How to check whether AI mentions your brand
Ask them yourself, properly, and write down the sources as well as the names. It costs an afternoon and no money, and the two best signals — your own server logs and Bing’s free AI performance report — are first-party and available to nobody but you.
The method, in six steps
Do these in order. Steps one to three cost an afternoon; steps four and five are two sources you already own and probably have never opened.
- Write down ten questions a buyer asks before they know you. Not your brand name. The questions someone types when they are still choosing a category, a shortlist or a price band.
- Ask each question in a fresh, logged-out session. Personalisation and chat history contaminate the result. Use a private window, no account, and do not follow up inside the same thread.
- Record three things per answer, in a spreadsheet. Whether your brand was named, in what position among the brands named, and the full list of sources cited. The source list is the part most people skip and it is the most useful.
- Open your server logs and search for the AI crawler user agents. A hit from a user-triggered fetcher means a real person's assistant opened your page mid-conversation. This is first-party, free, and available to you and nobody else.
- Open Bing Webmaster Tools and read the AI performance report. Free, first-party, and the only honest source of Copilot data there is. It covers your own site only, which is a limitation and also a guarantee of accuracy.
- Repeat the whole thing next month with the same questions. A single run is not a measurement. Engines are non-deterministic, so only a series on a fixed question set tells you anything.
The six sources, cheapest first
1. Ask the engines by hand
Free, available today, and covers every engine including the ones with no API. It is also the method most people get wrong, in four specific ways:
- They search their brand name. That measures a question nobody asks. A buyer who knows your name does not need a recommendation. Ask the questions people type before they have heard of you.
- They stay logged in. Chat history and personalisation contaminate the answer. Use a private window with no account.
- They ask once. Engines are non-deterministic. Our own two runs a minute apart gave identical headline numbers but moved per-domain citation counts by one or two.
- They record only whether they were named. The source list underneath is more useful: it is the literal list of pages you are competing against, and getting onto one of them is often easier than outranking it.
2. Read your own server logs
Underrated, entirely free, and first-party. Every AI crawler identifies itself in the user agent. Search your access logs for these strings:
| Look for | What a hit tells you |
|---|---|
PerplexityBot, OAI-SearchBot, Claude-SearchBot | Your page is being indexed as a candidate source. It can be cited. |
ChatGPT-User, Perplexity-User, Claude-User | A real person’s assistant opened this page during a conversation. This is the closest thing to a click that AI search produces. |
GPTBot, ClaudeBot, CCBot | Training collection. Interesting, but it tells you nothing about today’s answers. |
Which pages get fetched, and how often, is a genuine signal about what the engines consider relevant to your category — and it is not available in any third-party tool, because it is your log. Crawler names change; the current list and what each one does is on AI crawlers and robots.txt.
3. Bing Webmaster Tools
Free, first-party, and the only non-speculative source of Copilot data that exists. Microsoft reports AI-surface performance for your own verified site. The limitation is real and worth stating: it covers your site only, so it cannot tell you anything about competitors. In exchange it is not an estimate.
4. Google Search Console, with a caveat
Search Console does not separate AI Overview impressions from ordinary search impressions in a way you can rely on. What you can watch is the pattern that usually accompanies an Overview appearing above your result: impressions holding or rising while clicks fall. That is a hint, not a measurement, and anyone presenting it as an AI Overviews report is over-claiming.
5. A free tier or a free check
Several trackers offer a free query allowance, and so do we — three questions on Perplexity, no payment. Used honestly these are a sample, not a monitoring programme, and their value is telling you whether it is worth setting one up at all. If the sample convinces you to buy something, tool, agency or in-house is the next decision.
6. Scripting it yourself
If you can write a little code, the Perplexity API will run your whole question list and return the citation array for each answer. That array is what our entire citation analysis is built from. A day of work gets you a repeatable measurement you own outright. Our method, including how we treat failed calls, is documented in full on how we measure AI visibility so it can be copied.
What none of these can see
- Your analytics cannot answer this question. A citation nobody clicked still shaped the answer and left no referral trace. Zero AI referral traffic is not evidence of zero AI visibility, and treating it as such is the most common reporting error in this field.
- Nobody can see the answers real users got. Every method above simulates a user. The distribution of what real people were told is not observable from outside the engine, by us or by anyone selling you a dashboard.
- Nobody can measure AI Overviews or Copilot through an API, because neither has one. Automated coverage of those is scraping, and it should be described as such.
What we can and cannot measure, stated plainly
Because the question on this page is one you should also be asking about us: Perplexity is the only engine we hold an API key for. Our ChatGPT and Claude adapters are built but not yet keyed. Gemini is not built. Google AI Overviews and Microsoft Copilot have no public API and are reviewed by hand for Scale clients. Any vendor in this category should be able to answer that question as directly.
What a good first result looks like
Probably bad, and that is normal. Ours was: named in 0 of 8 answers, 0 of 90 citations, a visibility score of 0 — published in full on our score page. The number is only useful as the start of a series, so take it, date it, keep the raw answers, and run it again next month against the same questions.
Common questions
How do I check if ChatGPT mentions my brand?
Ask it, in a logged-out private session, the questions a buyer types before they have heard of you rather than your brand name. Record whether you were named, your position among the brands named, and the full list of sources cited. Then search your server logs for the OAI-SearchBot and ChatGPT-User user agents, which tell you whether your pages are being indexed and opened. There is no ChatGPT API that reports citations, so anyone offering automated ChatGPT tracking is simulating a user the same way you would.
Is there a free way to track AI citations?
Yes, three of them. Ask the engines by hand and log the sources they cite. Search your own server access logs for AI crawler user agents, which is first-party and available to nobody else. And open the AI performance report in Bing Webmaster Tools, which is free and the only non-speculative source of Copilot data. None of these costs anything.
Can Google Analytics tell me if AI is citing my site?
No, and this is the most common reporting error in the field. A citation that nobody clicked still shaped the answer and left no referral trace at all. Zero AI referral traffic is not evidence of zero AI visibility. Server logs are far better, because a crawler hit is recorded whether or not a human ever clicked.
How often should I check?
Monthly, on a fixed set of questions. Weekly tells you mostly about engine noise: our own two runs a minute apart produced identical headline numbers while individual domain counts moved by one or two. Changing the questions between runs breaks the comparison entirely, so fix the list before you start.
Or have the first two steps done for you
Three of your buyers’ questions on Perplexity, free, with every source it cites. It is the same method as below, run for you.
3 questions on Perplexity, free. A domain checked in the last week is served from that stored run rather than asked again, and the daily free allowance resets at midnight UTC.