You cannot measure your visibility in AI answers precisely, because no assistant reports how often it cites you. You can track three honest signals: visits referred from assistant sites, AI crawler activity in your server logs, and the results of a fixed panel of test questions you run each month. Together they show a trend, which is what you need to judge whether your work is paying off.
This post explains each signal, what it can and cannot tell you, and a simple routine to run.
Why measurement is hard
Classic search gives you impressions, clicks and positions through tools such as Google Search Console. AI answers do not have an equivalent for most assistants. Answers are generated per request, vary with the wording of the question, and often satisfy the reader without a click. There is no public log of who asked what and which sources were used.
So be wary of any report that claims an exact share of AI answers for your brand. It is a sample of prompts someone chose, taken at one moment. It can still be useful, as long as you read it as a sample.
| Can measure | Cannot measure reliably | |
|---|---|---|
| Traffic | Visits referred from assistant sites | Answers that mention you with no click |
| Crawling | Which AI crawlers fetch which pages | How a fetched page was used |
| Answers | What a fixed set of test prompts returns today | What every user sees, every time |
| Share | Your own trend over time | A precise market share of AI answers |
Signal one: referral visits from assistants
When a user clicks a cited link in an assistant, the visit often arrives with the assistant's domain as the referrer. In your analytics, create a segment for known assistant and AI search domains and watch it over time.
- What it tells you: that you are being cited and that the citation is useful enough to click. Which pages earn those clicks is especially valuable.
- What it misses: every answer that cites or paraphrases you without a click, and visits where the referrer is stripped, for example from some apps.
Treat the number as a floor. A rise usually means real gains; a flat line does not prove nothing is happening.
Signal two: AI crawler activity in your logs
Your server access logs record the user agent of every request. AI providers publish the names of their crawlers, so you can filter for them: for example GPTBot, OAI-SearchBot and ChatGPT-User from OpenAI, ClaudeBot and Claude-User from Anthropic, and PerplexityBot from Perplexity. Our guide to robots.txt for AI crawlers lists the main ones and what each is for.
- What it tells you: which pages AI systems fetch, how often, and whether a user-initiated fetcher visited a page. That last one is a direct sign that someone asked an assistant about content on your site.
- What it misses: how the page was used afterwards. A fetch is not a citation.
User agent strings can be faked, so for decisions that matter, check the provider's published IP ranges where they offer them. Also check your own rules: a crawler you block in robots.txt should stop appearing, which confirms your policy works.
Signal three: a fixed prompt panel
Write a list of 20 to 30 questions your customers really ask, the kind your pages are meant to answer. Once a month, ask each question in the assistants that matter to you and record three things: whether you are cited, whether you are mentioned, and which competitors appear.
- Keep the questions and wording fixed, so months are comparable.
- Use a clean session, and note the date, the assistant and your location.
- Record facts, not impressions: cited yes or no, position of the citation, the linked page.
This is manual, and that is the point. It is cheap, it is transparent, and it shows you exactly what a customer would see. It also surfaces content gaps: a question where a competitor is always cited is a page you should write or improve. See how to write content AI assistants quote.
Reading the signals together
Each signal is weak on its own. Together they tell a story. Suppose your prompt panel shows you are now cited for a question about delivery times. In the same month your logs show more user-initiated fetches of your delivery page, and analytics shows a small rise in assistant referrals landing on it. Three independent signals point the same way, so you can be fairly confident the change you made to that page worked.
The reverse is useful too. If AI search crawlers fetch a page often but you never see it cited in your panel, the page is being read but not chosen. That usually points to the content: the answer is buried, unclear or contradicted elsewhere. If the crawlers never fetch it at all, the problem is access or discovery, such as a robots.txt rule, a missing link, or text that only appears after JavaScript runs.
What to report, and what not to
Report trends and specific wins: more assistant referrals to a key page, a new user-initiated fetch pattern, a prompt where you are now cited. Do not report a single "AI visibility score" as if it were a precise fact. The systems vary too much for that, and a number that moves with the prompt wording will mislead the people who read it.
Fix the inputs while you measure
Measurement only helps if you act on it. The inputs you control are the same ones covered in what is answer engine optimization: crawl access, clear answers, structure, structured data and freshness. The technical side is quick to check.
Next step
Before you measure, make sure AI systems can reach and read your site. The free agent-readiness scan checks your crawler rules, llms.txt, Markdown and structured data in one pass. If you want the reporting routine built into your own tools, see our automation and workflows service.