Brand Mentions in ChatGPT: A Diagnostic, Not a Mystery

What we measured

Brand mentions in ChatGPT, Claude, and Gemini answers can fail at any of six independent links. I re-checked crawl access on trustgrowth.ai on 2026-08-19 UTC (the live robots.txt is also Last-Modified: Sat, 15 Aug 2026). That check covers Link 1 only. I did not run a sampled prompt test for this revision, so this article does not report mention counts.

This is a reader-run framework. Every check below is meant to be repeated against your own domain, with your own dates logged.

Method

• Data sources: robots.txt fetch and a crawler-user-agent HTTP response on trustgrowth.ai; Google Search Console Coverage and URL Inspection report definitions; vendor crawler documentation cited per link.
• Scope: one domain (trustgrowth.ai) for the crawl-access check. No anonymized third-party domain is used as a results sample in this version.
• Date range: crawl-access re-check performed 2026-08-19 UTC. The served robots.txt carries Last-Modified: Sat, 15 Aug 2026.
• Excluded: branded or category prompt sampling. A sampled-prompt check is described as a procedure you can run; it is not a result I am publishing here.
• Tooling: curl for HTTP checks. Chat interfaces are named only as tools you may use when you log your own sample.

What "brand mentions in ChatGPT" actually means

A base-model answer and a browsing-augmented answer come from different mechanisms. Non-browsing responses draw on patterns learned during pretraining up to a documented cutoff.

Browsing or grounding — ChatGPT with browsing enabled, Gemini with grounding, Claude with tool use — instead retrieves live web pages at query time. A brand absent from pretraining can still appear in a browsing answer if it is crawlable and ranks well now; a brand present in pretraining can be absent from a browsing answer when retrieval surfaces competitors instead.

No public API exposes exact training weights, retrieval ranking, or citation-selection criteria for these systems. Anything claiming to "know" why a specific brand was or wasn't mentioned is inference from sampled outputs, not verified ground truth from OpenAI, Anthropic, or Google. A matched symptom in this diagnostic is a correlation with a likely cause, not proof of causation. That is also why generative-engine visibility deserves separate measurement from classic keyword rankings, and why GEO belongs in a growth score at all: a logged sample, with model version and date, is how you move from anecdote to evidence. I am not attaching such a sample to this article.

The six links, in the order a request travels

Each link maps to one testable question and one primary evidence source.

Link What it means Primary evidence source Typical symptom if it breaks Crawl access Can a bot fetch the page at all robots.txt, server response to a crawler user-agent Zero visibility regardless of content quality Indexation Is the URL a confirmed entry in a search index Google Search Console Coverage and URL Inspection Invisible to retrieval-augmented answers Citation density Do other sources reference the brand Backlink/mention export, Wikidata entry Little for a model to synthesize from Content clarity Can a model extract and disambiguate what you do Organization/Author schema, one plain "what we do" sentence Attribution goes to the wrong entity Retrieval freshness Is the model's knowledge current for your brand Model version and documented training cutoff, browsing/grounding state Real but outdated presence Sampling Did you actually test it Logged prompt set with model version and date False negative, not a measured gap

Units: categorical. Source: mechanism mapping from the primary documentation cited per link below.

Link 1: crawl access

If a crawler cannot fetch your page, nothing downstream matters. Check robots.txt for disallow rules against GPTBot (OpenAI documentation), ClaudeBot (Anthropic documentation), Google-Extended (Google Search Central), and CCBot (Common Crawl). Treat any documentation reference as verified only as of the date you check it.

I am not restating the full crawler-verification method here. For user-agent, robots.txt groups, and log-level IP checks, use Check if AI crawlers can access your site. A tool that reports "blocked" can also be failing to reach you; how to tell whether an SEO tool measured your site or just failed to reach it covers that case.

Worked example, first-party, trustgrowth.ai, 2026-08-19 UTC:

curl -sS -D - https://trustgrowth.ai/robots.txt
curl -A "GPTBot" -I https://trustgrowth.ai/

The robots.txt response was HTTP 200. Body:

User-agent: *
Allow: /
Disallow: /dashboard
Disallow: /onboarding
Disallow: /monitoring
Disallow: /auth
Disallow: /rails/
Disallow: /scan

No rule names GPTBot. The homepage request with User-Agent: GPTBot returned HTTP 200. The served robots.txt header was Last-Modified: Sat, 15 Aug 2026. That confirms Link 1 is open for GPTBot on this domain as of 2026-08-19; it says nothing about the downstream links.

Link 2: indexation

Being crawled is not being indexed. Google Search Console's Coverage report distinguishes the two: pages can sit in Crawled – currently not indexed or Discovered – currently not indexed, meaning Google has seen the URL but declined to add it to the index used for search results and for many retrieval-augmented answers. Run URL Inspection on your top entity pages (homepage, about, product). For the full methodology see the Search Console audit for trust and visibility.

Link 3: citation density

Models and retrieval layers synthesize from what other sources already say about you, so backlinks, press mentions, directory listings, and Wikidata/Wikipedia entries are raw material. Inside TrustGrowth we treat "under roughly 10 referring domains and no structured entity record" as an internal working threshold for "almost nothing to draw on" — it is not a published OpenAI, Anthropic, Google, or Ahrefs cutoff. See how domain authority actually works for the underlying mechanics, and authority and backlink analysis for one way to pull that export.

Link 4: content clarity

A model can extract what is unambiguous. A generic brand name, missing Organization or Author schema, and the absence of one clear "what we do" sentence all suppress correct attribution. This ties back to the E-E-A-T signals Google actually checks: if your name collides with a common word or a more prominent entity, a model is likelier to answer about the other entity.

Link 5: retrieval freshness

Every model version has a documented training cutoff, and cutoffs differ across and within model families. A brand that launched or changed after a model's cutoff will not appear in that model's non-browsing answers no matter how strong its current signals are — a real gap, not a defect to fix. Log the model version and test date alongside any result. Do not name a model version you did not actually query.

Link 6: sampling

Extrapolating from a single anecdote is not measurement. A usable sample is a logged prompt set: the same 5–10 category-level questions (brand name omitted), run across the chat interfaces you care about, with model version, date, and browsing/grounding state written down.

I am not publishing mention counts from such a run in this article. If you run one, keep the transcript. A table with invented dates or withheld "anonymized" counts is not evidence.

How to test each link yourself

Each check has an explicit pass condition. Treat a link as failing until its condition is met with dated evidence.

• Crawl access — fetch robots.txt and search for GPTBot, ClaudeBot, Google-Extended, CCBot disallow rules; render the page with JavaScript disabled. Pass: no matching Disallow and primary content present without JavaScript.
• Indexation — pull the Coverage report and run URL Inspection on key pages. Pass: the page reports as indexed.
• Citation density — export a backlink/mention list (Ahrefs, Semrush, or equivalent) and check for a Wikidata entry. Pass: enough distinct referring domains that a human editor would call the entity documented, plus at least one structured entity record. Do not treat "~10 referring domains" as an industry law; that number is our internal working threshold.
• Content clarity — validate structured data and confirm Organization/Author markup resolve, and that one plain sentence states what you do. Pass: schema validates and the sentence is unambiguous.
• Sampled AI visibility — run the same 5–10 prompts across ChatGPT, Claude, and Gemini, logging model version, date, and browsing/grounding state. Pass: a logged prompt set exists. A missing log is not a mention rate of zero; it is an unrun test.

A GSC-verified site audit can automate steps 1–4. Why visibility gets its own score line explains treating it as a distinct signal. You can also run a free scan to see this evidence format applied to your own domain.

Where the chain usually breaks

Each break point produces a distinct symptom rather than a generic "low visibility" result.

Break point Typical symptom What it looks like on inspection Blocked crawler Zero visibility regardless of content quality robots.txt disallow rule matches GPTBot or Google-Extended Thin or unindexed pages Invisible to retrieval-augmented answers Coverage shows Discovered – currently not indexed Low citation density Little for a model to synthesize from Backlink export is sparse; no Wikidata entry Ambiguous naming Attribution goes to the wrong entity Brand name collides with a dictionary word or larger known entity Stale training data Real but outdated presence Model describes an old product name, pricing, or feature set Never tested False negative, not a measured problem No logged prompt history across any model

Units: categorical. Source: pattern synthesis from robots.txt inspection and GSC Coverage report definitions. No sampled-prompt counts are attached.

From diagnosis to fix

Each link maps to a distinct fix category. A blocked crawler is an engineering fix: edit robots.txt, resolve server errors, or server-side render the affected templates.

Ambiguous naming and missing schema are content-strategy fixes: write an explicit "what we do" sentence a model can lift verbatim, and add Organization and Author JSON-LD. Low citation density is an authority-building fix that takes months, not a config change. Turning a diagnosis into a scoped task rather than a vague "do more content" directive is what Groundcrew agent roles are built to automate, and an agent found, fixed, and verified a real SEO defect shows one instance of that detect-to-fix loop.

Limitations

• Link 1 only was run here. The first-party evidence in this article is the 2026-08-19 trustgrowth.ai crawl-access check. Indexation, citation density, clarity, freshness, and sampling are methods, not results from a TrustGrowth-run sample.
• No mention counts. I deleted the previously drafted 6-prompt table. Those counts are not in this version because they were not a test I can stand behind.
• No training or citation logic. This method identifies broken links; it does not guarantee future mentions. Fixing a blocked crawler restores the possibility of visibility; it does not force a model to cite you.
• No ranking or revenue claim. Neither was measured here. A finding that cannot show its robots.txt snapshot, its Coverage report, or its prompt log with model version is an assertion — the same standard behind public proof pages with dated evidence.

FAQ

How do I make my brand appear on ChatGPT?

There is no submission form or listing process for a base model. Work the six-link check in order: confirm GPTBot isn't blocked in robots.txt, confirm key pages are indexed, build genuine third-party citations, add explicit schema and a clear "what we do" sentence, then sample prompts across model versions to measure your current baseline before changing anything.

How do you get your business listed on ChatGPT?

There is no listing or directory that ChatGPT reads from directly. Visibility depends on whether your pages are crawlable, indexed, cited by other sources, and unambiguous enough for a model to extract and repeat — the same six-link check above.

How to make ChatGPT suggest your business?

Suggestion in a non-browsing answer depends on your brand appearing with enough frequency and clarity in the pretraining corpus; suggestion in a browsing-enabled answer depends on ranking well for the query at retrieval time. Test both separately, since a fix for one does not guarantee the other.

How to get ChatGPT to mention you?

Mentions in non-browsing mode reflect pretraining-era signal strength (citations, clarity, disambiguation); mentions in browsing mode reflect current crawlability and ranking. If you run a sampled-prompt test, log model version and browsing state before assuming either is the cause.

How to get your brand cited by ChatGPT?

Citations in browsing-augmented answers come from pages the model's retrieval layer can fetch and rank at query time, so the fix overlaps with crawlability, indexation, and topical authority, plus the disambiguation work in Link 4.

Why is my business not showing up on ChatGPT?

Usually one of six things: a blocked crawler, unindexed pages, near-zero third-party citations, an ambiguous or generic brand name, a training cutoff that predates your current offering, or simply never having tested it with a logged, repeatable prompt set.

How to get ChatGPT to promote your business?

ChatGPT has no promotional-placement mechanism; see the OpenAI usage policies, which describe no paid-mention or promotional-listing feature. Any answer that names a brand does so based on training-data patterns or live retrieval, not a commercial relationship.

How do you get your product on ChatGPT?

"On ChatGPT" usually means appearing in an answer, not being integrated as a plugin or connector. For appearing in answers, the same six-link diagnostic applies.

Key takeaways

• Brand mentions in ChatGPT, Claude, and Gemini answers have six distinct possible points of failure, not one universal cause.
• Base-model recall and browsing/retrieval answers are different mechanisms with different fixes; test both separately with model version and date logged.
• No public API exposes exact training or citation logic; a matched symptom is a correlation with a likely cause, not proof of causation.
• Check robots.txt for GPTBot, ClaudeBot, Google-Extended, and CCBot disallow rules first — a blocked crawler nullifies every downstream signal.
• I confirmed Link 1 open for GPTBot on trustgrowth.ai on 2026-08-19. I am not publishing sampled mention counts in this version.
• A finding maps to one of three fix types — engineering, content strategy, or authority building — and conflating them wastes effort on the wrong team.