Why an ordinary rank check does not work here
An AI answer has no positions. It names zero, one, or a handful of businesses, in an order that can vary each time, and the answer changes if the question is phrased differently. A number from 1 to 10 does not exist. What does exist is: is your client mentioned, for which questions, and where does that mention come from.
The protocol we use
Per client, we fix a list of ten to fifteen local questions, worded the way people actually ask them. Not "best dentist", but "I'm looking for a dentist in Zwolle who is taking new patients". We ask those same questions again every month, to ChatGPT, to Google AI Overviews and to Gemini.
For each question we record three things: which businesses are mentioned, in what order, and which sources the answer cites. That last part is the most useful, because it shows where the assistant's picture comes from — often a comparison site, an industry directory or a local news site, and by no means always the business's own website.
What we found in practice
Three things stood out when measuring on our own locations and test profiles.
Answers are not stable. The same question, two days later, regularly produces a different set of businesses. One measurement therefore says little; a series taken over months says something. Anyone who sells you an AI visibility score based on a single measurement moment is selling you noise.
The sources are often third parties. For local questions, assistants strikingly often draw their picture from comparison and aggregator sites rather than from the business's own site. That explains why businesses with a mediocre website still get mentioned — and why consistent listings elsewhere carry more weight than most people assume.
How the question is phrased matters a great deal. Add one condition — "that's open in the evening", "that you can see without a referral" — and the set of businesses named changes completely. That is useful: it shows what information the business is publicly missing.
What you can do about it
There is no single switch. What there is, is a set of conditions that increase the chance an assistant recognises the business and is willing to name it: consistent details everywhere, schema that describes the business as a single entity, sameAs references to the profiles and registers it already appears in, and content that answers the question in the wording it is actually asked in.
What we do not do is promise positions. See AI visibility for how we set up the measurement protocol, and citations and NAP consistency for the groundwork underneath it.
Bronnen
- Our own monthly measurements on our own sites and test locations, with a fixed set of questions per location.
- Google, documentation on AI Overviews and on how business information is used across Google products.