AI brand mention: definition, and why it is not a citation

By 2026, asking an assistant before choosing a supplier is ordinary, and a quiet thing has been happening to your brand: it gets named in those answers, or it does not, in conversations you never see. No click, no referrer, no line in your analytics. That event is the AI brand mention, and it is routinely confused with the AI citation.

Just your name, said or not said, at the exact moment somebody was deciding. The two are not the same, and the difference decides what you can measure and what you can change. For the discipline that sits behind all of this, see our agence GEO pillar and the GEO (Generative Engine Optimization) definition.

AI brand mention: the one-sentence definition

An AI brand mention is any moment an AI engine such as ChatGPT, Perplexity, Claude or Google's AI features writes your brand name inside an answer. It counts whether or not a source is attached, because it measures presence in the answer, not attribution.

Read that second sentence twice, because it is the whole term. A brand mention is deliberately indifferent to sourcing. If an assistant is asked which firms handle a given need and it writes three names, all three brands were mentioned. Perhaps one of them got a little footnote link. Perhaps none did. The naming happened either way, in front of a reader who was asking because they intended to act.

The term exists precisely because the naming and the sourcing come apart so often. In classic search there was no gap to name: you were linked or you were absent, and the link was the visibility. An assistant breaks that equivalence. It can hold your brand in its answer while linking to nobody, and it does this constantly. So we needed a word for presence without attribution, and this is it.

Mention vs citation: what is the difference?

A mention is your name in the answer. A citation is your name with a source pointing back to you. Every citation contains a mention; most mentions contain no citation. The gap between the two sets is where brands misread their own AI visibility.

Think of it as one circle inside another. The outer circle is every naming. The inner circle is the small share of namings that come with a source. Our companion page on the AI citation covers that inner circle in depth, including how engines choose which sources to show. This page is about the outer one, and about the part of the outer circle that lies outside the inner: the mentions that name you and source nothing.

This is not a vocabulary quarrel. It changes what you conclude from the same evidence. A team looking only at referral clicks sees near zero and decides the assistants ignore them, when in fact they may be named regularly and simply never linked. The reverse error is just as common: a brand delighted to be mentioned everywhere assumes the pipeline is working, when nothing traceable ever reaches the site. Both readings are wrong for the same reason. They collapse two different events into one number.

Where does an AI brand mention come from?

A mention has two possible origins. Either the engine retrieved live pages from a search index and named a brand it found there, or it named a brand from its own training memory, without looking anything up. The origin decides how quickly you can influence it.

The retrieved case is the one most people picture, and it is well documented by the engines themselves. Google describes how its AI features draw on the web and link out to it [2], Anthropic documents Claude's web search and the display of the sources it used [3], and Perplexity publishes guidance on how its search retrieval behaves [4]. The pattern is the same across all three: fetch pages, then write from them. When a brand is named this way, it was named because a page said so a moment ago.

The other case is stranger and gets discussed far less. Ask an assistant a general question with no search triggered, and it will still name brands. Those names come from its training data, from what the web collectively said about a market long before the model shipped. Nothing was looked up. The model simply carries the association. This is why a well-known brand keeps surfacing in answers even when its own site is thin, and why a genuinely better competitor can stay invisible for a long stretch despite excellent pages.

I ran into this the slow way in spring 2026, testing real business queries by hand for SME clients. Two questions, same client, five minutes apart. The first triggered a search, and the client appeared, correctly sourced from a page we had reworked. The second stayed general, no retrieval, and the assistant named three competitors and not our client. Same brand, same day, same engine, opposite outcomes. Nothing was broken. The two questions were being answered by two different mechanisms, and we had only ever worked on one of them. That afternoon changed how I audit: the first thing I now establish is not whether a brand is named, but whether the naming came from retrieval or from memory, because the answer determines whether the fix takes weeks or a very long time.

The four mentions you will meet

Mentions are not all worth the same. In practice they arrive in four shapes: sourced, bare, comparative, and qualified. Knowing which one you are getting tells you what to fix.

• The sourced mention. Named, with a link back. This one is also a citation, and it is the only kind that can send a click. Best case, and rarest.
• The bare mention. Named, no source, no link. The most common by a distance. It carries real recommendation weight and gives you nothing to measure, which is exactly why it gets ignored.
• The comparative mention. Named alongside rivals, in a list or a comparison. Here position and framing matter as much as presence. Being third in a list of three is a mention, and not a good one.
• The qualified mention. Named with a judgement attached, whether flattering or not. An assistant that names you and then adds a caveat has mentioned you in a way that actively costs you. Presence is not automatically an asset.

That last shape deserves a beat of honesty. Counting mentions as if they were all wins is a mistake I see often. A count is not a score. Ten bare mentions in third place with a caveat attached are worth less than one sourced mention that opens the answer, and no dashboard that tracks volume alone will ever tell you that.

How do you earn AI brand mentions?

There is no trick here, and anyone selling you one is selling you something else. Mentions come from being consistently nameable, in both mechanisms at once. The levers are few and they are unglamorous.

• Be an unambiguous entity. One consistent brand name, described the same way everywhere, with clean structured data. A model that is unsure who you are will not risk naming you. This is the work of entity SEO, and it feeds the memory side more than anything else you can do on your own site. Our note on how Google's patents treat entities in LLM retrieval goes further into the mechanism.
• Write so a machine can lift you. Answer in the first sentence of a section, then develop. That extractable sentence is what gets retrieved and reused. The E-E-A-T signals underneath decide whether it gets trusted.
• Bring first-hand figures. The original academic work on generative engine optimization, a 2024 paper from teams including researchers at Princeton and the Allen Institute for AI [1], tested this across a benchmark of real queries: adding cited statistics and credible sources lifted visibility in generative answers by roughly 40 percent, while keyword stuffing did nothing at all (Aggarwal et al., arXiv, 2024). Your own numbers are what a model cannot invent and a rival cannot copy.
• Get named where your buyers already read. Memory-based mentions are built from the wider web, not from your site. Being written about elsewhere is the only real lever on that half.
• Mark your markup up properly. Structured data is not decoration here: Google's own introduction to structured data [7] sets out what the markup is for and how engines read it. Our read of the Ahrefs study on schema markup and AI citations covers what the data does and does not support.

If you want the measurement side first, which is the honest order, our how to run a GEO audit checklist lays out the method step by step, and the complete GEO audit method and scorecard gives the scoring grid we use ourselves.

Alexis Dollé

CEO & Founder of Cicero Studio

I test AI visibility the slow way, by hand, one real business question at a time, across hundreds of sites. The habit I would pass on is this one: never count a mention without asking where it came from. A name the model remembered and a name it just looked up are two different facts wearing the same clothes, and telling them apart is most of the job.

LinkedIn →

Why do brand mentions matter now?

Mentions matter because they are where buying research now happens invisibly. An answer names two or three brands and links to none, the reader forms a shortlist, and nothing about it ever reaches your analytics. Absence at that moment costs you, and you never get the invoice.

The stakes concentrate on precise questions rather than broad ones. Across the 4887 French keywords Cicero Studio has analyzed, 34% draw fewer than 100 searches a month (Cicero Studio internal data). Those are not scraps. They are the specific, high-intent phrasings someone types the moment before choosing, and they are exactly the questions an assistant answers in two sentences naming two brands. Being one of the two is not a traffic play. It is being in the room.

The pattern I see is remarkably consistent. Across the 1210 SEO and GEO audits Cicero Studio has produced (Cicero Studio internal data), the recurring gap is almost never a shortage of content. It is content that a machine reader cannot cleanly lift or trust, published by a business that is convinced it is already doing the work. The question is rarely whether to write more. It is why nothing already written can be quoted, and whether the market has ever heard the name in the first place. Those are two different problems, which is the argument this whole page has been making, and they are the two that a serious look at your AI visibility has to separate before anything else.

What an AI brand mention is not

A definition earns its keep at the edges, so here is what this one does not cover. These are the misreadings I correct most often, and each one quietly burns money when it goes unchallenged.

One honest note to close on. This page reflects what I observe in the field and the public research available when it was last updated, on 17 July 2026. The engines do not publish how they weigh a name, and the memory side in particular is opaque to everyone outside the labs, myself firmly included. Where handling personal data enters the picture, the CNIL's guidance on AI is the reference in France. I would rather mark the edge of what I know than sell a certainty that does not exist.

Going further

We document the method in the open, because that is the best proof we have. Each resource below takes one angle further, from the citation mechanics to the specific assistants. Pick whichever matches the question you are actually holding.

Sources

Every factual claim above is tied to a named, dated source below. The engine documentation covers how retrieval works, the academic work covers what earns reuse, and the regulators cover the ground rules.