Every result we show is obtained by asking AI engines the questions a real customer would ask, on a published list of questions, on published dates, and recording what each engine answered. Nothing is estimated.
01Which engines we query
We query six engines, chosen because they are the ones people actually use, not because they are easy to query: ChatGPT, ChatGPT Search, Google Gemini, Google AI Overviews, Google AI Mode and Perplexity. Each is used the way a customer uses it: the public product, a fresh session, no account history, no system prompt, one question at a time. Microsoft Copilot is not in the set because it could not be collected reliably during our tests; it will be added when it can be.
02Which questions we ask
Thirty per market. A market is one type of business in one city. The questions are written before the first collection and do not change during a measurement, so that month-to-month comparisons compare like with like. They cover the ways a real customer asks: by need ("best place for veneers in Miami"), by neighborhood ("cosmetic dentist near Brickell"), by attribute ("med spa in Scottsdale with a nurse practitioner"), by comparison ("is X or Y better for Invisalign in Dallas"), and by budget. We do not ask questions that contain the client's name except in three control questions, reported separately, which check whether the engines know the business exists at all.
03How we read an answer
Four things are recorded for every answer, with the same definitions in every report:
- Named — the answer mentions the business by name, in any position.
- Recommended — the business is among those the engine suggests going to, not mentioned in passing or as a comparison.
- Position — where it appears when the answer is a list: first, top three, or lower.
- Sources — the pages the engine cites or links. Who gets cited gets recommended, which is why the source table is the working document of the whole service.
Names are matched by a person, not only by string: "Dr. Smith's practice" and "Smith Dental Arts" are the same business when they are, and two businesses with similar names are kept apart.
04When and how often
The first collection is the baseline. Then the same thirty questions go to the same six engines once a month, on dates published in the client's report. Engines change their answers from day to day; a single collection is a snapshot, three are a trend, and we say which is which. Nothing is extrapolated from one run.
05Where the method comes from, and what we add
The engine set, the counting rules and the principle that questions are published before collection are those of Visibilitas, an independent observatory of AI answers founded by the same founder. Visibilitas publishes rankings and open data under a CC BY 4.0 license and sells nothing; BrandPrompted applies the same method to individual businesses, privately, for a fee. BrandPrompted has no input into the observatory's results, and the relationship is stated on the observatory's transparency page.
What we add is the source table and the recommendation: an observatory records; an agency decides what to do about it.
06What the numbers cannot tell you
They cannot tell you why an engine named someone: the engines do not explain themselves, and we do not pretend to read their minds. They cannot promise a position next month. They can tell you, exactly, whether you were named today, who was, and from which sources — which is enough to know what to do, and enough to know afterwards whether it worked.