AI brand mentions

An AI brand mention is an occurrence of a brand's name in the text an AI assistant generates in reply to a prompt. It is a property of the generated answer, not of a search results page, and it exists whether or not the answer links to anything.

That distinction is the whole reason the term needs its own definition. Three things get collapsed into "visibility" and they are separately measurable: the brand appearing in the prose (a mention), a source document being linked (a citation), and someone arriving on your site afterwards (a visit). An answer can carry any one of them without the other two.

Where a mention comes from

Two different mechanisms produce one, and they fail differently.

A parametric mention comes from the model's weights. The brand was frequent enough in training data that the model associates it with a category, and it writes the name without fetching anything. Nothing you publish today changes it, and the model may attach stale attributes — an old price, a discontinued feature, a founder who left.

A retrieved mention comes from documents fetched while the answer is being written. This is the RAG path: a search step returns pages, some enter the context window, and the model writes from them. Here your current page is eligible, and the mention can be traced to a document.

Only the second is reachable by publishing. Only the second usually carries a link. And the second has a hard prerequisite: the agent has to be allowed to fetch. OpenAI documents OAI-SearchBot as the agent that surfaces websites in ChatGPT's search features, and states that sites opted out of it will not be shown in ChatGPT search answers — a robots.txt line is therefore a mention-rate control, with roughly a day of lag before it takes effect. See robots.txt.

How mentions are measured, and what that costs in accuracy

Almost every tool in this space measures the same way: put a fixed set of prompts to a set of models on a schedule, store the answers, and match your brand name against the text.

The matching step is where the error lives. Attensira's own published limitations describe its detection as a case-insensitive substring match of the brand name or domain against the answer, with no entity resolution and no word-boundary check. The consequence is systematic rather than random: a brand named Notion matches "notional", a brand named Arc matches "march" and "search", and averaging does not remove it. If your name is a common English word, read the stored answers before quoting your own number.

Two further properties follow from how the data is collected:

• Models are not deterministic. The same prompt to the same model twice can produce two different answers, so a rate computed from one draw per prompt is thin. Sampling more draws per reading is the only fix; declaring a change without a significance test is not.
• Not measured is not zero. A model that was never queried, or a surface with no available reader, produces an absence. Drawing that as a measured zero invents a rejection that never happened.

Failure modes worth naming

1. Reading a mention rate as a share. Rates for you and your competitors run over the same denominator of answers. One answer naming four brands raises four rates. They can sum past the whole because they are not slices of anything.
2. Treating position as a rank. Where a URL sits in a model's citation list is a property of that citation list, not a leaderboard placement against competitors.
3. Counting mentions without reading them. A mention can be an unfavourable one, or about a different company with your name. The count alone cannot tell you which.
4. Attributing revenue. Nothing links the answer a person read to the session they later started. That connection is drawn by you, with your analytics, and should be stated as an inference.
5. Chasing the parametric path. No amount of publishing this quarter changes weights fixed last year. The addressable surface is retrieval — see GEO and AI search.

What actually moves a retrieved mention

Be fetchable by the named agents, publish passages that answer the prompt in the passage itself, and keep the facts a model would state about you — pricing, positioning, category — literally present as text on pages that are cheap to fetch. A model can only name you with attributes it can find; if the correct attribute is only implied, the stale one in its weights wins.

Frequently asked questions

Is a mention a citation?

No. A mention is the name in the prose; a citation is a link to a source. Either can appear without the other.

How is a mention detected?

Usually by string matching against the stored answer text, which is why common-word brand names carry noise.

Can I be removed from AI answers?

From ChatGPT's search answers, yes — OpenAI documents that opting out of OAI-SearchBot excludes a site from them.

Do mentions sum to a market share?

No. Competitor rates share a denominator and can total more than the whole.