Define your prompt sample and count each observation
Freeze buyer-task prompts, record the same context and count brand mentions separately from website citations. Use repeated runs to describe variation within the declared sample. A failed run or missing answer remains in the denominator record rather than disappearing from the report when it makes the findings less convenient.
Worked example — illustrative data.
| Run context | Prompt and surface | Observation fields |
|---|---|---|
| Sample A; date/time blank | Compare scheduling tools for a small studio; chosen AI surface | Brand mentioned? Citation URL? Answer saved? |
| Sample B; same stated settings | Repeat exact prompt | Different sources or wording? |
| Sample C; chosen market recorded | Where can I book a pottery workshop? | Business identity and location accuracy |
Count the declared sample
For an illustrative six scheduled-run sample, suppose five answers were captured and one run failed. If the brand appeared in two captured answers and the website was cited once, report 2/5 mentions and 1/5 website citations among captured answers, plus the separate 1/6 failed-run count. Keep the scheduled total visible. Do not treat a failed request as a displayed answer excluding the brand.
Use the monitoring investigation register for follow-up triggers and report changes; this page defines the sample and counting rules.
Choose prompts from real buyer tasks and freeze their wording before the reporting period. Record market, language, account context, surface, model where shown and whether a fresh conversation was used. Save the answer and its cited pages so another reviewer can inspect the observation. Keep failed or unavailable runs in the log. Compare like-for-like samples and disclose changes to the prompt set instead of quietly removing unfavorable questions.
Use a status codebook before counting the sample
Agree what counts as a mention, an own-site citation, an uncertain identity and an unassessable capture before collection. In the six-attempt example, leave the failed run outside the five assessable answers while showing it in the scheduled total. If an answer is cut off before the source list, mark citation status unassessable rather than entering a convenient zero.
Decision: Use repeated samples to find factual errors or content questions worth investigating. Do not convert the log into a guaranteed ranking position, a market-share estimate or a sales claim.
A mention can cite a source other than the brand website: A correct brand mention can cite a third-party page rather than the brand website. Count the mention and citation as different fields, and check whether the cited page supports the claim.
Can I divide referral visits by these sampled citations?
No. Your chosen answer captures and the website’s referral visits describe different populations. Report them separately unless a valid link and compatible denominator exist. A citation in your sample does not establish that it generated a particular visit.
Continue with Guides for stores, software teams and local businesses, AI search audit for stores and SaaS brands and Monitor AI answers without promising rankings.
For Google AI supporting pages and eligibility, see Google AI features documentation (checked November 2026).