Brand Mentions in AI Search

Two questions hiding inside one claim

A brand mention is any occurrence of a brand's name — the name a company trades under, and the associations attached to it — in text on a page that brand does not control. An unlinked mention is one with no hyperlink pointing back: the name appears, nothing points anywhere. Under classic SEO an unlinked mention was a curiosity, something to convert into a link. Since 2025 it has been sold as the primary input to AI search visibility, on the reasoning that language models consume text rather than link graphs.

That reasoning bundles two questions that have different mechanisms, different evidence and different timescales. The first: do unlinked mentions shape what a model knows, through its training data? The second: do they shape what a grounded answer says, through retrieval at the moment the answer is written? The second has a clean mechanical answer. The first does not, and the industry has been answering it with confidence for two years on evidence that does not support confidence.

A third measurement gets folded in with both and should not be. Whether an answer names your brand is separate from whether it cites a page of yours. Brand awareness in an AI answer and citation in an AI answer are two outputs, and they come apart in both directions.

The retrieval mechanism, which is real

When an AI surface runs a live search and reads pages, the text of those pages enters the model's context window. If a third-party page names your brand, that name sits in front of the model at generation time whether or not it links to you. The model can therefore name your brand in an answer assembled entirely from pages you do not own.

This is not speculation. It follows directly from retrieval behavior the platforms themselves describe: Google's documented query fan-out across subtopics and data sources for AI Overviews and AI Mode, OpenAI's documented rewriting of a question into targeted queries sent to search providers, Perplexity's numbered citations of the pages it fetched. Retrieved text is read; a name in retrieved text is available to be spoken.

Two things follow. Where your brand appears on the kind of third-party page retrieved for the questions you care about — a comparison, a category roundup, a review — the mention is in play on a timescale of days, not model releases. Where it does not, on-site work is no substitute, because the model can only speak names it has in front of it. This is the version of the brand-mention claim a publisher can act on within an ordinary planning horizon, and it is not the version most often sold.

The training-data mechanism, which is an inference

The other story runs through pretraining. Text crawled from the web becomes training data; a brand named repeatedly across that text becomes something the model holds statistical associations about, retrievable with no search at all. No link is needed, because pretraining consumes text rather than link graphs.

This is a sound inference — a conclusion that follows from how the system is built, as distinct from one demonstrated by measurement — and in general terms it is almost certainly true. What cannot be established from outside is anything specific enough to act on. No operator publishes training-corpus composition. Nobody outside the labs can say which corpora a given model used, whether a given site was in them, how many mentions shift an association, or how long the lag runs between a mention appearing and a model reflecting it. Model release cadence alone makes that loop months long and unattributable.

The distinction is not academic, it is commercial. An agency selling mentions on the promise that they will train the model to recommend you is selling an effect on a loop that is slow, invisible, and impossible to attribute to the work when it arrives. Describing a plausible mechanism as an observed one is the most common piece of dishonesty in this subject, and it is usually not deliberate — the two stories genuinely sound alike until you ask what would count as evidence for each.

What 0.664 actually measured

One number carries this entire discourse. Ahrefs, on 26 May 2025, ran a Spearman correlation across 75,000 brands between off-site footprint and AI Overview visibility and found branded web mentions leading at 0.664, against branded anchors at 0.527, branded search volume at 0.392, Domain Rating at 0.326 and backlinks at 0.218. Its top quartile for web mentions averaged 169 AI Overview mentions against 14 for the next quartile down. Ahrefs repeated the exercise on 12 December 2025 across ChatGPT, AI Mode and AI Overviews and got branded web mentions at 0.656 to 0.709, with YouTube mentions highest of all at about 0.737.

The measurement is real and it replicated. What it means is a different question, and Ahrefs answered it honestly in its own write-up: correlation is not causation, and the factors it tested showed moderate to very weak correlations. The 0.664 travels; the caveat does not.

Here is why the caveat matters. This is a cross-section of brands, and large, well-known brands have more web mentions, more AI visibility, more branded search, more backlinks and more YouTube coverage than small ones. Every one of those correlates with every other because all correlate with being a large, well-known brand. A coefficient computed across that cross-section cannot separate mentions producing visibility from both being produced by size. Seer Interactive's earlier study, published 7 January 2025 on more than 300,000 finance and SaaS keywords, has the same shape and points at a different variable: brands ranking on page one of Google correlated at roughly 0.65 with mentions in LLM answers, backlinks weak or neutral.

Both studies were funded by companies selling tools whose value proposition is affected by the result. That does not make either wrong, but the framing and the choice of what to publish were made by an interested party, and neither study was designed to answer the causal question.

Mentioned, cited, both, neither

Because naming and citing are separate outputs, four states exist, and two of them break most measurement programs.

Cited but not named. Seer Interactive named this the ghost citation in March 2026: your page is used as a source, the link appears, your brand is never spoken. Seer's framing is that the content is trusted enough to be sourced but the brand is not known well enough to be mentioned.

Named but not cited. The answer recommends you and links somewhere else. Semrush's 2026 AI Visibility Index, released 26 June 2026 across 126 million US prompts, found the overlap between brands mentioned and domains cited on Gemini can fall as low as 30%.

Seer's larger finding, across 541,213 responses covering 20 brands on 6 platforms, is that the two states travel together: when a brand is mentioned in a response, its citation rate is 53.1%; when it is not mentioned, the same brand's citation rate is 10.6%. That is a wide gap and it is the strongest single observation in this subject, but note what it does not settle. Mention may drive citation, citation may drive mention, or both may follow from the same underlying familiarity. Seer says as much itself, describing its work as behavioral evidence rather than proven architecture. Note also the sample: twenty brands, with Claude and Meta AI excluded because they return no citation URLs at all.

The practical consequence is that a dashboard counting only citations, or only mentions, will report success while the other half collapses.

Numbers that a product was launched on

Two figures circulate more widely than anything above and neither survives a look at its source. AirOps' offsite-signals research is credited with the claim that 85% of top-of-funnel AI visibility comes from domains a brand does not own, and with a companion figure that brands earning both citations and mentions are 40% more likely to stay visible across multiple AI responses. The AirOps pages reviewed on 29 August 2026 state the figures and the framing without publishing sample composition, date range or method. AirOps launched Offsite, a product built on precisely this premise, on 3 February 2026. Both numbers should be treated as vendor marketing until a methodology appears.

A third figure is more interesting and equally thin. Peec AI, which sells AI visibility monitoring, reported across 37,804 AI responses spanning ChatGPT, Gemini, Perplexity, AI Overviews and AI Mode that polished keyword-style prompts and casual conversational questions often surfaced the same brands. If it replicates, it undercuts a substantial amount of prompt-variation consulting. On one vendor study whose design is not stated in the coverage available, it is a hypothesis rather than a finding.

The pattern is worth naming. Generative engine optimization — the practice of shaping content and off-site presence for inclusion in AI-generated answers — has an evidence base composed almost entirely of research funded by the companies selling the remedy. Occasionally that cuts against interest, which is when it is most credible: a brand monitoring vendor reviewing this literature conceded in its own write-up that it contains no matched cohorts, no difference-in-differences, and no randomized placement experiments.

What the platforms say, which is nothing

This is the documentary finding and it deserves stating flatly rather than working around. No platform documents brand mentions, linked or unlinked, as an input to retrieval, ranking, citation or generation.

Google's AI features documentation, updated 10 December 2025, covers indexation and states that no special optimization is required; it says nothing about off-site brand text. OpenAI documents only that ChatGPT ranks results using multiple factors and that placement is not guaranteed. Perplexity says it searches authoritative sources and defines nothing further. Microsoft comes closest, and its guidance is about your own pages rather than other people's: the Bing Webmaster Guidelines rewritten on 26 February 2026 instruct that entity names should be clear and consistent, with no ambiguous references. That is guidance on naming yourself unambiguously, not a statement that other people naming you is an input.

Every claim in this subject that goes beyond that one sentence rests on third-party correlation or on inference from model architecture. Repeated searching on 29 August 2026 found no operator statement on brand mentions as a signal, at any date.

What would settle it

The experiment is not hard to describe and it has not been published. Take a randomized set of comparable brands, place mentions for the treated group on comparable third-party pages, hold everything else steady, and measure mention rate in AI answers before and after against untreated controls. A matched-cohort, difference-in-differences design of exactly this shape already exists in the adjacent literature — Ahrefs used one to test whether adding schema markup changed citation rates, and published a null result that cut against its own commercial interest. The design is available. Nobody has pointed it at mentions.

Until someone does, the honest position has three parts. The retrieval mechanism is sound and actionable: get named on the third-party pages retrieved for your category, and the name is available to the model immediately. The training mechanism is plausible and unmeasurable from outside, and should be labeled as inference every time it is used. The correlation, the only thing anyone has actually measured, is consistent with mentions mattering and equally consistent with mentions being a symptom of the brand size that produces AI visibility by other means.

One more constraint applies to everything here. Whatever produced a mention last month is not guaranteed to produce one this month: Semrush's weekly tracking caught ChatGPT cutting Reddit citations from about 60% of responses to about 10% inside a fortnight in September 2025, with no notice from the operator. Nothing in this subject is durable enough to build a fixed program around.

Frequently asked questions