This page is the whole method — not a summary. Anyone, including your current agency, can re-run our measurement and get the same answer. That is the point.
One AI answer is one sample. Independent research finds a given prompt returns the same brand list less than 1% of the time — which is why the industry's standard evidence, a single screenshot, is worth roughly nothing. So we re-run every tracked query on a fixed cadence: up to 100 queries across three engines, daily, for about 9,000 responses per brand per month. Naver AI Briefing is not yet in production; its planned spec is described below. Every number we publish carries the sample it came from, and every response is stored and handed back as raw data.
ChatGPT, Gemini and Perplexity are sampled daily through official APIs. For Naver AI Briefing, our own September 2026 measurement (60 clinic queries) found a briefing on every query, but it named a clinic in only 33% of recommendation-type queries. The same run returned 5.0 citations per query across 68 unique domains. So the planned Naver spec treats it as a citation-source map rather than a mention metric — it is not yet in production and its start date will be announced. Engines we do not measure are named as not measured, and we do not reproduce the summary text itself.
Mention rate is the share of sampled responses in which the brand appears, counted separately for branded queries (your name in the question) and unbranded queries (the question a new customer actually asks). Share of voice is your mentions divided by all brand mentions in the panel. Confidence intervals are computed for every rate, and two brands whose intervals overlap are reported as tied — we do not sell a difference that the data cannot support.
Every source an engine cites is extracted and resolved to its domain. Aggregated across a panel, this produces the citation source map: which blogs, review platforms, directories, and sites feed AI answers in your category. One market's map recently resolved 374 unique citation domains.
A measurement that stops at a number is half a product. Each cycle we compare two things: where you are absent in the citation map, and the conditions our own runs show actually produce a naming — in the September clinic run, a brand was named when the cited source carried that brand in its title, and not otherwise. From that we write page specs, draft copy and an application guide: one priority per cycle, not a backlog.
Three limits hold this in place. We write drafts; the provider approves and publishes them, and each approval is recorded against a document version. Banned-phrase rules are applied to every draft before it leaves us. And when we test whether our own prescription moved a number, we hold back untouched control queries — advice that cannot be checked against a control is not evidence, it is a story.
Before any work ships — yours or your vendor's — the baseline is preserved. After it ships, we re-measure with the identical panel and call the result: moved, flat, or worse, with the confidence interval shown. Where panels allow, treated queries are compared against untouched control queries, so ordinary market drift is not mistaken for effect.
Every report ships with its raw data: query text, timestamp, engine, full response text, and every citation URL — for every run, exportable as CSV or JSON. Nothing we publish is un-re-checkable.
We publish source domains, not provider rankings. For each city × vertical we release how often each cited domain appears — that ordering is of websites, not of clinics. An individual brand's own figure goes only to that brand. Published query sets are paired with unpublished verification queries, rotated on a schedule, so a map cannot be gamed by targeting its questions. These figures record how often AI services output a name across a defined query set and period; they are not an assessment of clinical quality and not a recommendation.
We do not operate your channels, sell content retainers, or run ads. We do not charge on results, and we do not guarantee AI visibility — anyone who does is selling something measurement cannot honestly promise. Our only guarantee is that you will know exactly what happened, with the data to check us.