How to Check if ChatGPT Mentions Your Brand

Start with six questions, one per buying stage. You want the shapes buyers type, in the buyer's words, and none of them should name your brand.

Then run them.

1. Open a fresh chat for each question. Memory and personalization bend ChatGPT's answers toward sites you already visit, which is exactly the skew you are trying to avoid.

2. Ask the question and read the whole answer, follow-ups included.

3. Log the outcome: named, cited with your domain, or absent. Also log which domains the engine cited instead, because that list is your to-do.

4. Repeat each question three to five times per engine and track the rate ("named in 3 of 5"). A single answer proves nothing; engines rephrase and re-retrieve on every run.

5. On Google, search the same six questions. An AI Overview does not appear for every query, so note which ones got one and who was named in it.

6. Re-run the set monthly and compare the rates month over month.

Perplexity shows its sources with every answer, which makes it the fastest engine to audit. ChatGPT and Gemini cite selectively, so read the text for your name even when no links appear.

Why an engine can name you and never cite you

Answers draw on two pools: what the model learned during training, and what a retrieval step fetched live while writing the answer. Your name can live in the first pool through reviews, directories, and forum threads while your own site never appears as a source.

Citations come from the second pool. Before an engine can cite your pages, its bots need permission to read them, and that permission lives in robots.txt. This is where the GPTBot question gets tangled, so it deserves its own section.

Should you block GPTBot?

GPTBot is OpenAI's training crawler. It reads your pages so future models can learn from them, and it sends almost nothing back: Cloudflare measured a crawl-to-referral ratio of 70,900 to 1 at peak in 2025. Blocking GPTBot stops that reading. It changes nothing about whether ChatGPT names you today, because today's answers come from a different bot.

ChatGPT's search answers are fed by OAI-SearchBot, and when a user asks ChatGPT to open a page, ChatGPT-User does the fetch. Anthropic splits the same way: ClaudeBot trains, Claude-SearchBot feeds Claude's search, Claude-User fetches on request. So the decision that moves your visibility is the one about the search bots. Current GEO practice blocks training crawlers and leaves the answer-surface bots open.

A baseline robots.txt for that policy looks like this. Treat it as a starting point, since paywalled content or staging hosts change the calculus.

# Training crawlers: stay out
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: CCBot
Disallow: /

# Answer-surface bots: welcome
User-agent: OAI-SearchBot
Disallow:

User-agent: PerplexityBot
Disallow:

An empty Disallow means allow everything. Group order in the file does not matter; a bot uses the most specific group that names it and ignores the rest.

Three special cases deserve attention. Google-Extended is a directive token, never a real crawler. A User-agent: Google-Extended block with Disallow: / opts you out of Gemini training, and AI Overviews ignore it because they follow Googlebot. Blocking Googlebot itself removes you from all of Google Search, so a site chasing visibility has no reason to do it. Bytespider ignores robots.txt entirely, so a rule against it is theater; a WAF rule or IP block is the fix that holds.

One more cleanup while you are in there. Older guides tell you to block anthropic-ai or Claude-Web. Both tokens are retired, and rules naming them do nothing today.

Read your robots.txt the way a bot reads it

Fetch https://yoursite.com/robots.txt and check three things.

First, which groups name an AI bot explicitly. A bot that finds its own group ignores the * group entirely, so a dedicated User-agent: OAI-SearchBot section overrides whatever the general rules say. Within a group, the longest matching path wins, which is how a specific Allow: /blog/ can beat a broader Disallow: /.

Second, what the * group says. Every bot you did not name falls back to it, so a blanket Disallow: / there locks out every compliant AI bot along with everything else you forgot to exempt.

Third, whether your rules match the token the bot sends. Applebot follows the Googlebot rules when no group names it, so a site that restrained Googlebot also quietly restrained Apple's answers without ever writing the word Applebot.

An llms.txt file at the root is a different, optional thing: a short curated index of your site, proposed by llmstxt.org, aimed at AI readers. Google has said plainly that no special files are required for its AI features, and the other engines manage without one. If you add it, keep it short and current, because a stale map is worse than no map.

Run the same check on a schedule (Selise GEO)

The manual method costs one evening per month. Selise GEO, our GEO product, automates the same loop end to end.

The free scan on selisegeo.com takes one URL and no account. It drafts the buyer questions, picks a searcher perspective, and asks the four engines shown on its landing page (ChatGPT, Gemini, Perplexity, Claude). The trimmed report covers overall score, per-engine citations, and a technical read of your site including per-bot robots.txt access and a check for llms.txt.

The paid product turns the one-off scan into brand tracking, which is where the manual method runs out of road. The same question baseline gets re-run against the same engines, so scores compare run over run. A gaps view lists every question no engine cited you for and who was named instead. A sources view shows the websites the answers drew on, split by site type, plus the searches the engines ran to build them. The AI crawler view checks each bot's access across your homepage and key pages, tests any single page live per bot, and compares your robots.txt with your competitors'.

What to fix first

1. Unblock the answer-surface bots. If OAI-SearchBot, Claude-SearchBot, or PerplexityBot meet a Disallow: /, the engines cannot read your site while composing answers, and content work cannot fix that while the block stands. The edit is one line.

2. Serve real HTML. Crawlers that do not execute JavaScript see nothing on a page whose text, headings, and structured data arrive only after the bundle loads. JSON-LD injected by JS is a known blind spot of exactly this kind.

3. Keep the basics honest: one h1, descriptive titles and meta descriptions, valid JSON-LD, a sitemap. Boring, weighted heavily by every audit that scores this stuff.

4. Publish where the engines already read. When the answers cite review sites and directories in your category month after month, and never you, the site-type mix is telling you where your missing presence belongs.

5. Add llms.txt last, or skip it. Nothing on this list does less for your visibility.

FAQ

Does blocking GPTBot remove my site from ChatGPT?

No. GPTBot trains future models. ChatGPT's live answers are fed by OAI-SearchBot, and page fetches on request use ChatGPT-User. Block those two and ChatGPT search stops reading your site. Block only GPTBot and today's answers stay as they were.

How many runs make a result trustworthy?

Three to five per question per engine, tracked as a rate with the sample size ("named in 3 of 5"). Engines vary their answers between runs, and a month-over-month rate is the signal, a single answer is noise.

What is the difference between a mention and a citation?

A mention is your name in the answer text. A citation is a link to your site as a source. Engines lean on third-party sites for sources, so a brand can be mentioned often and cited never.

Do I need llms.txt?

No. It is an optional proposal from llmstxt.org, and Google states that no special files are required for its AI features. Add one if you want a curated index for AI readers, and keep it accurate if you do.

Run the check tonight: six questions, four surfaces, three runs each. Log the results, fix any robots.txt lines that block the answer-surface bots, and ask the same questions again in a month.