouski - GEO-16 citation audit (arXiv 2509.10762, 2025)](https://arxiv.org/pdf/2509.10762) - 1,702 citations across Brave, Google AIO, Perplexity; page quality odds ratio 4.2; metadata/freshness, semantic HTML, and structured data the strongest correlates (observational preprint, B2B SaaS scope)
- {otterly-llmstxt} [OtterlyAI - The llms.txt experiment](https://otterly.ai/blog/the-llms-txt-experiment/) - only 84 of 62,100+ AI-bot visits (0.1%) requested /llms.txt over 90 days
- {sej-llmstxt} [Search Engine Journal - Google says llms.txt comparable to the keywords meta tag](https://www.searchenginejournal.com/google-says-llms-txt-comparable-to-keywords-meta-tag/544804/) - Google representatives state Google does not use llms.txt
- {sej-llmstxt-300k} [Search Engine Journal - llms.txt shows no clear effect on AI citations across 300k domains](https://www.searchenginejournal.com/llms-txt-shows-no-clear-effect-on-ai-citations-based-on-300k-domains/561542/) - large-scale corroboration of no measurable citation effect

**Crawlers**
- {openai-bots} [OpenAI - Bots and crawlers documentation](https://developers.openai.com/api/docs/bots) - GPTBot (training), OAI-SearchBot (ChatGPT search surfacing), ChatGPT-User (user-triggered fetch) are distinct with different robots.txt behavior
- {google-extended} [Google Search Central - Google crawlers and Google-Extended](https://developers.google.com/search/docs/crawling-indexing/google-common-crawlers) - Google-Extended controls Gemini/Vertex training and does not affect Search inclusion or ranking
- {almcorp-claude} [ALM Corp - Anthropic Claude bots and robots.txt strategy](https://almcorp.com/blog/anthropic-claude-bots-robots-txt-strategy/) - roles of ClaudeBot, anthropic-ai, and Claude-Web

**Performance**
- {webdev-vitals} [web.dev - Web Vitals](https://web.dev/articles/vitals) - LCP 2.5s, INP 200ms (replaced FID in 2024), CLS 0.1 as the "good" thresholds at the 75th percentile

**Case study**
- {engagement} Anonymized Holbrook Solutions client engagement (a family-owned auto detailing shop). First-party figures (Lighthouse scores, LCP, the indexing and ranking baseline, the AI-discoverability score, and the niche strategy) are from the engagement's internal SEO baseline dated 2026-06-03. The business name, domain, address, and contact details are withheld; the structured-data example above uses placeholder values.

---

# Understanding IPOs: How They Work and What It Means to Invest in One
*A definitional walk through initial public offerings — what they are, why a company does one, how the process runs from S-1 to opening trade, the two prices and the first-day pop, how to read the headline numbers, what breaks a deal, lockups, and the access rules that decide whether an ordinary investor can participate at all.*

Source: https://theholbrookreport.com/reports/understanding-ipos/ · Published: 2026-06-01

An <span class="term" data-def="The first sale of a private company's stock to the public, after which the shares trade on an exchange.">initial public offering</span> (IPO) is the first sale of a private company's stock to the public, after which the shares trade on an exchange. The mechanics of that event are consistent across deals, and they determine both what a buyer is actually purchasing and which price that buyer can reach. The sections below build from the ground up: what an IPO is and why a company does one, how the process runs, the two prices that exist on the first day, how to read the headline numbers, what causes a deal to break, the lockup calendar, the access rules that govern whether an ordinary investor can participate, and the alternative routes a company can take to the public market.

Every claim is labeled as [FACT] (a sourced data point), [CHARACTERIZATION] (a descriptive label placed on those facts), or [PROJECTION] (a conditional "if X then Y" statement, never a prediction). The tags are interactive: a fact links to its source, and a characterization or projection gives its specific reasoning on hover or keyboard focus. The severity pills mark how central a mechanism is to understanding the offering — <span class="pill high">Central mechanism</span>, <span class="pill med">Important context</span>, <span class="pill ok">Counterexample</span> — not how dangerous it is.

## What an IPO is and why a company does one

A company is private when its shares are held by a closed set of owners — founders, employees, and the funds that backed it through successive private rounds — and there is no continuous public market in which to buy or sell them. An IPO converts the company to public status: a tranche of shares is sold to outside investors and the stock begins trading on an exchange, where its price is set continuously by supply and demand.

Companies pursue this for a small set of reasons. The offering can raise fresh capital when newly issued shares are sold, the cash going to the company itself. It provides <span class="term" data-def="A way for existing shareholders to convert their holdings into cash by selling into a public market.">liquidity</span> for early investors and employees, whose previously unsellable shares become tradable. It establishes a public, market-tested valuation rather than a privately negotiated one. And it gives the company publicly traded stock to use as an acquisition currency and a recruiting tool. The flip side is the burden of being public: audited quarterly reporting, regulatory scrutiny, and a share price that reacts to every disclosure.

For an outside investor, the relevant consequence is that the IPO is the first moment shares can be bought at all — and, as the rest of this piece develops, the terms of that first moment are set by the company and its bankers, not by the buyer.

## How the process works

<div class="finding">
  <div class="finding-head">
    <div class="name">From registration to opening trade</div>
    <span class="pill med">Important context</span>
  </div>
  <p>The sequence from private company to traded stock is standardized. A company files a registration statement — for a U.S. IPO this is <span class="term" data-def="The registration document a company files with the SEC before a traditional U.S. IPO; it contains audited financial statements, a business description, the share structure, and a Risk Factors section.">Form S-1</span> — with the <span class="term" data-def="The U.S. Securities and Exchange Commission, the federal agency that regulates securities markets and reviews IPO filings.">SEC</span>, which reviews it. A public S-1 with audited financials follows, then an amended S-1 carrying a proposed <span class="term" data-def="The proposed band of possible offer prices, for example $85 to $95, disclosed in an amended S-1 before pricing.">price range</span>. Management then conducts a <span class="term" data-def="A series of presentations where management pitches the offering to large institutional investors before pricing.">roadshow</span>, during which the <span class="term" data-def="The investment banks that organize the IPO, set the price, and sell the shares; they control who receives an allocation.">underwriters</span> run a <span class="term" data-def="The traditional process where underwriters collect institutional orders during the roadshow to decide the final price and who gets shares.">bookbuilding</span> process — collecting orders from institutions to gauge demand. The offering is <em>priced</em> the night before trading, the stock opens the next morning, a lockup expires roughly 90 to 180 days later, and the first public earnings reports arrive after that. [FACT: valuethemarkets]</p>
  <p>The practical division for a long-horizon buyer is that the first day is marketing and the first year of filings is information. The <span class="term" data-def="The single fixed price, set the night before trading, at which underwriters sell allocated shares.">offer price</span> is set by the underwriters and the company; an individual buyer almost never transacts at that price and instead buys in the open market once trading begins, frequently at a higher number. [CHARACTERIZATION: a statement about the typical case, not a measured rate — it follows from how allocation routes offer-price shares to institutions, but deals vary (some grant small retail allocations, and not every IPO opens higher), so it describes the central tendency rather than a certainty] There is no obligation to participate on the first day.</p>
</div>

Two earlier stops on this path deserve their own definition because press coverage frequently conflates them. The S-1 is the most information-rich document a company going public produces: audited financial statements, a business description, the share structure, and a legally required Risk Factors section. A <span class="term" data-def="An early version of the S-1 submitted privately to the SEC for review; the public cannot read it, no offering can be priced from it, and it is not a commitment to sell shares.">confidential draft S-1</span> is an earlier version submitted privately to the SEC, permitted under the 2012 JOBS Act for <span class="term" data-def="A JOBS Act category for companies under $1B in annual revenue at filing, allowed confidential filing and scaled-back disclosure.">emerging growth companies</span> (firms under $1B in annual revenue at filing). The public cannot read it, no offering can be marketed or priced from it, and it is not a commitment to sell shares; it starts the SEC's review clock and preserves the option to go public. Anthropic's 2026-06-01 step was this confidential draft. [FACT: anthropic]

<div class="callout"><span class="label">What this really means in practice:</span> Until a public S-1 is filed, there is nothing to fundamentally analyze — no audited revenue, no share count, no price. The figures circulating in the press before that point are private marks and run-rates, not the audited numbers the public filing will contain. When it arrives, that filing — not a headline — is the document in which the recurring failure causes (a rich multiple, insider supply, control structure, related-party revenue) actually become visible.</div>

## The two prices and the first-day "pop"

<div class="finding">
  <div class="finding-head">
    <div class="name">The first-day open is, on average, the most expensive entry</div>
    <span class="pill high">Central mechanism</span>
  </div>
  <p>Two prices coexist at launch. The offer price is fixed the night before trading and sold to a chosen list of buyers. The <span class="term" data-def="The first price at which the stock actually trades on the exchange, set by live supply and demand.">opening price</span> is the first price at which the stock trades the next morning, set by live supply and demand, and is frequently well above the offer price.</p>
  <p>IPOs are, on average, deliberately priced below where they open. <span class="term" data-def="Setting the offer price below what the market will pay, which creates the first-day jump.">Underpricing</span> is the practice of setting the offer price below the level demand will support; the <span class="term" data-def="The jump from the offer price to the higher opening or first-day closing price.">"pop"</span> is the resulting jump from the offer price to the opening price. Long-running research by Professor Jay Ritter (University of Florida) puts the average U.S. first-day return at roughly 17 to 18 percent from 2001 through 2023, having spiked to about 65 percent during the 1999–2000 dot-com period. [FACT: ritter]</p>
  <p>The mechanism determines who captures that return. Institutions that receive an <span class="term" data-def="The batch of shares underwriters hand out at the fixed offer price before public trading begins; it goes overwhelmingly to large institutions.">allocation</span> at the offer price hold the gain and often sell into the retail demand that drives the open higher. By the time an individual account can transact, it is frequently buying from those sellers at a price that already embeds the excitement. [CHARACTERIZATION: the research measures the average first-day jump; naming the retail buyer as the one who pays it is the author's reading of who sits on each side of that trade — a directional conclusion the literature supports, not a measurement of any specific buyer's loss]</p>
  <div class="callout"><span class="label">What this really means in practice:</span> The pop is not money handed to the person buying on the first day; it is the cost that person pays. The discount embedded in the offer price was already delivered — to allocation holders — before the first public trade. Buying the open means buying after the discount has been distributed.</div>
</div>

The cautionary cases are the offerings that "worked" spectacularly on the first day and still charged anyone who bought the open. The figures below are first-day moves measured from the offer price — the discount the open-market buyer did not receive.

| IPO | Offer price | Open / first-day close | First-day move |
|---|---|---|---|
| Snowflake (Sep 2020) | $120 | opened $245, closed ~$254 | +112% [FACT: snowflake] |
| Airbnb (Dec 2020) | $68 | opened $146, closed ~$144.71 | +112% [FACT: airbnb] |
| Visa (Mar 2008) | $44 | closed ~$56.50 | +~28% [FACT: visa-ipo] |

In each case the institutions allocated at the offer price captured the move. Snowflake and Visa were strong businesses — a buyer at Snowflake's open is reported down roughly 41 percent more than five years later [FACT: snowflake-drawdown], while Visa proved a strong long-run holding [FACT: visa-longrun] — yet the first-day open buyer paid the full markup over the offer price in all three. The point is not that these were bad companies; it is that the first-day open repeatedly charges the open-market buyer the underpricing the insiders already captured. [CHARACTERIZATION: a label drawn from the offer-to-open spreads in the table, not a claim about company quality — two of the three were strong businesses, so "charges the buyer" refers only to the first-day entry price, not to long-run returns]

<div class="callout"><span class="label">Counter-receipt:</span> A 2025-vintage study argues the classic underpricing figure may be overstated by as much as 40 percent because a small set of enthusiast buyers skews the first-day data. [FACT: globeandmail] Even under that revision the directional point holds: the open is set by the most eager marginal buyer, not the most disciplined one. [CHARACTERIZATION: the conclusion holds even if the study's ~40% downward revision is right — a smaller average pop would still be set at the margin by the most willing buyer, so the point about who sets the open price doesn't depend on the disputed magnitude]</div>

Allocation is the sorting step that explains the asymmetry. Underwriters distribute offer-price shares overwhelmingly to large institutions and favored clients; ordinary retail accounts rarely receive a meaningful allocation on a hot deal. Some brokerages run IPO-allocation programs that grant retail clients a few offer-price shares, but on in-demand deals the amounts tend toward tiny or zero. The structural consequence is that most individuals transact at the opening price, which is precisely the number the pop has already inflated.

## Reading the numbers

A handful of figures dominate IPO coverage, and each is easy to misread. Three pairs of terms separate the honest reading from the headline.

### Run-rate versus audited revenue

<div class="finding">
  <div class="finding-head">
    <div class="name">Run-rate is a pace, not a year of collected revenue</div>
    <span class="pill high">Central mechanism</span>
  </div>
  <p>A <span class="term" data-def="A recent short period of revenue multiplied out to a full year; an estimate of current pace, not audited annual revenue.">run-rate</span> takes a short recent stretch of revenue — say one month — and multiplies it out to a full year. If a company books about $3.9B in a single month, twelve times that is roughly a $47B run-rate, even though it has not collected $47B over a real year. A run-rate differs from <span class="term" data-def="The actual revenue recognized over a real 12-month period under accounting rules, checked by independent auditors.">audited annual revenue</span> in three ways: it assumes the recent pace holds for twelve straight months (which for a fast-growing company overstates the real <span class="term" data-def="The actual figure summed over the most recent real 12 months; backward-looking and concrete.">trailing-twelve-month</span> total); it is typically reported by the press or the company, not checked by independent auditors; and audited revenue is the actual money recognized over a real 12-month period under accounting rules.</p>
  <div class="callout"><span class="label">What this really means in practice:</span> When a figure such as Anthropic's press-reported "~$47B run-rate" (late May 2026) appears, read it as "running at a pace that would total $47B a year if this exact month repeated twelve times — per the press, unaudited." [FACT: yahoo-anthropic] The gap between that sentence and "$47B in revenue" is the gap worth tracking. Any valuation ratio built on a run-rate where audited revenue belongs is distorted by the same gap. [CHARACTERIZATION: reasoning from the definitions, not a sourced figure — because a revenue multiple divides value by revenue, feeding it an annualized run-rate instead of audited revenue understates the multiple by the same proportion the run-rate overstates revenue]</div>
</div>

### Post-money valuation versus market capitalization

<span class="term" data-def="A company's value immediately after a funding round, including the new cash just raised.">Post-money valuation</span> is what a company is judged to be worth immediately after a fresh private investment round closes; <span class="term" data-def="A company's value just before a funding round, excluding the incoming cash.">pre-money</span> is the value just before that cash arrives. Worked example using Anthropic's press-reported figures: a "~$965B post-money on a $65B Series H" (a Series H is simply the eighth lettered funding round) implies a pre-money of about $900B, with the new investors owning roughly $65B ÷ $965B, about 6.7 percent. [FACT: fortune-anthropic] A post-money figure is a number a handful of investors negotiated in one private deal — a <span class="term" data-def="A valuation negotiated privately with a few investors, as distinct from a price thousands of public buyers and sellers set daily.">private mark</span>, not a market-clearing public price. [CHARACTERIZATION: a distinction about how the number was set, not whether it is too high or low — a post-money figure is agreed by a few investors in one negotiation, while a market price is discovered continuously by many buyers and sellers]

The public-market cousin is <span class="term" data-def="A public company's total value: share price multiplied by shares outstanding.">market capitalization</span> — share price times shares outstanding, set continuously by the open market rather than by one round. Ten billion shares at $100 each is a $1 trillion market cap. A private mark is not a promise the public market will agree; the IPO is where a continuously tested price is discovered, and it can land above or below the private mark.

### P/E ratio versus revenue multiple

Two ways to express how expensive a stock is. The <span class="term" data-def="Share price divided by earnings per share — how many dollars you pay for one dollar of annual profit.">price-to-earnings (P/E) ratio</span> is the share price divided by per-share annual profit. An "85× P/E" means paying $85 for $1 of annual profit, which only pays off if profits grow fast — pairing "85×" with "slowing growth" was the warning in Facebook's case. The <span class="term" data-def="Company value divided by annual revenue, used when a company has little or no profit yet.">revenue multiple (price-to-sales)</span> divides company value by annual revenue and is used when a company has little or no profit, so a P/E would be negative or meaningless. A company valued at $100B with $10B of revenue trades at a 10× revenue multiple. A revenue multiple is only as honest as the revenue figure fed into it — which is why the run-rate distinction above is load-bearing.

## What breaks an IPO

<div class="finding">
  <div class="finding-head">
    <div class="name">Priced too high, or structurally flawed</div>
    <span class="pill med">Important context</span>
  </div>
  <p>The reverse of underpricing is an offering that comes public too expensive, or with structural weaknesses, and trades below its first reference price within weeks. Roughly 21 to 25 percent of recent U.S. IPOs posted negative first-day returns — they closed the first day below the offer price. [FACT: duke]</p>
</div>

| IPO | What happened | Stated cause(s) |
|---|---|---|
| Facebook (May 2012) | $38 offer; below offer within days ($34.03, then $31); under $18 by Sept 2012 | 85× P/E into slowing growth; mobile-revenue softness disclosed selectively to large investors; ~57% of IPO shares sold by insiders; a Nasdaq trading glitch on debut [FACT: facebook] |
| Lyft (Mar 2019) | opened ~$87 vs $72 offer; fell ~25% within a month | priced into decelerating growth [FACT: lyft] |
| Uber (May 2019) | $45 offer; opened $42; closed first day −7.6% | opened below offer [FACT: uber] |
| Rivian (Nov 2021) | $78 offer at ~$66.5B valuation; below $78 within weeks; reported down ~90% since | rich valuation, no path to profit [FACT: rivian] |
| Robinhood (Jul 2021) | $38 offer; dipped below offer in the first week | weak subsequent quarter [FACT: robinhood] |
| Peloton (Sep 2019) | ended first day ~11% below the IPO price | broken debut [FACT: peloton] |
| WeWork (2019) | IPO pulled entirely | heavy losses, tangled structure, weak governance, founder-control concerns [FACT: wework] |

The recurring causes are a small set: a price embedding too much future growth, heavy insider supply, governance or control problems, and no clear path to profit. [CHARACTERIZATION: a common thread found by inspecting the cases in the table, not a measured frequency — these causes recur across the documented breakages, but the piece doesn't claim a statistical rate at which each appears] None is visible from a headline valuation; all of them live in the S-1.

Two of those causes are specific structural features worth defining, because they recur and because both are disclosed in the filing rather than the press.

**Dual-class and super-voting shares.** Many founder-led technology companies go public with two share classes: the public buys <span class="term" data-def="Shares with one vote each, typically held by the public.">Class A shares</span> with one vote each, while founders hold <span class="term" data-def="Shares carrying multiple votes each (often ten), typically held by founders, concentrating control.">Class B / super-voting shares</span> with many votes each (often ten). The effect is that a public holder can own a real economic slice of the company while holding little say in how it is run. If founders hold 15 percent of the shares but those shares carry ten votes each, they can control a majority of the voting power while owning a minority of the economics. [CHARACTERIZATION: an illustrative example using round numbers (15% of shares at ten votes each) to show how super-voting mechanically separates control from economics — not any specific company's actual share split] It is not automatically adverse — it can let founders pursue a long-term plan — but it is a structural feature to price in rather than discover later. Anthropic is reported to combine founder control with a Long-Term Benefit Trust, a governance body intended to steer the company toward its mission; the share-class details are unknown until the public S-1. [FACT: yahoo-anthropic]

**Related-party revenue.** This is revenue a company books from a counterparty that is also one of its investors, owners, or affiliates. The concern is that such demand can be partly circular: an investor supplies cash, some of which returns as the investor's own spending and is recognized as revenue. If a backer invests $10B and in the same period spends $3B buying the company's product, part of the reported revenue is funded by the company's own investor and may not reflect independent customer demand. Amazon and Google are reported as both large strategic investors in Anthropic and as cloud/compute partners — exactly the relationship the S-1's <span class="term" data-def="Disclosures revealing revenue from investors or affiliates, and how dependent revenue is on a few large customers.">related-party and customer-concentration disclosures</span> exist to reveal. [FACT: fortune-stake] It does not make the revenue fake; it means independent demand should be discounted until the filing quantifies those flows. [CHARACTERIZATION: guidance on how to weigh the revenue, not a finding that any of it is circular — related-party flows may be entirely legitimate, but until the S-1 sizes them, how much of the revenue proves independent demand is unknown, so the prudent reading discounts it]

## Lockups and lockup expiry

<div class="finding">
  <div class="finding-head">
    <div class="name">A scheduled supply event roughly 90 to 180 days after listing</div>
    <span class="pill med">Important context</span>
  </div>
  <p>A <span class="term" data-def="A contractual promise by insiders not to sell shares for a set window after the IPO, most commonly 180 days.">lockup</span> is a contractual promise by insiders — founders, employees, early-investor funds — not to sell for a set window after the IPO, most commonly 180 days, sometimes staggered (for example 25 percent released at 90 days and the balance at 180). Its purpose is to keep that large pile of insider shares from flooding the <span class="term" data-def="The number of shares actually available to trade in the public market, excluding locked-up insider shares.">float</span> on the first day. When the lockup expires, that supply becomes eligible to sell, and academic evidence consistently shows small negative <span class="term" data-def="Price moves larger or smaller than the overall market would explain; used to isolate an event's effect.">abnormal returns</span> around unlock dates, commonly in the low-single-digit-percent range. [FACT: valuethemarkets]</p>
  <p>The lockup expiry is therefore a known, scheduled supply event. It is not guaranteed to produce a lower price — a strongly performing stock can absorb the supply — but it is a feature of the calendar that the first-day coverage never mentions. [CHARACTERIZATION: the unlock date is a fact; calling it a softer entry point is conditional — the added supply weighs on price only if demand doesn't absorb it, which a strong stock can, so this flags a tendency around unlocks, not a prediction that a given stock falls]</p>
</div>

## Access: who can reach which price

<div class="finding">
  <div class="finding-head">
    <div class="name">The realistic options for an ordinary investor</div>
    <span class="pill high">Central mechanism</span>
  </div>
  <p>This is the part an investor most needs and the part the coverage least explains. The blunt summary is that an ordinary individual generally cannot buy at the offer price, generally cannot reach the private markets that precede the IPO, and is left with the open market — the one place where the pop has already been paid. The concrete routes:</p>
</div>

**Offer-price allocation — usually closed.** Allocation, defined above, is why an individual usually cannot buy at the offer price; underwriters route those shares to institutions. Some brokerages run IPO-allocation programs that may grant retail clients a few offer-price shares, but on in-demand deals the amounts tend toward tiny or zero.

**Pre-IPO secondaries — gated by the accredited-investor rule.** Before an IPO, shares of a still-private company can sometimes be bought on <span class="term" data-def="Platforms such as Forge, EquityZen, and Hiive where shares of still-private companies change hands.">pre-IPO secondary markets</span> such as Forge, EquityZen, and Hiive. These are generally open only to accredited investors. The <span class="term" data-def="An SEC category determining who may buy certain private securities without the protections of a public offering.">accredited investor</span> rule determines whether those markets are open at all. As of 2026 an individual qualifies mainly by income — over $200,000 a year alone, or $300,000 jointly with a spouse, in each of the last two years with the same expected this year — or by net worth over $1 million excluding the value of a primary residence; certain license holders (Series 7, 65, or 82) also qualify. [FACT: sec-accredited] The net-worth test excludes the primary residence. [FACT: sec-networth] These thresholds have been unchanged since they were set; legislation that would add a qualification-by-exam path independent of wealth has passed the House and is not yet law, so it cannot be relied upon today. [FACT: sec-accredited] [PROJECTION: conditional on the House-passed exam-qualification bill becoming law — if it does, wealth would no longer be the only route to accredited status; since it hasn't, this is a possible future path, not something an investor can use today]

**Diluted public proxies — available to anyone, but thin.** An ordinary investor can buy a public company that itself owns a stake in the target. Owning Amazon or Google is an indirect, heavily diluted way to hold Anthropic exposure: if such a holder owns a low-double-digit-percent stake, a move in the private company's value reaches the proxy's share price only after being diluted by everything else that company does. [CHARACTERIZATION: reasoned from the reported stake size, not a computed sensitivity — a parent holding a low-double-digit-percent stake passes only that fraction of the target's value change into its own price, further muted by the rest of its business; the direction is clear, the exact pass-through is not quantified here]

**The open market after listing — the default.** For most individuals the realistic path is to buy on the exchange once trading begins, on the buyer's own timeline. Because there is no obligation to buy on the first day, a buyer can also wait — past the opening pop, and past the lockup-expiry supply event — rather than transact at the open. [CHARACTERIZATION: a description of the timing the open market leaves available, not advice to use it — since nothing compels a first-day purchase, waiting past the pop or the unlock is possible; the piece notes these options exist without recommending the reader choose them]

## Routes to going public

<div class="finding">
  <div class="finding-head">
    <div class="name">The route changes what "the first day" means</div>
    <span class="pill med">Important context</span>
  </div>
  <p>A <b>traditional IPO</b> sells newly issued shares at a fixed offer price set by underwriters, raising fresh cash; it is the default this piece mostly describes. A <span class="term" data-def="Going public by letting existing shares trade with no new shares sold and no fixed offer price.">direct listing</span> skips the underwritten sale: the company lists existing shares and lets them trade, with no fixed offer price and (classically) no new money raised. Coinbase used a direct listing in April 2021 — which is why its first-day drop is measured from the open ($381 to ~$328, −14%) rather than from an offer price. [FACT: coinbase] A <span class="term" data-def="A 'blank-check' shell company that raises public money first, then merges with a private company to take it public.">SPAC</span> is a shell that goes public first with no operating business, then merges with a private company to bring it public through the back door; WeWork reached the public market via a SPAC in 2021 after pulling its traditional IPO. [FACT: wework]</p>
  <p>Because a direct listing has no offer price and no allocation step, the offer-price-to-open gap that defines the pop does not exist in the same form. Knowing which route a deal uses tells a reader how to interpret its first-day numbers. [CHARACTERIZATION: a consequence that follows from the route definitions, not a measured effect — a direct listing has no offer price, so its first-day move is measured from the open, which means the same "−14%" means something different than it would after a traditional IPO]</p>
</div>

### The counterexample: a modified Dutch auction

<div class="finding">
  <div class="finding-head">
    <div class="name">The first-day spread is a design choice, not a law</div>
    <span class="pill ok">Counterexample</span>
  </div>
  <p>Google's 2004 IPO used a <span class="term" data-def="An IPO where investors bid and a single clearing price is found by auction rather than set by bankers.">modified Dutch auction</span>. In a Dutch auction, investors submit bids stating how many shares they want and the price they will pay; the company finds the single <span class="term" data-def="The highest price at which all offered shares can be sold; everyone who bid at or above it pays this one price.">clearing price</span>, and everyone who bid at or above it pays that one price. Because the broad market — not the underwriters' bookbuilding — sets the price, the format tends to compress the first-day pop. Google cut its range from $108–$135 to $85–$95, priced at $85, and rose a comparatively modest ~18 percent on the first day, the muted pop the format is designed to produce. ("Modified" means it retained some underwriter involvement.) [FACT: google-dutch]</p>
  <p>This is the structural counterexample: when the crowd sets the price rather than a banker's order book, more of the value stays with the company and small buyers. It is rare — most deals still use bookbuilding — but it demonstrates that the first-day spread is a feature of how most deals are structured, not an iron law. [CHARACTERIZATION: drawn from one well-documented case (Google 2004) as an illustration, not a statistic — a single auction compressing the pop shows the spread isn't inevitable, but one case can't establish how reliably auctions do so, hence "possible," not "always"]</p>
</div>

## What an informed investor understands

Put end to end, the mechanics resolve into a coherent picture rather than a list of warnings. An IPO is the engineered transition of a company from a closed set of private owners to a continuously priced public market, undertaken for capital, liquidity, and a market-tested valuation — and the terms of that transition are set by the company and its underwriters before any outside individual can act. The process is legible in advance: the S-1 is the document where the business, the share structure, and the risks become visible, and until it is public the circulating figures are private marks and run-rates, not audited results.

The first day carries two prices for a reason. The offer price embeds a deliberate discount that is delivered to allocation holders before the first trade; the opening price is where that discount has already been paid. The headline numbers each have an honest and a flattering reading — run-rate versus audited revenue, post-money mark versus market-clearing price, revenue multiple versus the quality of the revenue under it — and the difference is knowable, not mysterious. The deals that break do so for a short, repeating list of reasons, every one of which is disclosed in the filing rather than the press. The lockup expiry is a dated supply event on the calendar from the day the company lists.

Most consequential for an ordinary investor is the access structure: the offer price is largely closed to individuals, the private markets that precede the IPO are gated by the accredited-investor thresholds, public proxies deliver only diluted exposure, and the open market — the one route open to everyone — is precisely where the pop has already been priced in. The route the company chooses changes what the first-day figures even mean, and the Dutch-auction case shows the first-day spread is a design choice rather than a fixed property of going public.

What an informed investor is left with, then, is not a verdict on any single offering but a way to read one: which price is being quoted and who already captured the discount, whether a number is audited or annualized, which structural features sit in the filing waiting to be read, and which of the few real access routes actually applies. The mechanics do not tell anyone whether to participate. They make clear exactly what participating would mean.

## Sources

The durable, load-bearing figures here — Ritter's first-day-return averages, the named case-study IPO numbers, and the accredited-investor thresholds — were spot-verified against primary and contemporaneous sources in June 2026. The Anthropic-specific figures are press-reported and provisional until a public S-1 exists.

**IPO mechanics, underpricing, and academic data**
- {ritter} [Jay R. Ritter — *Initial Public Offerings: Updated Statistics* (University of Florida)](https://site.warrington.ufl.edu/ritter/files/IPO-Statistics.pdf) — average first-day returns by era (~17–18% for 2001–2023; ~65% for 1999–2000)
- {ipohub} [IPOHub — IPO underpricing causes](https://www.ipohub.org/article/ipo-underpricing-causes) — who captures the first-day gain
- {globeandmail} [The Globe and Mail — study says underpricing data skewed by enthusiasts](https://www.theglobeandmail.com/investing/markets/stocks/FIG/pressreleases/34235615/ipo-pops-overstated-new-study-says-enthusiasts-skew-underpricing-data/) — the ~40%-overstatement claim
- {duke} [Duke FinReg Blog — negative first-day returns / IPO overpricing](https://sites.duke.edu/thefinregblog/2022/10/26/the-ipo-overpricing-phenomenon-debunking-the-determinants-of-negative-first-day-ipo-returns-in-the-us/) — the ~21–25% negative-first-day rate
- {valuethemarkets} [Value The Markets — IPO lockup period and expiry](https://www.valuethemarkets.com/analysis/investing-ideas/understanding-the-ipo-lockup-period-and-expiry) — 90–180-day lockup mechanics and lifecycle timing
- {cornell} [Cornell INFO 2040 — Dutch auction vs bookbuilding](https://blogs.cornell.edu/info2040/2012/09/26/going-public-is-a-dutch-auction-ipo-more-efficient-than-traditional-book-building/) — clearing-price mechanics

**Case studies**
- {snowflake} [CNBC — Snowflake debut](https://www.cnbc.com/2020/09/16/snowflake-snow-opening-trading-on-the-nyse.html) and [CNBC — Snowflake left $3.8B on the table](https://www.cnbc.com/2020/09/17/snowflakes-first-day-pop-means-ipo-left-3point8-billion-on-the-table-the-most-in-12-years.html) — offer $120, opened $245, +112% first day
- {snowflake-drawdown} [24/7 Wall St. — Snowflake IPO buyers still down ~41% years later](https://247wallst.com/investing/2026/04/22/if-you-bought-snowflake-ipo-youre-still-down-41-more-than-5-years-later/) — long-run drawdown
- {airbnb} [CNBC — Airbnb soars 112% in debut](https://www.cnbc.com/2020/12/10/airbnb-ipo-abnb-starts-trading-on-the-nasdaq.html) — offer $68, opened $146
- {visa-ipo} [CNN Money — Visa IPO priced at $44](https://money.cnn.com/2008/03/18/news/companies/visa_ipo.fortune/index.htm) — offer price and size
- {visa-longrun} [The Motley Fool — Visa long-run return](https://www.fool.com/investing/2020/02/13/if-you-invested-100-in-visa-ipo-this-is-how-much-m.aspx) — long-run holder outcome
- {google-dutch} [McLane — Google's 2004 Dutch auction IPO](https://www.mclane.com/insights/googles-dutch-auction-ipo-is-there-a-take-away-lesson-for-the-rest-of-us/) — range cut from $108–135 to $85–95, priced $85, ~18% first day
- {facebook} [Wikipedia — Initial public offering of Facebook](https://en.wikipedia.org/wiki/Initial_public_offering_of_Facebook) and [CNBC — Facebook below offer price](https://www.cnbc.com/2012/05/21/facebook-shares-fall-below-ipo-offering-price.html) — $38 offer, decline, causes
- {uber} [CNBC — Uber ends first day down ~7%](https://www.cnbc.com/2019/05/10/uber-ipo-stock-starts-trading-on-the-new-york-stock-exchange.html) — $45 offer, opened $42
- {lyft} [Yahoo Finance — Lyft ~25% first-month drop](https://finance.yahoo.com/news/uber-ipo-to-begin-trading-on-nyse-135445721.html) — decelerating-growth context
- {rivian} [CNBC — Rivian prices at $78](https://www.cnbc.com/2021/11/09/rivian-prices-ipo-at-78-a-share-valuing-company-at-66point5-billion.html) and [Nasdaq — Rivian below $78](https://www.nasdaq.com/articles/rivian-stock-just-fell-below-its-ipo-price-of-$78-per-share:-time-to-buy) — offer $78, below offer within weeks
- {robinhood} [CNBC — Robinhood below IPO price](https://www.cnbc.com/2021/10/27/robinhood-drops-10percent-to-below-ipo-price-as-investors-worry-about-bleak-outlook.html) — below $38 offer
- {peloton} [Yahoo Finance — Peloton ~11% below offer](https://finance.yahoo.com/news/pelotons-stock-takes-spill-even-173555397.html) — broken debut
- {wework} [HBS Working Knowledge — WeWork, the IPO that shouldn't](https://www.library.hbs.edu/working-knowledge/wework-the-ipo-that-shouldn-t) — governance, pulled deal, later SPAC route
- {coinbase} [Yahoo Finance — Coinbase closes 14% below opening price](https://finance.yahoo.com/news/coinbase-direct-listing-stock-cryptocurrency-bitcoin-172655492.html) — direct listing opened $381, closed ~$328

**Regulatory definitions**
- {sec-accredited} [SEC — Accredited Investors](https://www.sec.gov/resources-small-businesses/capital-raising-building-blocks/accredited-investors) — income/net-worth thresholds, license-holder paths, and pending qualification-by-exam proposal
- {sec-networth} [SEC — "Accredited Investor" Net Worth Standard](https://www.sec.gov/resources-small-businesses/small-business-compliance-guides/accredited-investor-net-worth-standard) — the primary-residence exclusion
- {carltonfields} [Carlton Fields — JOBS Act: Emerging Growth Companies](https://www.carltonfields.com/insights/publications/2012/jobs-act-emerging-growth-companies) — EGC definition (under $1B revenue) and confidential-draft eligibility
- {jdsupra} [JD Supra — JOBS Act confidential review option for EGC IPOs](https://www.jdsupra.com/post/contentViewerEmbed.aspx?fid=28294cae-1127-4886-8afe-fed9a97b6edd) — confidential draft S-1 mechanics

**The Anthropic filing (press-reported; to be re-confirmed against the public filing)**
- {anthropic} [Anthropic — confidential draft S-1 announcement](https://www.anthropic.com/news/confidential-draft-s1-sec) — filing status; no price or share count set
- {fortune-anthropic} [Fortune — Anthropic confidentially files IPO at $965B valuation](https://fortune.com/2026/06/01/anthropic-confidentially-files-ipo-965-billion-valuation/) — valuation, Series H
- {yahoo-anthropic} [Yahoo Finance — Anthropic files confidential S-1](https://finance.yahoo.com/markets/stocks/articles/anthropic-files-confidential-1-joins-161008569.html) — ~$47B run-rate, Google ~14% stake, Long-Term Benefit Trust
- {fortune-stake} [Fortune — Google/Amazon AI profits and the Anthropic stake](https://fortune.com/2026/04/30/google-amazon-ai-profits-anthropic-stake-bubble-earnings-2026/) — related-party / holder context