How to measure AI visibility without fooling yourself

Measure five things separately: whether the answer names your business (mention), links to your site (citation), recommends you as a choice (recommendation), sends a measurable visit (referral visit), and produces a stored enquiry. Report each with its denominator, such as “named in 4 of 12 completed unbranded checks,” and keep the raw log.

Five measures, five definitions

Measures and their limits
MeasureDefinitionLimit
MentionThe answer text names your businessSays nothing about tone or accuracy on its own
CitationThe answer links to a page on your siteA citation is a count, not a rank
RecommendationThe answer presents you as a suggested choiceWording varies; record the excerpt
Referral visitA session your analytics attributes to an AI platformMany AI-influenced visits arrive without a referrer
EnquiryA contact your system stored, with a source you can trustSelf-reported sources need a clear question on the form

A sample observation log

QTypeMarketLangPlatformModeDateRunMentionedRecommendedOwn URLThird-party URLCompetitorsNotes
CL-01unbrandedCalgaryENChatGPT Searchsearch on18 Nov1nono—directory.example2 namedAnswer names two other brokers
CL-01unbrandedCalgaryENChatGPT Searchsearch on18 Nov2yesno—directory.example2 namedNamed in list, third-party cite
CL-01unbrandedCalgaryENPerplexitydefault18 Nov1nono—none cited3 namedNo sources cited: recorded as none
CL-04brandedCalgaryENGoogle SearchAI Overview shown18 Nov1yesn/aown site——Accuracy check: hours correct

Unbranded result: named in 1 of 3 completed checks. Branded accuracy is reported separately.

Sample observation log (illustrative data)Each row is one run of one question: question ID, branded or unbranded, market, language, platform, mode, date, run number, whether the business was mentioned, whether it was recommended, any link to its own site, any third-party URL cited, competitor names, and notes. From these four rows: named in 1 of 3 completed unbranded checks; one answer cited no sources.

Rules that keep the numbers honest

  • Separate branded questions from unbranded discovery questions.
  • Separate your own visibility from a client’s if you run both.
  • Record platform, mode, date, language and the city written into the question.
  • Keep inconclusive results and failed runs in the log.
  • Report Google Search Console and Bing Webmaster Tools data under their own product names. Neither measures ChatGPT.

For clinics and other regulated practices

Measure discovery and access (was the right location named, did the link lead to booking), not clinical claims. Keep patient information out of logs entirely. See the clinic example.

Primary sources

Checked on 18 November 2026. Platform behaviour changes; we re-check these references when we revise the page.