An AI visibility check is a repeated sample of answers, not a stable ranking report. The same question can produce different citations across wording, date, location, account state, and engine version.
Build a reproducible prompt panel
Choose a small set of real buyer questions across tasks: explain, compare, troubleshoot, and choose. Save the exact wording, market, language, date, engine/product mode, and whether web search was enabled. Run neutral paraphrases at a planned cadence and retain the full answer and cited URLs.
Score separate outcomes:
- Was the brand named?
- Was the site cited, and which page?
- Did the source support the statement attached to it?
- Was the description accurate, incomplete, or wrong?
- Did the answer mention a competitor instead?
Report the denominator and the number of runs. “Cited in 3 of 10 checks” is interpretable; “ranked #1 in AI” usually is not. Do not combine engines or prompt types into one score unless the method explains the weighting.
Turn absence into a content question
When a relevant page is not cited, inspect whether it is crawlable, answers the exact task, includes evidence, and is corroborated elsewhere. Then make one change that improves usefulness for people—not a phrase inserted solely to influence an answer—and rerun the same panel plus a holdout question.
AI answer samples are volatile and do not prove organic discovery or business impact. Track referrals and qualified actions separately; a citation is not a visit, and a visit is not a conversion.
Optional AI-assisted work brief — AI visibility observation protocol
Run the same neutral queries across a defined set of AI/search surfaces, locale, language, account state, and date. Record prompt, full response, citation URL, brand mention, and whether the page actually supports the answer. Repeat enough times to observe variability and label this as sampled observation, not rank. Never treat absence in one response as proof of invisibility.
Worked example — Record variability honestly
Ask the same neutral question several times across dates and locales. One answer may cite the portfolio and another may not. Record sample size, phrasing, engine, location, and citation. Report the pattern as an observation with variability, not a stable position or promise. Improve the source page when it lacks clear evidence, then recheck later.
What Google’s AI guidance does and does not promise
Google says its AI Search features rely on Search fundamentals and that eligibility does not guarantee display. It also rejects special markup, llms.txt, and formulaic chunking as required shortcuts. Track sampled answers as observations, not rankings or guaranteed citations. Official guide.
AI disclosure: AI tools assisted with drafting. No visibility or citation result is claimed.