AI brand visibility: across 75,000 brands, the top correlate was not on the site

What is AI brand visibility, and what did this study actually measure?

The study selected brands rather than pages. Ahrefs filtered for domains with a Domain Rating above 40, took each domain's highest volume keyword where that keyword carried at least 800 monthly searches, and treated the result as a brand, a method they describe in their own methodology as not a perfect calculus but one that worked well enough to find 75,000 brands. They then analysed what they describe as millions of AI responses using their brand monitoring product, and correlated each brand's search metrics against how often it was mentioned.

Three details in that methodology change how the numbers should be read, and all three are stated on their page rather than inferred here. The correlation is Spearman, so it measures rank agreement rather than a linear relationship. The question pools differ by platform: ChatGPT, Copilot, Gemini and Perplexity share one pool while AI Overviews and AI Mode share another, and the AI Overviews figures are carried over from earlier research for benchmark comparison rather than collected fresh alongside the rest. And the authors print the disclaimer themselves, that correlation is not causation and that improving these metrics will not automatically raise anyone's visibility.

The unit being predicted is a brand mention in a generated answer. That is not the same unit as a citation, and it is not the same unit as a fetch of your page. An assistant can name a company from what the model already holds, from a third party page it retrieved, or from your own page, and a mention count does not separate those three. We have written before about why a citation count is not answer influence, and the same gap applies here one level up: a mention is presence in an answer, and presence is not attribution to a source.

The distinction matters for anyone buying a tool on the strength of these numbers. A brand mention study describes the selection layer, where an engine decides which companies are worth naming. A scanner describes the access layer, where a crawler either gets your bytes or does not. Generative engine optimization as a discipline spans both, and most of the disagreement about what it means comes from people standing on different layers. It is also worth knowing which engines in a given tool can retrieve at all: reading our own registry on 29 August 2026 we found that three of the eight engines we ask cannot search the web, so a blended visibility percentage can mix live retrieval with recall.

Every signal at the top of the ranking sits off your own site

Sorted by strength, the ranking is unusual for a document published by an SEO company, because the things it puts at the top are not things an SEO controls on the page. YouTube mentions came first at about 0.737, defined in the study as a brand name appearing in a video title, transcript or description. YouTube mention impressions, the same mentions weighted by how many views each video received, came second at about 0.717. Branded web mentions, meaning the brand named on other people's pages, ran from 0.66 to 0.71 depending on platform.

Below those sit branded anchors, the visible clickable text of a link that carries the brand name, at 0.628 on AI Mode against 0.511 on ChatGPT and 0.527 on AI Overviews. Branded search volume follows at 0.466, 0.352 and 0.392 on the same three. Branded traffic is lower again at 0.357, 0.235 and 0.274. Domain Rating on ChatGPT sits at 0.266, which the study describes as a mid tier factor much weaker than the branded signals.

Two negative results are the more useful half. Number of site pages correlated at about 0.194, which the authors call almost no relationship, and they name the practice this argues against directly: investing in programmatic content to raise AI visibility. Link metrics, meaning number of backlinks and URL rating, are reported as very weak across all three systems, and the study prints no coefficient for them, so neither does this post.

Every factor above 0.4 in that list is measured somewhere a site owner does not own. Video titles and transcripts belong to YouTube. Mentions belong to whoever wrote the page. Anchor text belongs to whoever chose to link. Branded search volume belongs to the people typing. This is the same boundary we ran into in five structural signals that tell an AI who you are, where six of the 30 pages we fetched as raw HTML on 26 July 2026 carried no JSON-LD at all, two of them developer.mozilla.org and theguardian.com, so whatever made those pages worth quoting was not markup on the page. It also sets a limit on what markup can do: structured data can make a page unambiguous about which entity it describes, which is worth doing, and there is nothing in this ranking suggesting it will make an unmentioned brand mentioned. The study's own summary of the hierarchy puts YouTube presence and brand mentions first, branded anchors and search volume second, and traditional authority metrics last.

One further figure belongs here because it constrains how differently the three engines behave. The study reports pairwise correlations between which brands each assistant mentions: 0.821 between AI Overviews and AI Mode, 0.769 between AI Mode and ChatGPT, and 0.749 between AI Overviews and ChatGPT. The engines weight signals differently and largely name the same companies anyway. A tool that reports three separate visibility scores is reporting three views of a heavily overlapping set. Entity confidence is the vocabulary we use for the underlying question of whether an engine knows which company a name refers to at all.

The layer underneath: a gate that grants nothing

None of the above says the page does not matter. It says the page is not where the variance sits among brands that already clear the bar. The bar itself is documented, and it is worth reading precisely because it is so much less demanding than the optimisation advice built on top of it.

Google's own documentation on AI features and your website, last updated 10 December 2025, states that to be eligible to be shown as a supporting link in AI Overviews or AI Mode a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements, and that there are no additional technical requirements. The same page says there is no separate ranking system for AI features and no special optimisation needed, and points at nosnippet, data-nosnippet, max-snippet and noindex as the controls that restrict what is shown.

Read that as an engineer rather than as a marketer and it describes a gate. Passing it grants nothing. Failing it removes the option entirely, and no volume of YouTube mentions restores it, because the correlations in the previous section are about which of the eligible brands get named. This is why a scanner and a brand monitor are complements rather than substitutes, and why we say so on the comparison page rather than claiming our measurement subsumes theirs.

What that gate is made of is what a scanner reads. Whether the crawler is allowed by robots.txt, whether the response is a 200 rather than a challenge page, and whether the text a human sees is present for a client that runs no JavaScript, which we call prose parity. Our scoring method is public, and the what GPTBot sees view exists so a reader can compare a crawler's copy of a page against their own without taking our word for the difference.

There is an honest asymmetry to state here rather than bury. Failing the gate is rare and catastrophic; passing it is common and buys nothing. That means for most sites the interesting number is not our score but the brand signal ranking above, and a reader whose pages already return 200 to every named crawler is better served spending the next month on the things in section two than on anything we sell. Where our measurement earns its place is the case where something is quietly broken, and those cases are not rare enough to ignore: our platform guide for Google AI Overviews exists because the eligibility rule above is met at the HTTP layer or not at all.

Counting brand mentions is itself a measurement, and ours was wrong

Every correlation in this study rests on a mention count. Something had to decide, for each of millions of generated answers, whether a given brand was mentioned in it. That decision is a measurement with its own error rate, and it is the one part of this subject where we can speak from our own record rather than reporting someone else's.

On 13 August 2026 our prompt tracker reported a 12.6 percent brand mention rate on its first live run: 19 counted mentions across 151 classified answers. The true figure was near 1 percent. Of the 19, eight were an engine defining the Tagalog adjective lantad, three resolved the name to a different business, and six were engines naming the brand only to say they had never heard of it. We published that correction as nineteen counted brand mentions, and about two were the company, and the fix was to split the question in two: a deterministic matcher decides whether the characters are present, and a separate judge decides what those characters refer to.

That failure is not evidence against the Ahrefs study, and it should not be read as one. Their brand selection deliberately screens for domains with a Domain Rating above 40 and a head keyword above 800 searches a month, which filters out exactly the ambiguous small name our tracker tripped over. A study of Nike and Apple has far less of this problem than a study of a scanner nobody has heard of. It is evidence about what the number means: a mention rate is a claim about string matching and reference resolution before it is a claim about anything else, and the reasonable thing to ask any vendor, us included, is how they separated the two.

The second inherited limit is sampling. A paper by Julius Schulte, Malte Bleeker and Philipp Kaufmann, posted to arXiv on 8 April 2026 as Don't Measure Once: Measuring Visibility in AI Search, argues that visibility should be characterised as a distribution across repeated measurements rather than a single point outcome, because repeated identical prompts return different sources. We wrote that one up as the finding that one run is one draw. Correlating across 75,000 brands averages a great deal of that noise away, which is a real advantage of working at that scale and one we do not have. It does mean the same coefficients would be poor evidence about any single brand, and any single brand is what a buyer is actually asking about.

What to check on your own site, and what to stop checking

The practical reading of this study, for someone who owns a site rather than a research budget, is that the on page work is a short list with an end, and the off site work does not have one. Finish the short list, then stop treating it as the lever.

The short list is the gate. Confirm that the crawlers you care about are not disallowed, which is what a robots.txt tester is for and which is worth re-checking after any platform migration, because the file is frequently generated rather than written. Confirm the tokens you are naming are the ones the operators actually publish, since a rule aimed at a token nobody sends does nothing; our crawler directory lists the fifteen we evaluate and links each operator's own documentation. Confirm the page returns your text without JavaScript, and confirm the response is a 200 rather than a bot challenge: requesting the home page of 391 hostnames on 9 September 2026 we recorded five captcha responses and twelve Cloudflare managed challenges, seventeen hostnames in all answering a scanner with something other than their page.

What the study argues against spending the next quarter on is more pages. Number of site pages correlated at about 0.194, and the authors say so in the plainest terms available to them, naming programmatic content as the specific practice the data does not support. Anyone weighing a large content programme against the cost of getting mentioned on other people's sites now has a published number on one side of that decision and should read the study's own caveat about causation before treating it as settled.

There is one thing we cannot help with at all, and it is the top of the ranking. We do not measure YouTube mentions, we do not hold an index of video transcripts, and we have no view of branded anchor text across the web. Ahrefs does hold that data, which is what made this study possible and is a straightforward reason to use their instrument rather than ours if the question you are asking is which brands get named. The question we answer is narrower and sits earlier in the chain: when an engine does go looking, can it read what it finds. Those are different measurements, and a report that claimed to be both would be worth less than either.