Why ChatGPT doesn't mention your brand — and the 6 fixes that actually work

· by boneLau

Short answer: ChatGPT doesn't mention your brand because it has no independent, machine-readable evidence that your brand is an answer to the question someone asked. It is not a ranking problem, and it is usually not fixed by writing more pages on your own site. What fixes it is a third-party source that says what you do, using the words your buyers actually type, plus a sentence on your own site that an AI can lift verbatim.

Everything below comes from running brands through six AI assistants — ChatGPT, Perplexity, Gemini, Qwen, GLM and Kimi — every week and recording what each one answers. (We now also cover Google AI Overviews; every number quoted below predates that addition.) Where we quote a number, the sample size and the date are in the Methodology section at the end. Where we don't know something, we say so.


First: this is not a ranking problem

Search engines rank pages. AI assistants assemble answers. The two use overlapping inputs but they fail differently:

Classic searchAI answer
What's retrievedpages that match the querypassages that could support an answer
What decides the outputranking orderwhat the model can ground the answer in
Why you're absentyour page ranks too lownothing in the retrieved context says you're relevant
What usually fixes itlinks, on-page relevancean independent source + a liftable definition

That's why "we rank #3 for our category keyword but ChatGPT has never heard of us" is a normal, not contradictory, situation. It also explains the most common failed fix: writing yet another page on your own domain. If the model doesn't already treat your domain as the authority for the category, a new page on that domain inherits the same problem.


"Why isn't my brand mentioned in ChatGPT?" — the short answer

People ask this in slightly different words — why isn't my brand mentioned in ChatGPT, why is my brand not showing up in ChatGPT, why are brand mentions missing from ChatGPT responses. The answer is the same in all three cases: the model has no retrievable, quotable statement that your brand is an answer to that question. It is not withholding you; nothing in the material it assembled says you belong there.

Two consequences follow. First, re-asking the same question in a brand-new chat rarely changes anything, because the answer is assembled from indexed material rather than from the conversation. Second, the fix is not "tell ChatGPT about your brand" — there is no way to do that. The fix is to put a statement where the retrieval step can find it: an independent page that names your category, your use case, and your domain in one quotable sentence.

The test is one line: ask five of your buyer questions without naming your brand. If the answer never mentions you, you have a retrieval gap, not a description problem — and those two need different fixes. The exact 20-minute setup is here.


What we measured

We run a fixed question set — discovery ("best X for Y"), comparison ("X vs Y"), objection ("is X any good") — against six AI assistants, every week, and record the raw answer, whether the brand appears, the position, and any source URL the model returns. (The runs quoted in this article predate our addition of Google AI Overviews.)

Three findings shaped the fixes below.

1. The same brand gets a different answer — and different sources — per engine

A single run for one brand (2026-08-24, 5 questions × 6 engines):

EngineMentionedWhat it cited
ChatGPT5/54 answers cited third-party pages (a community site, a directory)
Perplexity5/5every answer cited the brand's own site
Gemini5/5cited sources, but the grounding URLs were masked
Kimi1/5the one mention was "no information about this brand"
GLM0/5nothing
Qwen0/5nothing — despite having mentioned this brand 16 times in earlier runs

Two engines can be answering from completely different evidence. So "does AI mention me?" is the wrong question; "which engine, on which question, citing what?" is the question that leads to a fix.

2. "Knows you, then forgets you" is a real state

Qwen had mentioned that same brand 16 times across earlier runs — then returned a clean zero. Nothing changed on the brand's side. Retrieval is not a permanent memory: it's a snapshot of what the model could find that day.

The practical consequence: a single before/after test can convince you that a fix worked, or that it failed, when neither is true.

3. Run-to-run noise is roughly ±10 percentage points

On mention rate, we measured about ±10pp of natural variance between runs of the same question set. A move inside that band is noise, not news — including the move you're about to celebrate after publishing one article.

This is the reason we compare each week against a baseline built from the previous six runs, and why the acceptance rule we use is multiple runs until two agree.


The 6 fixes

1. Measure the baseline before you change anything

You cannot fix an unknown state. Pick 5 questions your buyers would actually ask (not your taglines), run them across the assistants that matter to you, and record: mentioned yes/no, which position, which sources cited.

How to verify: save the raw answers. You'll need them in three weeks to tell a real move from noise (see fix 6).

2. Describe yourself in category words, in one liftable sentence

Models answer category questions ("best X for Y"). If your homepage says "We help ambitious teams unlock growth", there is no sentence connecting you to a category — so there is nothing to retrieve.

Write one plain sentence, in your first paragraph, of the form: "[Brand] is a [category] tool for [who], that [does what]." Then use the same category words in your title, H1 and pricing page. Consistency across your own pages is the cheapest fix on this list and the one most often skipped.

How to verify: ask an assistant "what is [your brand]?" — a good answer will reuse your category words rather than inventing a different one.

3. Let a third party say what you do (this is the actual fix)

Your own site is structurally the weakest citation for an AI answer: it's the one source that always has a stake in the answer. In our runs, the engines that mentioned brands most reliably were citing third-party pages — community threads, directories, review sites, comparison articles.

So the work is: get an independent page to state your category and use case in its own words. Practical versions that actually get indexed and retrieved:

  • a listing on directories that AI assistants already retrieve (they show up in citations constantly)
  • a genuine comparison or "alternatives to X" article on a third-party site
  • a review or community thread where a real user describes the use case
  • a guest piece or data contribution where someone else publishes your numbers

How to verify: after a few weeks, look at the cited sources in your report. If citations are still 100% your own domain, nothing has changed no matter what your mention rate says.

4. Fix per engine, not in general

Different engines lean on different evidence. Treat them as separate channels:

EngineWhat we typically seeWhere to spend effort
Perplexitycites the brand's own site heavilyon-site clarity: definition sentence, pricing, comparison pages
ChatGPTcites third-party pagesdirectories, community threads, third-party writeups
Geminicites broadly, sources often maskedbroad coverage + structured data
Qwen / GLM / Kimipatchy, category-dependent, unstablelanguage- and region-appropriate sources; re-measure more often

A single "AI visibility" number hides this. A per-engine view tells you where to work this week.

5. Build a second source, not a second landing page

The instinct after not being mentioned is to rewrite the landing page. In our data, that rarely moves citations, because the missing input isn't your description of yourself — it's someone else's. Budget one hour a week for the third-party work in fix 3 and you'll out-produce a month of on-site rewriting.

6. Re-run and require agreement before you conclude anything

Given ±10pp noise and "knows you, then forgets you", a single measurement is an anecdote. Run the same question set at least twice (ideally on different days) and only act when the two runs agree. We use this rule on our own product; it has saved us from shipping "fixes" for noise at least twice.

How to verify: keep a log of runs per question per engine. The pattern — not the point — is the signal.


What to measure (and what not to)

MetricHow to get itWhy it matters
Mentioned per engine per questionyour own run logtells you where the gap is
Cited sources (own vs third-party)the source URLs in answerstells you why the gap exists
Accuracy of the descriptionread the answersa wrong mention is worse than none
Indexed pages, impressionsGoogle Search Consolethe search-side baseline
Visitor UVyour analyticswhether any of this reaches humans
Signups / leadsyour own DBthe only number that pays for the work

Don't measure: a single composite "AI score" with no per-engine breakdown, and any before/after taken from one run.


FAQ

How long does it take to get mentioned by ChatGPT? Indexing and retrieval of a new third-party source typically takes weeks, not days. Treat 4–8 weeks as the first honest checkpoint, and measure with repeated runs rather than one.

Do I need to publish more content on my own site? Usually not first. On-site clarity (fix 2) is necessary but rarely sufficient. The missing input is normally a third-party source (fix 3).

Can I pay to be mentioned? No. There is no ad slot in an AI answer. What looks like "paid placement" is usually a well-indexed third-party page.

Does llms.txt help? It's cheap and harmless, and it helps assistants understand what your site is — but do not expect a ranking or citation lift from it alone. Treat it as documentation, not distribution.

Why does one engine mention me and another doesn't? Because they retrieve from different evidence. Fix them separately (fix 4) rather than averaging them into one number.


Methodology

  • What we run: a fixed set of 5 questions (discovery, comparison, objection) across six assistants — ChatGPT, Perplexity, Gemini, Qwen (Alibaba), GLM (Zhipu), Kimi (Moonshot) — via OpenRouter.
  • What we record: the raw answer text, whether the brand appears, its position, and every source URL the model returns.
  • Sampling: each question × engine pair is run repeatedly and averaged; run-to-run variance on mention rate measured at roughly ±10pp.
  • Comparison rule: each week is compared against a baseline built from the previous six runs; moves inside the noise band are reported as noise, not news.
  • Sample sizes quoted here: the per-engine table is one brand, 5 questions, 6 engines, 2026-08-24; the "16 earlier mentions then zero" is that brand's full run history on Qwen. The 30-answer / 6-engine figures come from our own weekly run in October 2026.
  • What we do not measure: answer quality beyond mention/accuracy/position, private assistant traffic, and any engine's internal ranking logic. Sources that can't be measured are reported as unmeasured rather than guessed.

BrandScope tracks how AI assistants describe your brand, every week. You can [run one brand through the live question set](https://www.brandscope.dev/) in about 30 seconds.

Want your own baseline? Run one brand through the live question set — about 30 seconds, no account needed. Check your brand →

← All posts