How Scanley measures AI search visibility
Scanley measures AI search visibility by asking the questions a metro or category actually asks, on ChatGPT, Gemini, Perplexity, Claude, and Google's AI Overviews, more than once each, and recording which institutions each answer named, recommended, or only mentioned. Every published figure states its engine, its date, its run count, and this method, so anyone can reproduce it.
| Engines measured | ChatGPT, Gemini, Perplexity, Claude, and Google's AI Overviews. | |
|---|---|---|
| Unit of measurement | Runs that named an institution, out of runs made, per question and per engine. | |
| Headline figure | Recommendations in written answers only; mentions and map listings are counted separately. | |
| Unrun engines | Reported as not run and excluded from every denominator. | |
What a research report is
Scanley Research publishes two kinds of measurement. A metro benchmark asks a metro's banking questions and reports which institutions the engines named, so a credit union in that metro can see the field it is competing in. A category scan asks one product category's questions across a set of institutions, so a marketing team can see who the engines already treat as the answer for auto loans, business checking, or first-time homebuyer programs.
Both use the same instruments as a client scan. The difference is scope: a client scan measures one institution against its metro, and a research report measures the metro or the category itself. Nothing in a research report is a claim about a client, and no client data is published without that client's decision to publish it.
How a number is made
Each question is asked on each engine we ran, and each engine is asked more than once, because generated answers vary run to run. A report states how many runs named an institution out of how many were made, not a single lucky screenshot. An engine we did not run is reported as not run and is excluded from every denominator; a missing measurement is never a miss.
Being named is not being recommended. When an answer names an institution in its written prose we record whether it recommended it, merely mentioned it, or named it adversely, and the headline counts recommendations only. A map or listings widget is recorded separately from the written answer, because appearing in a map pack means the institution has a branch nearby, not that the engine chose it.
What every report states
The engines run, the date of the run, the run count per engine, the question set, and the method sentence, printed on the page next to the figures. Reports that use invented or illustrative figures to show a layout are labeled as such and are kept out of the search index. A report whose figures were measured carries Dataset structured data so the measurement is citable as data, not only as prose.
Reports are drafted by the internal Scanley research instance and reviewed by a person before publication. When a report is corrected, the correction and its date are noted on the report itself and the updated date changes; the original publication date does not.
Questions this report answers
Why does the same question give different answers on different days?
Generated answers are sampled, not looked up, and each engine holds different evidence and weighs it differently. That is why every Scanley figure is a count of runs rather than a single answer, and why every figure names its engine and date.
Can a credit union reproduce a Scanley figure?
Yes. The report states the question, the engine, the date, and the run count. Asking the same question on the same engine around the same date should produce answers in the same range; the method exists so that it can be checked.