Why does ChatGPT recommend the wrong business?

Three of five assistants named the wrong company

In August 2026, First Mention put its own name and domain to five AI assistants, ChatGPT, Claude, Perplexity, Gemini and Google's AI Overviews, and asked what the business was and who ran it. Three of the five answered about a different company. They resolved firstmention.co.uk to firstmention.co, an American home-services SEO firm with no connection to Devon or to this consultancy, and described that firm's work as ours. One went further and attributed that company's published content to First Mention.

Nothing about this looked like a fault. The site was live, indexed and perfectly readable. The assistants had two plausible candidates for one name, and the evidence attached to the American company was older and more densely linked, so they chose it. A model does not tell you it is uncertain. It writes a confident paragraph about whichever entity it settled on.

The reason it happened to us is not mysterious. First Mention is new. The domain is new, the external record behind it is thin, and firstmention.co has been trading and publishing for years. Given two candidates for one name, the assistants took the one with more evidence attached to it, which is the reasonable thing to do on the information available. An SEO consultancy finding this on its own domain is uncomfortable to write down, and it also makes the point better than a borrowed example would: the fault has nothing to do with the quality of the site, only with a thin record standing next to an established namesake. Every new business, every rebrand and every company that shares a name with a bigger one starts in the same place.

The failure was not confined to one product, and the two assistants that answered correctly cannot be relied on to keep doing so, because the same question can resolve differently on a different day.

One clarification belongs in plain text on this page, because it is the fix the rest of the article describes. First Mention is a search and AI search consultancy based in Devon, England, trading at firstmention.co.uk. It has no connection to firstmention.co, a home-services SEO firm in the United States.

That is the whole problem in one example. The failure is silent, it happens before any question about quality or ranking is reached, and the business at the wrong end of it has no way of noticing during ordinary trading. It is also a diagnosable fault with identifiable causes: a near-identical domain, a name that is a common English phrase, and a thin external record.

What is entity resolution?

Entity resolution is the step where an assistant decides which real thing your name refers to. Before a model can say anything about a business, it has to match the words it has been given to one entity in its picture of the world, then retrieve the facts attached to that entity.

Google names two mechanisms behind its own generative features: retrieval-augmented generation, sometimes called grounding, which means the model fetches supporting material rather than answering from memory alone, and query fan-out, where one question is broken into several sub-queries (Google Search Central, published 15 May 2026, updated 10 July 2026). Both depend on the resolution step. Fetch material about the wrong entity and every sentence built on it is wrong, however well written.

Consider two firms trading as Ashworth Legal, one in Plymouth and one in Manchester. A customer asks an assistant about “Ashworth Legal's conveyancing fees”. The model has to pick. It will weigh which one has clearer, more consistent, more corroborated evidence attached to its name, and answer about that one. The Plymouth firm never sees the question and never sees the answer.

How does entity resolution break?

Entity resolution breaks through ambiguity, not through error. Four causes account for most of it: inconsistent naming, unclear location, markup that disagrees with the page, and a thin external record.

Take a builders' merchant near Tavistock registered as Tamar Building Supplies Ltd. The website header says Tamar Builders Merchants. The Google Business Profile says TBS Tamar. An old directory entry says Tamar Supplies. A human reads four labels for one yard. A model reads four candidate entities, each with a fragment of the evidence, none of them strong.

Location does the same damage. A law firm that says only “based in Plymouth” has told an American-trained model very little, because Plymouth is also a town in Massachusetts. The fix is unglamorous: the visible text says Plymouth, Devon, England, and says it in a place a crawler will reach.

Then there is markup. Structured data, which means machine-readable code describing what a page is about, is only useful when it is generated from the words visible on the page. Markup naming one organisation while the page names another is a contradiction, and a contradiction is worse than silence. Google's guidance is explicit that no special schema is required for its AI features and warns against over-focusing on structured data, so the value here is in agreement between the code and the text, not in volume of code.

The fourth cause is quieter than the other three. If nearly everything a model can find about a business sits on that business's own website, there is nothing to corroborate it, and a claim that appears in one place carries less weight than the same claim appearing in several. This is where an older, better-linked namesake wins: not because its evidence is better argued, but because there is more of it, in more places, agreeing with itself.

Why doesn't a ranking report show this?

Because a ranking report measures a different thing. It answers the question “where does this URL appear for this query”, and an entity failure happens on questions that never involve your URL at all. A dental practice can hold first place for “dental implants Exeter” all month and be absent from every AI answer about implant dentists in Exeter, with a rank tracker showing green throughout.

The two are not alternatives, and both sit on the same foundations: a page has to be indexed and eligible for an ordinary snippet before it can appear in a Google AI answer at all. The difference is that one of them is measured by the reports most businesses already receive, and the other is not measured by anything unless somebody sets out to measure it.

How do you check whether it is happening to your business?

By asking, repeatedly, and writing down what comes back. There is no report to request and no dashboard that surfaces an entity failure, so the check is a deliberate one: a fixed set of questions, asked across several assistants, run enough times to distinguish a pattern from a one-off answer.

The repetition matters because the answers move. The same question put to the same model twice can name different businesses. Repeat runs are the method, not an optional extra: five answers to the same question give a proportion that can be recorded and compared month to month, where one answer gives an anecdote. First Mention runs a fixed query set monthly and asks each query five times within the run.

Several assistants matter for the same reason. ChatGPT is the largest single destination by some distance, with 1.8 billion UK visits in the first eight months of 2025 against 368 million in the same period of 2024, but it is not the whole picture: over the year to August 2025 Ofcom recorded Gemini up 146%, Claude up 138% and Perplexity up 100% (Online Nation 2025, published 10 December 2025). Checking one assistant tells you about one model's picture of your business, and the pictures differ.

For a hotel in Torquay, the query set would cover the obvious customer questions (“family hotels in Torquay with parking”, “is [hotel name] dog friendly”, “who runs [hotel name]”) and each answer would be recorded against four things: was the business named, was it named accurately, was it placed in the right town, and were the facts attached to it actually its own. Those are four separate failures, they have separate causes, and each one needs different work.

What does fixing it involve?

Making one entity easy to identify and hard to confuse. In practice that means one canonical form of the business name used everywhere, the same wording in the header, the footer, the Google Business Profile, Companies House and every directory entry that can be reached and edited.

Location and country go into visible text instead of being treated as obvious. Organisation markup is generated from that visible text, so the two never disagree. External profiles need to be verified and not merely present, because corroboration from a source the model already trusts does more than another sentence on your own site.

The work also includes checking that AI crawlers are allowed in. Crawlers are the automated programs that fetch pages, and robots.txt is the small file telling them what they may take. GPTBot, PerplexityBot and Google-Extended are frequently blocked by a platform or host default that nobody chose, and blocking them removes the business from those assistants entirely. This is a publishing decision rather than a security measure, and it is worth knowing which way it has been set.

One thing that will not help: llms.txt. Google states it does not use these files, and lists them among the machine-readable extras site owners can ignore for Google Search. A supplier presenting one as an AI visibility route is selling something the search company has published as ineffective.

What makes an assistant cite you once it has identified you?

Facts it can lift. Correct resolution only gets a business considered, and what turns consideration into a citation is whether the page carries specific, plainly worded, attributable information rather than description.

The first controlled experiment on this, by Aggarwal and colleagues at ACM SIGKDD in August 2024, tested 10,000 queries and found that citing sources, adding quotations and adding statistics each raised visibility by 30 to 40% on position-adjusted word count. The 40% is a maximum, not an average. The same study found keyword stuffing scored roughly 8% below doing nothing, which is a useful result: a long-standing search habit actively harms AI visibility.

A more recent preprint by Yu and colleagues, published 31 March 2026, held the wording identical and changed only structure, hierarchy and chunking, and recorded citation rates rising 17.3% across six generative engines. Two limits apply: it is an arXiv preprint that has not yet been peer reviewed, and the abstract does not name the six engines.

Query fan-out is what makes this practical. A guest typing “family hotel near Salcombe with parking and a dog-friendly room” is not running one search. The assistant may run four, covering parking, dog policies, family rooms and the area separately, then assemble one answer. For the hotel, that means the parking arrangements and the dog policy each need to exist as findable, plainly worded text, because each may be retrieved by a different sub-query. A PDF tariff and a paragraph implying dogs are welcome will not survive that process.

Questions people ask

More answers are on the First Mention FAQ hub.

← Back to all guides