Why AI Names Some Brands and Not Others: The Evidence

ChatGPT knows 99.4% of startups when you name them. Asked which tools are worth using, it names 3.32%. Same models. Same brands.

The gap nobody prices

A study published in January 2026 took 112 Product Hunt startups and ran 2,240 queries at two models. It asked each question twice. Two shapes.

Note what the two models were, because it matters later: gpt-4o-mini with no web access, and Perplexity sonar with search on. One is the pure weights case. The other is weights plus retrieval.

First by name. Do you know this product?

Two answers: ChatGPT knew 99.4% of them, Perplexity 94.3%.

Then the same products, asked the way a buyer asks them. What are the best AI tools launched this year?

Two more: ChatGPT named 3.32%, Perplexity 8.29%.

The author does the arithmetic. "That's a gap of 30-to-1 for ChatGPT."

Sit with what that separates. Recognition is not recall.

The model has heard of you. It does not think of you.

Every audit in Part II assumed one thing: a system that had already found your page. This chapter is about the step before that. Almost nobody measures it.

Because your competitor did not beat you on a page. They were already in the answer before any page was read.

One more number, from another study. It decides whether this chapter is about you.

Ahrefs ran a sample against AI Overviews in May 2025. Established brands, real traffic, nobody's first week.

26% of them had zero mentions there. Not few. Zero.

So a quarter of credible companies are not in the answer at all. Is that you? You do not know yet. Artifact 8.5 finds out.

Why the model knows what it knows

The mechanism is not mysterious. It has been measured. The result is one of the cleanest findings in the field.

Researchers counted how many pretraining documents mentioned each fact. Then tested whether models could answer questions about it.

Their finding, in one sentence: "a language model's ability to answer a fact-based question relates to how many documents associated with that question were seen during pre-training."

Now the part that matters. The relationship is log-linear, and it holds across five pretraining corpora and three model families.

Each tenfold increase in coverage buys a roughly constant step of accuracy. Not a proportional one.

So the curve is the chapter. Coverage compounds slowly, and it compounds on a log scale.

Ten mentions to a hundred is ninety more. A hundred to a thousand is nine hundred more.

Same step on the chart. Very different invoice.

And you cannot wait for bigger models to fix it. The same authors priced that too. The number is a quintillion parameters.

Which is not a roadmap. It is a way of saying no. Politely.

What the model does instead of naming you

An absence has to be filled with something. So watch what fills it.

A 2023 paper tested it. What happens when a fact's two halves rarely appear together in training? Their finding: models "struggle to recall facts whose subject and object rarely co-occur in the pre-training dataset." Even when finetuned on them.

Then the mechanism, and it should be familiar. In 95% of those failures the model reached for a higher-co-occurrence word. It substituted.

Read that as a business fact. Asked for a category and a quality, the model returns whichever name has sat next to that quality most often. Not the best one.

Not the best one. The most co-occurrent one. Which one are you?

That substitution is what happened the last time an engine named your competitor in a sentence describing you.

Try it while you read this. Ask any engine for the best tool in your category, for the thing you do best. One prompt.

Then look at the name that came back, and ask a different question. Not whether they are better.

Ask how many independent articles pair that name with that quality, and how many pair yours.

If their count is an order of magnitude above yours, you did not lose a comparison. You lost a count, and the model never ran the comparison at all.

Scale does not rescue this either. The authors are plain. The bias "remains despite scaling up model sizes or finetuning."

A Meta team built a benchmark to size the same effect across 18,171 questions. They split entities into head, torso and tail by cumulative popularity.

GPT-4 accuracy, averaged across all four domains: 40.3 on head entities, 33.4 on torso, 19.0 on tail.

Now the hallucination rate, which runs the other way. 23.3% on head questions, 19.7% on torso, 15.9% on tail.

Read that inversion carefully. We did not expect it either.

The model invents most freely about the entities it knows best.

Which is the trap, and it is not the one people expect. Being in the tail does not mostly get you described wrongly. It gets you left out.

Two channels, and only one of them is yours

Here is where most of this field goes wrong. OpenAI documented the fix in a page nobody in marketing reads.

They run separate crawlers, and the separation is the point.

So the honest split. Retrieval: the part you can move this quarter.

Recall decides something else: whether the model reaches for you when nothing is retrieved at all.

Anthropic makes the lag legible in a way nobody else does. Their model page carries two columns. A training data cutoff, and a reliable knowledge cutoff.

For one of their models those columns sit five months apart, and reliable is the earlier of the two.

Trained on data through July 2025. Reliable on it only through February 2025.

Which is the same finding a 2023 paper reached from the other direction. Facts seen once, in one form, can be memorized and still not be extractable. The paper puts the extraction number at zero.

Not low. Zero, "regardless of subsequent instruction fine-tuning."

What makes a fact extractable: being said many times, in many forms, by many people.

Which is a description of coverage. It is also, precisely, the thing you cannot buy in one quarter. Or four.

That split hands you one live decision, and a lot of companies got it backwards in 2025.

Should you block GPTBot?

Read the two rows again. Blocking GPTBot says your content "should not be used in training." Blocking OAI-SearchBot removes you from ChatGPT search answers, though you can still appear as a navigational link. Different doors.

They are separate directives, and most robots.txt files written in a panic blocked both.

So take the decision on its own terms. Is your business publishing, with models trained on your archive competing with you? Then blocking training is a defensible commercial choice.

Does your business sell something else? Then blocking GPTBot opts you out of the only channel this chapter is about. You are asking not to be remembered.

Which settles something Chapter 3 left open. It recommended the training block, correctly, on the grounds that blocking a pure-training crawler costs you nothing in citations.

True, and citations are retrieval. Chapter 3 also said nobody had measured the other cost.

This chapter is that cost. Not citations: recall.

So go and edit the file. Chapter 3 shipped you a robots.txt with a hard GPTBot block in it, and if you are not a publisher, that line should come out.

Check yours today. Two lines in one file. Plenty say something nobody decided.

Then the planning question. How long?

Nobody publishes a retraining schedule, so the honest answer is assembled from cutoffs. OpenAI's current models carry a February 2026 cutoff. Anthropic publishes theirs per model.

Google publishes none at all for Gemini, which is its own kind of answer.

The observable cadence is roughly annual, per model family. Plus a release lag.

Which sets the unit of measurement for this chapter. Coverage published this quarter reaches a model that ships next year. Reliable recall lands months later.

Two years is not pessimism. It is the mechanism, and what does your plan assume?

What actually correlates with being named

Now what moves it. The honest answer: the two best datasets disagree.

One thing both studies agree on, and one thing only the smaller one tested. Both are worth money.

Off-site mentions outrank technical authority. Backlinks and domain rating sit near the bottom of the Ahrefs table. Under branded mentions. Well under YouTube.

Then on-page GEO work, which only the 112-startup study measured. It is explicit: no correlation with actual discovery rates.

Its recommendation reads oddly for a paper about AI. "Build the SEO foundation first and LLM visibility will follow."

Then the disagreement, and do not smooth it over. On the larger sample, off-site mentions track ChatGPT visibility at 0.664.

On the smaller one, with search switched off, nothing tracked visibility at all.

Two readings survive. Either the effect is real and 112 startups is too small to see it. Or it is real for large brands and absent for small ones.

Notice that both readings are bad news for the same reader. The company that most needs to be named is the one the evidence is thinnest for. Read that twice.

Then a second disagreement, about how much room there is at the top.

One 2026 study ran 250 queries across three models. Category concentration came back at a Gini coefficient of 0.28, top three brands taking 48.6%. Crowded, but not closed.

A vendor study of 126 million prompts found the top three taking 82.9% in news. Consumer electronics: 76.9%. Closed.

Do not average those. Different prompts, different engines, different definitions of visibility. And one sells visibility software.

What survives both is the shape. Whatever the exact number, a small set of names takes most of the answer. So the distance between third and fourth is worth more than the distance between tenth and eleventh. Where are you?

Four readers, four different problems

That gap means something different depending on where you sit.

First, the incumbent. You are named already, in most runs, in your own category.

Your problem is not acquisition. It is that a log curve protects you. The coverage behind your name took a decade. A challenger cannot buy that quickly.

So your work is defensive. And cheap. Measure quarterly. Watch for a rival closing. Do nothing else here.

Second, the challenger in a category that already exists. You are the reader this chapter is written for. The honest news is mixed.

The curve says the gap is real and slow. The 112-startup study says nobody has found a lever that closes it. Not on the closed-book arm, at least.

So your move is three things: measure the gap, fund coverage as a multi-year line, spend this quarter on Part II. That is where things move.

Third, the company creating a category nobody searches for yet. Your problem inverts.

There is no established set of names for a model to reach for. So co-occurrence is not against you. It is not working at all.

So your prompts have to be the job, not the category. And your coverage has to teach the category before it can teach your name in it.

Fourth, the local or regional business. Your category has a place in it. Literally. Place is the thing training data is thinnest on.

Check whether the engines name any local provider or default to national ones. If they default to national, this chapter's fix is not available to you at any price. The panel work in Chapter 7 is where your money goes.

Traced to source

Chapter 5 traced five writing prescriptions. Chapter 6 traced five entity claims. Chapter 7 traced five schema claims.

Fourth time, and the shape does not change.

Row one deserves the walk, because it is the most quoted number in this industry.

The paper is real, peer reviewed, and its authors are careful. It tests nine writing methods. The best of them lift a visibility score by up to 40%.

Now read the setup. The engine is a pipeline the authors built. The metric counts position-weighted words.

And every source being tested is already inside a five-document context window.

A 2026 survey of the field states the consequence plainly. The result is "conditional on a source already being present in a fixed context." It establishes "neither organic discoverability nor durable traffic effects." Neither.

The number is not wrong. It answers a question about editing. It is sold as an answer about being found.

What you cannot buy

Wikipedia is where this gets concrete. It is the one node where the rules are written down.

Start with the size of it. In one open training corpus, Wikipedia and Wikibooks are 4.3 billion tokens. Out of 3,059 billion.

That is 0.14%.

Which should end the idea that a Wikipedia article teaches the model about you by volume. It does not. Whatever it does, it does by being clean, linked and repeated elsewhere.

Now the rules, and they are written to stop exactly what a marketing team would try.

The standard: "significant coverage in multiple reliable secondary sources that are independent of the subject." Then the guideline lists what does not count.

The list is the modern PR stack. All of it.

Press releases. Anything "substantially based on such press releases even if published by independent sources," which the guideline calls churnalism. Any "paid or sponsored articles," including contributor posts at named publications that take outside submissions.

Their sentence, and it is four words: "Only unpaid sources count."

Syndication does not help either. One wire story reprinted in forty papers is one source. A series by the same journalist is one source.

Different articles in the same newspaper are one source.

And the guideline says why it is built this way. The bar exists "to prevent gaming of the rules by marketing and public relations professionals." Their words, not ours.

Then the sentence to read to anyone selling you a page. "Only a small percentage of the world's organizations meet the requirements for a Wikipedia article."

Clear it, and one more thing. You do not get to run it.

Paid editing must be disclosed: it is "a requirement of the Wikimedia Foundation." Direct editing is "strongly discouraged."

And once an article exists? You have "no right to control its content, or to delete it outside the normal channels."

The guideline also notes that offers to guarantee a page survives deletion, for money, "are always fraudulent."

So price it honestly. Wikipedia is downstream of coverage, not a substitute for it. The only route in is the slow one.

Chapter 6 left one thread here, and this is where it lands.

That chapter told you to pull 30-day Wikipedia pageviews for every entity competing for your name. If a rival beat you by an order of magnitude, it said, you were arguing with the training data. Not with a competitor.

Here is why that instrument works. The research that established the popularity effect used exactly that number as its proxy. Wikipedia monthly pageviews, standing in for "how frequently the entities are likely to be discussed on the web."

The pageview count was never about Wikipedia. It is the cheapest public estimate of one thing: how often the world writes about you.

One agency study puts the correlation between 7-day Wikipedia views and model visibility at 0.810. That is the highest coefficient in this chapter's evidence.

Chapter 6 already priced it: 27 tracked athletes, every one of them already carrying a Wikipedia article, and no published selection criteria.

Treat it carefully.

Take the instrument. Hold the coefficient loosely.

Which gives you a definition worth stealing, whatever you think of the encyclopedia.

Count one unit of coverage as this: one publisher, one journalist, one story, unpaid. Written for their reasons, not yours.

By that count a wire pickup across forty outlets is one. A three-part series by the same writer is one. Your sponsored column is zero.

Most marketing dashboards would score that same quarter at fifty. Which number would you rather plan against?

Where the canceled budget actually goes

Chapter 6 routed canceled line items here four separate times. So this chapter owes an answer. It is not "buy coverage."

Three things are worth funding. They are ordered by cost and reversibility, not by evidence, because the evidence does not rank them.

First, and it is nearly free: make every mention land on one string. Chapter 6 built that audit. It belongs here because of the co-occurrence finding.

A model reaches for one name: whichever co-occurs most with the category. Split your mentions across two spellings, a product name and a legal entity, and you have divided your own count. By four.

Nothing accumulates on four names. How many are you using?

Second, and this is where the budget goes if you fund this at all. Fund the thing that produces coverage, rather than coverage itself.

What clears Wikipedia's bar is what compounds in a training corpus. The overlap is not a coincidence. Both reward independent people writing about you for their own reasons.

So the line item is not placements. Ever. It is whatever makes a journalist, a researcher or a practitioner want to write. Which is a product question before it is a marketing one.

Third, and last, because it is the one you control least: the platforms where mentions accumulate.

YouTube topped the largest correlation table in this chapter at 0.737. Reddit was the strongest Perplexity predictor in the smaller one.

Neither is a placement you buy. Both are Chapter 9.

Now what not to fund, on this chapter's evidence. Press release volume: the origin study counts crawler hits, with no baseline.

Paid contributor posts, which Wikipedia excludes by name and which nobody has shown do anything else.

And any tool whose output is a rank position. Repeat runs put the odds of two identical orderings near 1 in 1,000.

What this costs, honestly

Nobody prices the coverage side either, so here is the shape of it.

Three costs, and only one appears on an invoice.

The visible one: whatever you pay a communications team or an agency. You already know that number.

The second decides whether any of it works. It is the time of the people a journalist actually wants to talk to. Usually a founder, a researcher or whoever runs the thing being written about.

That time competes directly with product. It is the real currency here. No budget line captures it.

The third cost is duration. On a log curve, a program that runs four quarters and stops has bought most of one step. One.

But the failure mode is not underspending. It is spending correctly for three quarters and then reallocating. That produces the cost with none of the accumulation.

Against that, price the alternative honestly. Everything in Part II is cheaper, faster and better evidenced. None of it touches the gap in Figure 8.0.

Which is the trade, stated plainly. Fast work on a surface you can move. Or slow work on the one that decides whether you are in the room.

One test tells you which side of that trade your company is actually on.

Look at your last four quarters of marketing spend. How much of it produced something a stranger would cite?

Not impressions. Not sessions. The unit is narrower: a thing with your name in it, written by somebody who does not work for you and was not paid to write it.

For most companies the honest answer is close to nothing, and that is the number Figure 8.1 is pricing.

Most companies should do the fast work and fund the slow one. In that order, and without letting the first be sold as the second.

The objections this chapter has to answer

"You have shown me a training-data law from 2023 and two correlational studies that disagree. Then you told me the fix takes years. What am I supposed to do on Monday?"

Fair, and the chapter's answer is that Monday belongs to Part II. Retrieval is the surface you can move this quarter. Chapters 4 through 7 are what to do there.

This chapter is about the other clock. Both are real. Only one is being sold to you as fast.

A second, and it is the strongest thing anyone can say against the whole argument.

"Grounding changed this. Engines retrieve now. My weights problem is somebody else's 2023 problem."

Partly right, and the source that proves it is the same paper this chapter is built on. Its authors say retrieval "can reduce the dependence on relevant pre-training information." And that it helps most on the rare entities recall fails.

So retrieval is the mitigation. Take it.

What it does not do is decide the shortlist. When a model answers a category question without retrieving anything, you are back on the training clock. Models still do that.

A third, which cuts against us harder than the second.

"A controlled 2026 study found brand identity explains 1.2% of variance in what a model recommends. Product attributes explain 82.4%. Your chapter is about a rounding error."

That study is real and we will not shrink it. But read what else it found.

Where a real brand appeared, it was recommended in 100% of 670 trials. And the fictional brands, unoptimized, received zero recommendations.

So the boundary is not known against unknown. It is whether the model has anything else to go on.

Where products differ on rating, price and reviews, the brand factor collapses to 1.2%. Where they do not, the known name takes all of it.

Their phrase for it: brand "works as a tiebreaker." Which is the whole of this chapter, priced honestly.

A fourth, from the reader who has done the arithmetic. "If coverage compounds on a log scale and the model retrains once a year, you are describing something I cannot influence inside a planning cycle."

Correct. That is the finding, not a failure to find one.

What follows from it is a different allocation, not a bigger one. Coverage is a slow asset with a long half-life. Fund it like one, rather than sprinting at it each quarter.

Then the one we raise against ourselves. Nobody has published a study in which coverage was added and unprompted recall was then measured to move. Nobody, anywhere.

We looked hard, and so did the field's own reviewers. A survey of 45 studies, published July 2026, states it. No technique shows "a stable, longitudinal, cross-platform causal effect on organic discoverability."

Here is how hard we looked, because an absence is only worth printing if somebody tried to break it.

We ran five search formulations: coverage added then AI mentions measured, longitudinal GEO intervention, digital PR lift, controlled field experiment, difference-in-differences on AI visibility. We chased every result that was not a vendor blog.

Two candidates came close, and both fail on inspection.

The press release test measured crawler requests inside six hours, with no baseline, on the testing firm's own announcement. Crawler hits are not answers.

The mention-effect study is good, and it measures the wrong end. Table 8.4 carries its number, and the number is real.

What it cannot carry is how the mention arose.

So the causal claim in this chapter is one we cannot make. What we have is three things: a measured law about document counts, two correlational datasets that disagree, and a rule about who carries the burden.

The disclosure this chapter owes you. Chapter 6 routed four canceled line items here, and Chapter 7 routed its canceled budget into this part. Earned coverage is a larger engagement than the retainers it replaced.

We sell into that, and we have sold it. Which is exactly why Artifact 8.5 returns a measurement rather than a proposal.

Ask us the question anyway. If our answer to "what will this move" is a number with no study behind it, we have become the thing this book has spent eight chapters describing.

What is left

Google's own line is the right place to stand. "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."

Read it next to the log curve and the two of them agree. There is no lever here, because it was never a lever.

It is an accumulation. And it belongs to the part of your business that earns coverage, not the part that publishes pages.

So the three things this chapter can defend. Recall tracks one thing: how many documents mention you, and it tracks it on a log scale.

Retrieval and training are separate channels on separate clocks, and the engines document them separately.

And nobody has shown that buying coverage moves the first one.

Which leaves the question of where coverage actually lives. Not the press release, on the evidence in Table 8.4.

Two signals topped the largest table in this chapter: YouTube mentions, and branded web mentions. Neither is a channel you can buy your way into cleanly.

Chapter 9 is about where those accumulate. Two platforms matter more than their traffic suggests.