Measuring your visibility in AI: an honest protocol

The honest answer to “do we show up in ChatGPT?” starts with another question: on which questions, in which assistant, and how many times have you tried? Measuring your visibility in AI is not done with a screenshot taken on a Tuesday. An assistant’s answer changes from one session to the next and from one week to the next.

This guide offers a protocol that an SME can run on its own, every month, and says what each reading does not measure. It also helps you judge a provider: serious tracking is part of what an AI SEO agency should deliver, method included.

Why a single measure is not enough

No data source sees everything, so you measure with three readings that complement each other:

1. What the assistants answer: a fixed list of questions, asked on a fixed date, several times.
2. What their crawlers read: the server or CDN logs, filtered on the crawlers’ names.
3. What arrives on the site: the visits whose source is an assistant.

The first tells you whether you are cited, the second whether you can be, the third whether it produces anything. Our guide to getting cited by ChatGPT sketches these three readings; this protocol makes them something you can keep up over time.

Reading 1: a list of questions, asked several times

Choosing the questions

Start from the questions a customer asks before contacting you, not from your sector’s keywords. “Which open-source CRM for a 30-person SME?” works better than “open-source CRM”. Three sources give good questions:

• requests received by form and by phone, in the customer’s own words;
• Search Console queries that take the form of a question;
• objections that keep coming up in sales meetings.

Keep between fifteen and forty questions: beyond that, nobody maintains the list every month. Classify each question by intent (find out, compare, buy): a brand can be absent from general questions and very present on comparisons.

Asking the questions properly

Assistants personalise their answers according to the account’s history and memory. For a measurement that can be compared from one month to the next:

1. 01Prepare.

   Use a session with no history and no memory switched on, always the same configuration.

2. 02Repeat.

   Ask each question at least three times, in three fresh conversations.

3. 03Note.

   For each try, record whether the brand is cited, which page is linked, which competitors appear, and whether the answer is correct.

That last point matters as much as the citation: an assistant that cites you with an out-of-date price flags a source page to correct, or to date more clearly.

Reading the result

You do not read “cited or not”, you read a frequency: cited in two tries out of three, on eight questions out of twenty. That frequency is what you track from one month to the next. A variation of one try means nothing; a trend over three months does.

Reading 2: the crawler logs

An assistant can only cite a page that its crawler was able to read, or that its engine has indexed. The server logs, or the CDN’s, tell you which crawlers visit, on which pages, and with which response (200, 403, 404).

Filter on the names each company publishes: Anthropic in its help page on crawlers, OpenAI in its list of crawlers. Separate training from search:

At Google, there is nothing to filter beyond Googlebot. The Google-Extended token, which controls the use of content for Gemini, has no user agent of its own according to Google: it does not show up in the logs.

Three things to look at every month:

• Refusals and absences. A run of 403s on a search crawler points to a firewall or a CDN rule that blocks without anyone having decided it. A crawler that reads only robots.txt, and nothing else, often points to a ban in that file: check the version served at the public address, not the one in the repository.
• The pages read. If the crawlers read only the home page, review the internal linking or the sitemap.
• Visits triggered by a user. A visit from ChatGPT-User or Claude-User to a page means a human asked a question that led there. It is the signal closest to a citation that you can see on the server side.

This reading tells you whether you are readable. It does not tell you whether you are cited.

Reading 3: visits coming from assistants

In your analytics, isolate the visits whose source is an assistant (chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com). Create a dedicated segment or channel group, so as not to drown them in “referrals”.

Two limits to keep in mind:

• It is a lower bound. Many citations produce no click: the answer is enough for the reader. And some visits arrive with no source passed on, filed under direct traffic.
• Google does not separate its AI answers. According to Google’s documentation on AI features, appearances in AI Overviews and AI Mode are counted in the Search Console Performance report, under the “Web” search type, with the rest of the traffic. The article on AI Overviews and traffic details what that implies.

Finally, add to your forms the question “How did you hear about us?”, with “an AI assistant” among the choices: it is self-reported, but it is often the only place where you can see that a prospect first asked ChatGPT.

A monthly table, not a dashboard

Every month, enter four figures in the same table, keeping in mind what each one says:

Count the volume before fixing the list: twenty questions asked three times to four assistants make 240 conversations a month. That figure decides between taking the readings by hand and using a tool.

What this protocol does not do

It does not give a precise “share of voice”, nor revenue attributed to citations: assistants’ answers are not a stable ranking.

Nor does this protocol replace the groundwork: measuring is there to know where to act. The llms.txt file, often presented as a lever, is more a matter of hygiene: see what you can really expect from an llms.txt file.

Where to start

Start small: twenty questions, three tries per question, and one month of logs. This first reading is the starting point; the following ones tell you whether an action had an effect.

If you would rather have this tracking run without you, it is part of our SEO and GEO content engine, which includes, every month, tracking of your positions in Google and of your citations by AI assistants.

Frequently asked questions

Sources

1. Its help page on crawlers (new tab)support.claude.com
2. Its list of crawlers (new tab)developers.openai.com
3. Google’s documentation on AI features (new tab)developers.google.com