Mentioned isn’t recommended: how we measure AI visibility
Why we report mentions, recommendations and citations separately, and the rules that keep AI visibility numbers honest.
“AI visibility” sounds like one number. It isn’t. When a patient asks an AI engine where to get care, several different things can happen to your name, and they are worth very different amounts. This note sets out how we measure them, and the rules we follow so the numbers mean something.
Start from a fixed question set
For each specialty and market we write questions the way patients actually ask them: with a city, a worry or a budget. The set stays frozen for a measurement period. If we change the questions, we start a new baseline instead of comparing old and new numbers. A rise that comes from easier questions isn’t progress.
Keep two kinds of question apart
Some questions name you: “Is [your clinic] good for implants?” Others don’t: “Where should I get implants in Shenzhen?” A question that names you will almost always produce a mention, so mixing the two inflates the result. We report them separately. Named questions check whether AI describes you correctly. Unnamed questions check whether AI thinks of you at all.
Mentioned, recommended, cited
For each answer we record three things:
- Mentioned. Your name appears anywhere, in any context, including a warning or a long list.
- Recommended. The answer lists you as a place the patient could go. We also record your position and the tone. This is our closest measure of being on the shortlist. It is an approximation, judged answer by answer, and we say so.
- Cited. The answer gives a page on your own site as a source. This is only possible on engines that search the web and show their sources.
The three can move in different directions. You can be mentioned more often and recommended less, for example when AI starts naming you in comparisons that favour someone else.
Missing is not zero
Sometimes an engine times out, refuses or returns an empty answer. We count that as missing, not as an answer that left you out. Rates use valid answers only, and we show how many are missing so you can see the range the true value could fall in.
Show the sample, and the range
“23%” means little without “of how many”. Every rate in our reports comes with its numerator and denominator and a 95% interval. With 120 answers, a change of three or four points can easily be noise. We only call a week-on-week change a rise or a fall when the intervals don’t overlap.
Compare like with like
We only compare periods that used the same questions, the same engines and the same model versions. When an engine changes its model, we note it and break the trend line instead of drawing straight through.
What the numbers don’t tell you
We state two limits plainly in every report.
- An answer from an API is not exactly what a user sees in an app. Apps add their own instructions, location and personalisation. We use API results to track change over time, and check samples in the real products.
- Being recommended is not a booking. We track enquiries from your website separately, with their source, so you can see whether visibility turns into patients.
None of this is complicated. It takes discipline to do every week, and to resist the better-looking number when it isn’t the true one.