Buyers ask assistants for recommendations, and whether you are named in the answer is a genuine commercial fact. It is also unusually easy to measure badly, because the output is prose, the systems are non-deterministic, and nobody can check your working unless you kept it.
Everything below is method rather than product. Run it by hand on a spreadsheet if you like; the discipline is what makes the number mean something.
The method
- Ask what a buyer would type'Best project management tool for agencies', not 'what do you think of Acme'. A question naming you measures whether the engine knows them, not whether it recommends them.
- Ask the same questions every weekA changing question set means every week-over-week movement is partly a change in what you asked. Add questions deliberately and note the week they entered.
- Force the answer to be groundedRequire a live web search. An ungrounded answer reflects what the model memorised during training: not what a buyer sees, not influenced by anything the client ships, not traceable to a source.
- Store the answer verbatim, and keep itEvery percentage should be traceable back to a sentence a person can read. Storing only your parsed summary means the first time somebody disputes a number, you have nothing.
- Separate a failure from an empty answerA rate-limited call produced no measurement and must be excluded from the denominator. An engine that answered and named nobody produced a finding and must be counted. Collapsing them makes every throttled week look like a collapse.
- Report per engineThe engines disagree with each other more than they agree. A single blended score hides the disagreement, which is usually the part worth acting on.
The failure mode to watch for in any vendor's dashboard: an engine that was unreachable rendering as 0%. That is their outage displayed as nobody naming you, and it is indistinguishable from a real collapse unless the tool tells you which it was.
What not to claim yet
Do not issue verdicts on AI-visibility changes the way you would on a search fix. Thresholds derived from Search Console's daily aggregates do not transfer to systems that answer the same question differently on a Tuesday.
Measure the week-to-week variance of your own question set for a couple of months before you conclude that a movement means something. Most published AI visibility movement is inside the noise band nobody has measured.
How many questions are enough?
Enough that one weird answer cannot move the percentage much. Below about ten, a single engine having an odd day swings the number more than anything you did.
Is being cited the same as being named?
No, and they should be counted separately. An answer can cite your page as a source while recommending somebody else, which is a different situation and often a more fixable one.
Try it on your own site, free
No card, and the free tier does not expire — the first verdict takes about six weeks to arrive.