Why isn’t ChatGPT mentioning my brand?

The instinct when an assistant does not name you is to assume it has formed a view. It almost never has. In the large majority of cases the model has never successfully read a page you published, or has nothing consistent enough about you to state with confidence.

Absence is an information problem far more often than a reputation problem. Seven causes explain nearly all of it, and they are not equally likely.

The first check

Before diagnosing anything, establish that absence is real rather than sampled.

Generated answers vary between identical runs. Ask the same question three times and you may get three different sets of companies. A single answer that omits you proves nothing at all.

Write fifteen to twenty questions a real buyer would ask about your category — not keywords, full conversational questions — and run each at least twice across ChatGPT, Gemini and Perplexity. Record who is named and what is cited. Consistent absence across that set is a finding. One missing answer is noise.

1. Your content is not readable

The most common cause by a wide margin, and the one almost nobody checks.

Googlebot renders JavaScript in a headless Chrome engine. AI crawlers do not. An analysis by Vercel and MERJ covering more than 500 million GPTBot fetches recorded zero JavaScript execution — GPTBot downloaded JavaScript files roughly 11.5% of the time and never ran them, with identical behaviour from ClaudeBot and PerplexityBot. They fetch raw HTML under tight timeouts, take what is there, and leave.

A page can rank in the top three on Google and show GPTBot nothing but an empty shell.

The check: disable JavaScript in your browser and load your ten most important pages. Five minutes. If the answer text is not there, this is your diagnosis and nothing else on this list matters until it is fixed.

2. Your crawlers are blocked

Second most common, and usually nobody decided to do it.

Check robots.txt first, then — more importantly — your CDN or WAF bot-management settings. Many platforms ship presets that deny AI user agents by default, so the block is inherited from a security configuration rather than chosen as a policy.

The check: request your homepage from outside your own network with a GPTBot user agent string and confirm a 200 response. A 403 here explains everything downstream.

3. No entity clarity

Models reason about entities — brands, products, categories — rather than matching strings. To name you as an option in a category, a model needs to know unambiguously that you are in that category.

Most brand pages fail this badly. They describe a mission and a feeling and never once state the category noun a buyer would use. A model reading that page cannot connect it to the question being asked.

The check: read your homepage and about page as a stranger. Can you tell, in one sentence, what category this company competes in and which alternatives it sits against? If not, write that sentence and put it somewhere permanent and crawlable, then mirror it in your Organization schema.

4. No third-party corroboration

This is the cause that surprises people who have done everything else right.

A model weighs what independent sources say about you far more heavily than what you say about yourself, and unlinked mentions count. Reviews, comparison posts, industry roundups, forum threads and news coverage all shape what a model believes — and Reddit is cited disproportionately often.

It is common to find a competitor ranking below you on Google being named first by ChatGPT, purely because they are described consistently in a dozen places you are absent from.

The check: look at which URLs the assistants cite when answering your category’s questions. Those cited domains are the corroboration map. If you appear on none of them, that is the gap, and it is more actionable than any amount of on-site work.

5. You are testing the wrong prompts

Sometimes the brand is visible and the test is not.

People test prompts phrased like keywords — “best CRM software” — which invite the model to name the largest, most-discussed players and nobody else. Real buyers ask longer, more constrained questions: “what CRM works for a 12-person nonprofit that needs Xero integration”. Those are the prompts you can realistically win, and frequently you already do.

The check: rewrite your prompt set to include the constraints your actual customers have. If you appear on the specific ones and not the broad ones, you are not invisible — you are correctly positioned in a niche, and the strategy is to widen from there rather than to attack the head-on question.

6. Your pages are unquotable

You are retrieved and then not used. This is a stage-two failure and it looks identical to stage-one failure from the outside.

Models lift passages, not pages. A page that answers six questions vaguely, buries its answer under four paragraphs of preamble, or hedges everything into mush gives the model nothing clean to extract — so it quotes someone else who wrote two clear sentences.

The check: copy three sentences from the middle of your key page and read them cold. Do they answer something on their own? If they depend on the previous paragraph or a pronoun with no antecedent, they cannot be lifted.

7. You are genuinely too small

Occasionally the honest answer is that there is not yet enough about you anywhere for a model to say anything with confidence.

This is not permanent and it is not a judgement. It means the corroboration work — getting accurately described in the places your category already gets cited from — is the whole job, and on-site optimisation will not substitute for it.

The productive move is to be specific rather than broad. A new company will not be named for “best project management software”. It can absolutely be named for the narrow, constrained version of that question where three sources describe it accurately.

Working the causes in order

The seven causes are not independent, and fixing them out of order wastes effort. A practical sequence:

1. Causes 1 and 2 first, because they are binary and cheap. If your content is unreadable or your crawlers are blocked, everything else is unmeasurable — you cannot tell whether your writing is unquotable when nothing has read it.
2. Cause 5 next, because it costs nothing and occasionally dissolves the whole problem. Rewriting the prompt set to match how buyers actually ask frequently reveals you were visible all along on the questions you can win.
3. Causes 3 and 6 together, since entity clarity and quotability are both editorial and both fixed in the same pass through your key pages.
4. Cause 4 last and continuously, because corroboration is the highest ceiling and the longest timeline, and it is wasted effort if a crawler still cannot read the site it points at.

How to tell which stage you are failing

There is one diagnostic that separates retrieval failure from selection failure cleanly. Ask an assistant a question so specific that only your page could answer it — your own product name plus a detail documented nowhere else.

If it answers correctly, your content is being retrieved and read. Your problem is selection or corroboration: you are in the candidate set and losing. If it cannot answer, or invents something, your content is not reaching the model at all and the problem is mechanical.

That single test saves a great deal of guessing, because the two failure modes look identical from the outside and their remedies share nothing.

When it mentions you but gets it wrong

A distinct and common finding: the assistant names you and states outdated pricing, a feature you removed, or a positioning you abandoned two years ago.

This is more tractable than absence. The model is drawing on something — usually an old page of yours, a stale third-party listing, or a review written against a previous version of the product.

Three fixes, in order. Publish the correct fact on a structured, crawlable page of your own, stated plainly enough to be extracted. Update or request correction on the third-party sources that carry the stale version, starting with the ones the assistants actually cite. Then re-scan on a schedule, because correction propagates at the speed of recrawling rather than immediately.

Whichever cause applies, the diagnosis is worth doing before the work. The remedies for cause one and cause four share nothing at all, and choosing between them by instinct is how a quarter gets spent on the wrong problem. Our guide to which changes to make covers the fixes in order of leverage.

Sources

• Vercel and MERJ — analysis of 500M+ GPTBot fetches finding zero JavaScript execution; same behaviour observed for ClaudeBot and PerplexityBot.
• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande — “GEO: Generative Engine Optimization”, ACM SIGKDD 2024 (arXiv:2311.09735).
• OpenAI documentation — GPTBot crawler identification and robots.txt directives.

Frequently asked questions