How to check if ChatGPT mentions your brand: a repeatable test
A manual, repeatable test for whether ChatGPT and other assistants mention, cite or recommend your brand, with a prompt set, a log and honest limits.
By Crawlito · Updated 9 min read
In short
- A mention, a citation and a recommendation are three different results, so log them in separate columns instead of recording one yes or no.
- Use 10 to 20 prompts written the way your customers ask, run each in a clean session, and save the date, the mode and the exact output.
- Treat one run as one sample: repeat each prompt on a fixed schedule and judge the trend, not a single answer.
- Search Console's generative AI performance report counts impressions in AI Overviews and AI Mode, and Bing's AI Performance report counts citations; neither describes click data.
- A manual test shows what a handful of prompts returned on one day; it cannot show how often you appear across everything people ask.
On this page
- Mention, citation and recommendation are three different results
- Build a small prompt set from real customer questions
- Run each prompt in a clean session and record everything
- Score what you saw
- Repeat on a fixed schedule and keep notes on variance
- Check first-party data too
- What a manual test cannot tell you
- Frequently asked questions
- Check the technical side first
To check if ChatGPT mentions your brand, write 10 to 20 questions your customers would really ask, run each one in a fresh session, and record exactly what comes back: whether your brand is named, whether your site is linked, and whether the description is correct. Repeat the same prompts on a schedule. That is a free test you can run this afternoon, and it needs no tracking tool.
One honest limit first. A manual test is a small sample of one assistant on one day. It can show you a problem or a pattern. It cannot tell you how often you appear overall, and nothing here makes an assistant more likely to name you. It only makes your observations trustworthy.
Mention, citation and recommendation are three different results
Before you run anything, decide what you are counting. People say "ChatGPT mentions us" and mean three different things.
| Result | What you see | What it tells you |
|---|---|---|
| Mention | Your brand or domain name appears in the answer text | The assistant knows the name and chose to say it |
| Citation | Your page is linked or listed as a source | A specific page was used or shown as support |
| Recommendation | The answer suggests you as a choice for the user's task | The assistant placed you among the options, ideally with a reason |
They can happen separately. An answer can name your brand with no link, link your page without naming the brand in the text, or list five competitors and leave you out. If you log only "yes" or "no", you lose all of that. Our post on AI citations vs referral traffic goes deeper on why these numbers should not be merged into one figure.
Build a small prompt set from real customer questions
The prompts decide whether the test means anything. Write them from what customers actually say, not from your own product vocabulary.
Good sources:
- Sales call notes and support tickets, where people phrase the problem in their own words.
- The "People also ask" boxes and autocomplete suggestions you see in a normal search.
- Questions you answer repeatedly by email.
- Your own site search log, if you keep one.
Aim for 10 to 20 prompts and group them by intent:
- Problem prompts: "How do I stop my invoices going to spam?"
- Category prompts: "What are good invoicing tools for a five-person agency?"
- Comparison prompts: "Compare tool A and tool B for freelancers."
- Brand prompts: "What is Acme Invoicing and who is it for?"
Brand prompts tell you whether the assistant describes you correctly. Category and problem prompts tell you whether you are in the consideration set at all, which is the harder and more useful question. Keep the wording fixed once you choose it. If you reword a prompt later, treat it as a new prompt, or your trend line stops being comparable.
Write the prompts down in a spreadsheet before you start so you are not tempted to tweak them after seeing results.
Run each prompt in a clean session and record everything
An assistant can use context you have forgotten about: earlier turns in the same chat, saved preferences, or a location. To reduce that, start a new conversation for each prompt, and consider a signed-out or fresh browser profile for a second pass. Using both gives you a feel for how much the answer depends on who is asking.
For every run, record:
- The date and time.
- The assistant and the mode you used, including whether web search was on or whether the answer showed sources.
- The exact prompt.
- The full answer text, pasted in, or a screenshot.
- Any source links shown, copied as full URLs.
A simple log is enough:
date, assistant, mode, prompt_id, brand_named, page_linked, description_correct, competitors_named, notes
2026-11-11, ChatGPT, web sources shown, P03, no, no, n/a, "Tool A; Tool B; Tool C", answer cites two review sites
2026-11-11, ChatGPT, web sources shown, P07, yes, yes, partly, "Tool A", pricing described is out of dateSaving the raw output matters. A summary such as "we showed up" is not something you can compare against next month. The pasted text is.
If you want to check more than one assistant, repeat the same prompts in each. Each assistant has its own crawlers and sources: OpenAI documents separate crawlers for ChatGPT search, training and user requests, and Perplexity documents its own PerplexityBot and Perplexity-User. A result in one says little about another.
Score what you saw
Once a run is logged, score it with a short, fixed checklist instead of a gut feeling.
- Named: is your brand or domain in the answer text?
- Linked: is one of your URLs shown as a source? Is it the page you would want people to land on?
- Described correctly: compare each factual claim about you against your own site. Note what is wrong, outdated or missing.
- Compared: which competitors appear, and in what role: as the main answer, as an alternative, or as a cautionary example?
- Source type: what kinds of pages are cited instead of yours, such as review sites, forums, vendor pages or news? That tells you where the assistant is looking.
The fifth point is often the most useful. If answers to your category prompts keep citing the same two comparison articles, the practical question is whether those pages describe you accurately, not whether your homepage needs another paragraph. AI search competitor gap analysis shows how to turn those observations into a content comparison.
A wrong description is a finding, not a failure of your test. Fix the source on your own site if it is yours, and record the date so you can see whether later runs change.
Repeat on a fixed schedule and keep notes on variance
One run proves very little. Answers from a generative assistant can differ between runs of the same prompt, and the sources can change with them. So:
- Run each prompt at least two or three times in the same sitting, in separate sessions, and log every run.
- Repeat the full set on a fixed schedule, for example the first Monday of each month. Weekly is fine if you are making changes and want to see their effect.
- Keep a notes column for anything that changed on your side: a new page, a robots.txt edit, a site migration, a product launch.
- Report results as counts out of runs ("named in 2 of 6 runs") rather than as percentages from tiny samples.
Be careful with before-and-after claims. If your numbers move, other things moved too: the assistant's own systems, the pages competitors published, the wording of the question. A change that follows your fix is consistent with the fix helping, but it does not prove it. For the reporting side, an AI search reporting dashboard covers how to show sampled results with their limits visible.
Check first-party data too
What an assistant says is one signal. What reaches your site is another, and you own it.
Server or CDN logs. OpenAI documents that ChatGPT-User may visit a page when a user's request calls for it, and that OAI-SearchBot is the crawler for ChatGPT search. Perplexity documents Perplexity-User for user-initiated visits. Search your logs for those tokens around the time of your tests, and verify the IP against the operator's published ranges, because anyone can fake a user agent. A visit is encouraging but not proof that your page was used in an answer, and a missing visit does not prove it was not. Measuring AI crawler traffic walks through the log commands.
Referral traffic in analytics. OpenAI's publisher FAQ says ChatGPT adds utm_source=chatgpt.com to referral URLs from ChatGPT search, so you can segment those sessions in Google Analytics or another tool. The tag only identifies visits that carry it, so treat the number as a minimum rather than a full count.
Google Search Console. Google says traffic from AI Overviews and AI Mode is included in the overall Search Console data and appears in the Performance report under the Web search type. Google also documents a generative AI performance report, rolled out to all sites as of August 31, 2026, that counts how often links to your site were shown in AI Overviews and AI Mode. At the time of writing, its help page describes impressions only, not clicks, and it excludes Search Labs experiments. Our post on the Search Console generative AI performance report covers how to read it.
Bing Webmaster Tools. Bing's AI Performance report, introduced as a public preview in February 2026, reports how often your site is cited in Microsoft Copilot, AI summaries in Bing and select partner integrations, with cited pages and a sample of the grounding queries. Bing describes citation counts as a measure of how often a page is cited, not of its importance, ranking or placement, and the announcement does not describe click data. In June 2026 Bing began a preview of Intents, Topics, Citation Share and Compare. Citation Share is your share of citations for a grounding query, and Bing calls it an observational metric, not a ranking. See Bingbot and AI search for setup.
What a manual test cannot tell you
Knowing the limits keeps you from over-reading a good week or panicking over a bad one.
- It is a sample. 15 prompts cover a sliver of what people ask. Your real questions may differ.
- It reflects one moment. The same prompt next week can return different sources.
- It does not capture personalization. Your session is not your customer's session.
- It does not measure influence. A brand can shape a purchase decision without ever being linked, and a link can appear without anyone clicking it.
- It does not explain why. The test shows what happened, not the reason. For the reason, check whether your pages are reachable at all: robots.txt rules, firewall challenges, content that only renders with JavaScript. Why ChatGPT can't see your website covers those causes, and AI search regression testing shows how to guard them on each release without depending on live answers.
If you later buy a tracking tool, the same rules apply to it: ask how many prompts it runs, how often, in which mode, and whether it shows the raw answers. A dashboard number without the underlying outputs is hard to trust.
Frequently asked questions
Can I just ask ChatGPT "does it mention my brand?"
That tests one question, not your visibility. Ask the questions your customers ask, and log the answers. Asking directly about your brand mostly tells you what the assistant says when your name is already in the prompt.
How many prompts and runs do I need?
Start with 10 to 20 prompts and two or three runs of each. That is enough to spot obvious gaps, such as being absent from every category prompt, without pretending to be a statistical measurement.
Why does the answer change when I repeat the prompt?
Generative answers can vary between runs, and the sources an assistant shows can vary with them. That is why you log every run and report counts out of runs rather than a single yes or no.
Does a visit from ChatGPT-User in my logs mean I was cited?
No. OpenAI says ChatGPT-User may visit a page when a user's request calls for it. A visit shows a page was fetched, not that it was quoted or linked in the answer.
Will fixing my site make ChatGPT mention me?
Fixing access and rendering problems removes technical barriers, but whether an assistant names or cites you depends on many factors outside your site. Use the test to see what changes, not to expect a particular result.
Check the technical side first
If your test shows you missing from answers, rule out the technical causes before anything else. The free check reads one page the way AI crawlers do, with no account needed: the robots.txt rules for each named AI crawler, the robots meta tag and headers, and whether the main text is in the raw HTML. It tells you whether a page is readable, not whether an assistant will cite it.
Filed under Strategy and audits. Topics: ai-search, analytics, checklist, strategy