Blog · Diagnosis

Why AI never mentions your brand

Published 30 March 2026 · Updated 12 October 2026 · 9 min read

Why AI never mentions your brand
Photograph: Mark Thompson / unsplash

The short answer: An assistant leaves you out because you were never a candidate, not because it read you and preferred somebody else. Four things cause that. The engine could not fetch or quote your page, it could not tell your company apart from a dozen similar names, it was answering from memory rather than from the live web, or nothing it already trusts repeats your claim. Each of those has a different fix, and guessing wrong costs a quarter.

The uncomfortable part is that all four look identical from where you sit. You ask the question, the answer names three other companies, and nothing tells you which gate you failed at. So the work starts with separating them.

The answer is finished before you are considered

Watch what happens between the question and the paragraph. Google documents that AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources to develop a response. That is Google describing Google's surfaces. The rest of the chain, documents coming back, being sliced into passages, and a few surviving passages becoming what the model writes from, is the working model we use for engines that publish little about their retrieval. It is not behaviour they have confirmed.

On a surface that works that way, your brand has to clear every stage to appear. Rank well on the original phrasing and you may still miss, because the fan-out asked something you never wrote about. Get retrieved and you may still miss, because the passage that came back was your navigation and a cookie banner. By the time the model is choosing words, the candidate set is already closed, and nothing about your positioning or your design gets a vote.

This is why "we rank third for that keyword" and "we are absent from the answer" sit together so comfortably. They are measurements of different machines.

The rule you set in 2023 that is still disqualifying you

Here is a failure we meet repeatedly in our own audits, and it is rarely on anyone's list. Our audits are a few dozen engagements, so read what follows as a pattern we keep running into, not as a rate measured across the web.

Google states plainly that a page must be indexed and eligible to show a snippet before it can appear as a supporting link in AI Overviews or AI Mode. Its robots meta documentation then says that nosnippet prevents content being used as a direct input to those features, and that max-snippet limits how much of it may be used. Plenty of sites added one of those tags years ago to stop aggregators lifting their copy. The page still ranks. The page is still indexed. On Google's AI surfaces the page is barred from the answer, and the rank and index reports we are usually shown at the start of an audit do not surface it. Both pages were checked on 12 October 2026 and say this today; other engines document their own snippet controls separately, and this one is Google's.

The second version of this lives in robots.txt. OpenAI documents separate agents with separate jobs, among them GPTBot for model training, ChatGPT-User for user-initiated fetches, and OAI-SearchBot for surfacing sites in ChatGPT's search features. The documentation is explicit that a site opted out of OAI-SearchBot will not be shown in the search answers themselves, while remaining available as a navigational link. Teams that wanted to opt out of training sometimes block all of them in a single ticket and remove themselves from a surface they meant to keep. That is a pattern we have seen in our own audits rather than a frequency we have measured, and these are OpenAI's agents under OpenAI's rules. Other engines publish their own agents, with their own names and their own effects.

Blocking is common enough to be worth checking rather than assuming. Cloudflare Radar data from 6 June 2025, published by Cloudflare on 1 July 2025, found that of the top 10,000 domains, 3,816 had a robots.txt, and 546 of those, about 14 percent, carried allow or disallow directives aimed specifically at AI bots. GPTBot was the most frequently disallowed, by 312 domains, 250 fully and 62 partially. That is one snapshot of large domains rather than of the web, and it cannot say how many of those lines were still intended by the time they were counted.

Two checks, fifteen minutes. Fetch your money pages and grep the response for nosnippet and max-snippet, in the meta robots tag and in the X-Robots-Tag header, which is where a CDN rule hides. Then read robots.txt as a stranger would and list every user agent you disallow, next to what that agent actually does. The self-check has the rest of the access list.

Memory and retrieval are two different sources

Ask an assistant that lets you switch web access off about your category, and the answer can differ from the one you get with retrieval on, because it is drawn from what the model absorbed during training. A brand that is absent in one mode and present in the other is telling you something, and most teams never run the comparison because they only ever test one mode. Some assistants do not expose that toggle at all, which limits where the pair can be run.

The readings below are how we interpret that pair in our own audits, not rules an engine has published. Absent with retrieval on and present with it off has usually pointed to an access or parsing problem: the live fetch is failing, or the page renders in the browser and arrives empty. Present with retrieval on and absent with it off tends to mean a recent arrival the training data has not met, which is normal and fixes itself slowly. Absent in both has been an identity problem, and it is the expensive one. Retrieval vs training data: which one is answering about your brand sets out how to run the pair cleanly.

Engines differ here more than the category admits. When we have put the same brand question to Gemini, ChatGPT and Perplexity on the same morning, we have got three answers built from three different source sets. That is an observation from a handful of prompt sets on single days, not a measured rate, and it is why we treat one engine's silence as weak evidence on its own. How Gemini answers brand questions covers one of those pipelines in detail.

You are a string, not an entity

Retrieval works on resolution. For a claim to attach to you, the company in your footer, the company in your schema markup, the company in the funding story and the company in the review roundup have to be treated as one organisation. Share a name with a clinic in Ohio, rebrand eighteen months ago, run two legal entities across three domains, and that gets harder. We have not found published documentation of how that resolution is performed, so this is a model we work from, consistent with Google's advice to keep site and organisation details consistent, rather than behaviour we can observe directly.

A model that cannot resolve you appears to hedge, and a hedge reads as absence. It writes "several providers offer this" instead of naming you. That is our reading of the output, not something we can see inside the system, and your analytics holds no record of it either way.

Identity breaks hardest during change, which is why a replatform or a domain move is a trigger we meet repeatedly for a brand going quiet in answers. The redirects can be perfect and the citations still fall, because the corroborating pages elsewhere on the web now point at a URL that no longer exists in the engine's cached view. Migrating a site without losing citations is about that window specifically.

Nothing the engine trusts repeats your claim

The last gate is the one you do not own. Judging by the sources engines actually show alongside their answers, they lean on material that already carries weight: documentation, encyclopedic references, trade press, forums with real threads, comparison pages written by somebody else. If your claim exists only on your own domain, it is a marketing statement. If several independent places say it, an answer has something to rest on. We cannot see the thresholds involved, so treat that as a direction rather than a count.

Note the asymmetry. We have watched a brand mentioned in a widely read comparison article get pulled into answers for queries that article never targeted, because the passage is what travels, not the page. That is slow, unglamorous work, and it is the reason corroboration takes quarters while access fixes take days.

Running the diagnosis in order

Do it cheapest first. Confirm the page is indexed and snippet-eligible. Read robots.txt and the snippet directives. Fetch the page with JavaScript disabled and check that the answer to the question is actually in the HTML. Run twelve buyer questions, three times each, across at least three engines, with retrieval on and off, and record only three things: were you named, was the sentence true, and which sources got cited. Patterns show up fast. Absent everywhere in every mode is access or identity. Present on awareness questions and absent on comparison questions is corroboration. If you were visible last quarter and are not now, diagnosing a drop in AI citations is the narrower version of this procedure.

Honest limits

The four gates above are a diagnostic frame, not a published description of how any one engine works: only the Google and OpenAI behaviours cited here are documented, and we have extended them to engines that document nothing. Our run counts are small and answers are unstable, so a brand appearing in two runs out of nine is not reliably a brand that moved, and we treat anything under roughly a third of runs as noise until a later sample agrees. The patterns we describe from our own work come from a few dozen audits rather than a sample designed to represent anything, and we have not measured how often each gate is the one that fails. The Cloudflare figures are a single June 2025 snapshot of large domains, so they cannot be read as the state of the web now. We also cannot connect a mention to revenue, because the referrers these surfaces pass are too sparse to attribute from, which means a conversion rate quoted off AI answers is modelling rather than measurement.

FAQ

Is being absent from AI answers a content quality problem? Sometimes, and less often than people assume. Access rules, rendering and identity account for most of the absences in the audits we have run, which are a few dozen engagements rather than a representative sample, and all three can coexist with genuinely good content.

Will adding more structured data get me mentioned? Google's AI features documentation says there are no additional requirements and no special optimisations needed to appear in AI Overviews or AI Mode. That is Google's statement about Google's own surfaces. Schema may still help an engine resolve who you are, which matters for identity, but we have no evidence it is a lever that inserts you into an answer.

How many runs before I believe a result? Three runs of the same question on the same engine, at minimum, and preferably on separate days. Single runs are anecdotes, and two matching runs can still be coincidence.

My competitor is named and we are not. Is that about their budget? More often it is about where they are mentioned. Check which sources the answer cites, then check whether you appear anywhere in those sources at all.

Sources

All four were opened and checked on 12 October 2026.

Related reading

Diagnosing a drop in AI citationsMigrating a site without losing citationsRetrieval vs training data: which one is answering about your brandHow Gemini answers brand questionsHow to check what engines say about you, properlyContradictions across your own pages