How to measure visibility inside AI assistants
Four readings are obtainable: whether assistants mention you, what they say, which of your pages their crawlers fetch, and how your search impressions move against your clicks. Volume inside an assistant is not obtainable, by anyone, and any dashboard showing it is modelling rather than measuring.
There is a lot of confident measurement being sold in this area, and most of it is a model dressed as a reading. Worth being precise about which is which, because a board that discovers a headline figure was estimated stops believing the honest ones too.
What can genuinely be measured
Ask a fixed set of buyer questions, in the languages of your markets, across the assistants your market uses, on a schedule. Record the answer text and any sources shown. This is sampling, so treat one run as a snapshot and the series as the reading.
Your own server logs, filtered by user agent. This tells you which pages are being read, how often, and whether the fetches succeed. It is the most reliable and most neglected source you already own.
In Search Console. Informational queries where impressions hold and clicks fall are the signature of an answer being given above your result.
The only reading that connects to revenue. Ask in the first conversation how they came across you and what they had already understood. It is qualitative and it is the one executives find persuasive.
What cannot be measured, by anyone
| Commonly offered | Reality |
|---|---|
| "Your share of AI answers" | Derived from a sample of prompts chosen by the vendor; it is a survey result, not a share |
| "Queries per month in ChatGPT" | No provider publishes query volume by topic. Any figure is modelled from search data |
| "Citations you received" | Only observable where citations are displayed, and displayed citations are an incomplete view of what was read |
| "AI traffic" | Referrer data from assistants is partial and inconsistent; much of it arrives as direct |
| "AI visibility score" | A composite the vendor defined. Useful for tracking direction if the method is published, meaningless as a level |
None of this makes the vendors dishonest, and a well-constructed sample tracked over time is genuinely useful. It becomes dishonest at the moment the output is presented as a measurement of the world rather than of the sample.
A number nobody can derive from a source is not a measurement. It is a position.
A monthly routine that holds up
- Twenty fixed questions, written as a buyer would ask them, in each market language. Do not change them, or the series is worthless.
- Run them across the assistants your market uses. Record the full answer text, not a score.
- Mark three things per answer: were you named, was what was said correct, were you shown as a source.
- Pull crawler fetches by user agent from your logs, by page.
- Pull impressions and clicks for question-shaped queries from Search Console.
- Write two sentences on what changed and why you think so. The sentences are the report; the tables are the appendix.
What to do with a bad reading
Absence usually has a boring cause. Before concluding that you need a visibility programme, check that your pages are fetchable without JavaScript, that the content is not behind a consent wall in the returned markup, that nothing in your robots file blocks a retrieval crawler, and that you have actually published a page that answers the question. In most cases at least one of those is the answer.
Questions this settles
No. No provider publishes topic-level query volume, so every figure of this kind is modelled from conventional search data and should be described that way.
Monthly is enough for direction and cheap enough to sustain. Weekly sampling mostly measures the variance of the models rather than any change in your position.
Sometimes, inconsistently, and often not at all. Treat referrer-based counts of assistant traffic as a floor rather than a figure.