Free GEO Audit: Find Out If ChatGPT Cites Your Brand

This article contains affiliate links: we may earn a commission. Learn more

To find out whether ChatGPT, Perplexity or Google cite your brand, ask ten questions three times on each engine: 90 answers to score in a spreadsheet, about two hours of work. Here is the five-step protocol, the margin of error of each rate and the technical checks to redo on your site.

We did not apply this protocol to our own site, and we did not use the free tools listed below. What we did check is the official documentation of the bot publishers (as listed on October 18, 2026) and the technical controls on our site. How we test explains why some articles stay informational.

For a definition of GEO and how it differs from SEO, see GEO: definition and differences from SEO.

What does it mean to be “cited” by ChatGPT, Perplexity or Google?

An AI names you when it writes your brand name, and cites you when it shows a link to your site as a source. The two are independent: a brand can be named without a link, or a site can be cited without the name appearing in the text. Score them separately, in two columns.

Can you check for free whether an AI cites your brand?

Yes, by hand and with no paid tool: ten questions asked three times on each engine give 90 answers to score in a spreadsheet. Allow one to two minutes per answer, so about two hours (an estimate, we did not time it). That is enough for a starting point, not for tracking over time, which means repeating the same measurement every month.

Heavier protocols exist. Edikka publishes an open protocol under a CC BY 4.0 license: 25 questions, 4 platforms, 3 repeats, so 300 tries. Ours is shorter, so you can do it in an afternoon.

Which questions should you ask the AIs?

Take ten questions split into three families: two about your brand, five about your category or competitors, three about the problems your customers are trying to solve. Write them once, in your customers’ language, and don’t change a word during the whole measurement: changing a question changes the measurement.

How should you ask the questions so the measurement is reliable?

Open a fresh conversation for every try, in a private window or without an account, with web search on when the engine offers it. Ask each question three times per engine, from the same country and in the same language, and note the date, the engine, the mode and the country. Answers vary with the engine, the session, the location and the date.

Three engines are enough to start: ChatGPT (web search on), Perplexity and Google (AI Overviews or AI Mode). Ten questions, three tries, three engines: 90 answers. Gemini can be added as an option.

How should you score the answers?

For each answer, fill in two boxes: is the brand named (0 or 1), is the site cited with a link (0 or 1). Add the competitors named and the domains cited. The cited domains are the most useful result: they show which pages feed the answers to your questions, and so where you would need to appear.

Then show the result as k/n per engine and per question family: “7 of 30 answers name the brand” beats “23%”, because n shows the sample size.

How many tries do you need for the rate to be reliable?

The more tries, the tighter the rate. In the worst case, 30 tries give a margin of about 18 points, 90 tries about 10 points, and 10 tries about 31 points. Published methods range from 3 repeats per question (Edikka, Wam) to about thirty for ChatGPT (Meteoria), without justifying the number.

These margins are a calculation, not a measurement: normal approximation, worst case (a 50% rate), tries assumed independent. Repeats of the same question are not necessarily independent. Use them as an order of magnitude: a gap between two measurements smaller than the margin proves no change.

Can the AIs read your site?

Check robots.txt first. According to its documentation, OpenAI recommends allowing OAI-SearchBot to appear in ChatGPT search, and Perplexity recommends allowing PerplexityBot. For Google, a page must be indexed and eligible for a snippet in search, with no particular schema.org markup; its July 10, 2026 guide adds that a site must be included in generative AI features in Search Console (a setting we have not checked). Each bot has a different role:

Sources: documentation from OpenAI, Perplexity, Anthropic and Google, as listed on October 18, 2026. OpenAI says a robots.txt change can take about 24 hours to be taken into account. Anthropic says its bots respect robots.txt directives.

Three commands redo the check on any site. The first reads robots.txt, the second requests a page with a bot’s user-agent (the expected status is 200), the third checks for an llms.txt:

curl -s https://exemple.fr/robots.txt
curl -s -o /dev/null -w "%{http_code}\n" -A "OAI-SearchBot" https://exemple.fr/une-page/
curl -s -o /dev/null -w "%{http_code}\n" https://exemple.fr/llms.txt

A 200 shows that no rule blocks this user-agent, not what a firewall would do with the IP address of a real bot. For Google, Search Console also shows the share of traffic that came from AI Overviews and AI Mode: Google’s July 10, 2026 guide points to a “Generative AI performance” report, and its December 2025 page counted these appearances in the Performance report, under the Web search type. We have not opened this report in our own account.

Does an llms.txt file help you get cited?

Not for Google: its guide to optimizing for generative AI, updated on July 10, 2026, says Google Search does not use llms.txt files and ignores them, so publishing one neither hurts nor helps in Google. For the other engines, nothing in the documentation we read says so. The llms.txt proposal site describes it as a help file for agents, and both OpenAI and Perplexity publish an llms.txt for their own documentation, but none of the pages about their bots (OpenAI, Perplexity, Anthropic) says those bots use it to choose sources. We have not measured it: an llms.txt costs nothing to publish, but don’t count on it as a guarantee. The details, with a survey of 26 sites and two studies, are in llms.txt: what is this file for.

What we checked on our own site

On October 18, 2026, the robots.txt of hayka-pacha.io allows everyone except /api/ and /go/ (our link redirects), and has no rule specific to AI bots: OAI-SearchBot, GPTBot, ChatGPT-User, PerplexityBot, ClaudeBot and Claude-SearchBot are therefore allowed. Our llms.txt is present. An article page returns 200 with the user-agent of each of these bots and that of Google-Extended, just as with a browser.

Limits: these requests come from our own machine with a declared user-agent, not from the IP addresses of the real bots. And we did not ask the engines the protocol’s ten questions for our own brand: so we publish no citation rate.

Which free tools exist?

Three free tool pages stand out. We did not use them; what follows is what their pages say.

• Ahrefs (page in French) offers a “free AI visibility checker”. We know neither its usage limits nor its results.
• Geoptie (page in French) offers a free GEO audit with “no sign-up”. It assesses a site’s readiness on six dimensions, it does not measure real citations.
• Neil Patel (page in French) offers an AI search visibility page. We did not read its details.

What about automating this tracking?

Repeating this measurement by hand every month gets long. If you would rather automate it, we compared four paid tools, their prices and their free trials in our comparison of AI visibility tracking tools. Our pick to get started is Hikoo, which listed a 7-day trial with no credit card on October 12, 2026.

What we have not verified

• Real demand for “AI visibility” and “ChatGPT cites my brand”: no measurable search volume, and we only measured France.
• The results of our own brand in the AIs, as the protocol was not applied to our site.
• The free tools, of which only the public pages were read.
• The real effect of an open robots.txt or an llms.txt on citations.
• The real time needed to run the 90 answers, estimated at about two hours.
• How the AIs behave outside France and in languages other than French.