How to measure AI visibility without misleading yourself

“Our AI visibility is 60%” leaves the important questions unanswered. Sixty percent of which answers? Was the brand named, cited, recommended, or mentioned as an option to avoid?

A useful report separates observed answers, identifiable website visits and qualified business outcomes. Start with the counting rules, then use the same rules when you compare periods.

Define one observation

One observation is one prompt, on one surface, under recorded conditions, at one time. Keep Google AI Overviews, Google AI Mode and ChatGPT search in separate series. Label API responses separately from consumer-product observations.

Save the prompt, date, location context, language, account state, search setting, response and cited URLs. If a capture fails, record the failure. If a valid search produces no AI answer, record that too. Those events need different treatment.

Build the question panel from customer research and sales records. Include discovery, comparison and suitability questions in proportions that fit your audit. Freeze the panel for comparisons and document later additions. A hand-selected panel describes those questions; it does not represent every AI conversation.

Keep mentions, citations and recommendations separate

Count a brand at most once per answer for a mention rate. Record whether it was recommended, described neutrally or discouraged. Use an alias map for product names and brand families, and keep the original response for inspection.

A domain citation is a separate event. An answer can mention your brand while citing someone else's site, or cite your site without recommending your product. Neither event establishes a visit or sale.

Metric	Calculation	What it describes
All-valid mention rate	Answers naming the brand ÷ valid observations	Presence across the observed surface
Conditional mention rate	Answers naming the brand ÷ observations with an AI answer	Presence when an answer occurs
AI-answer occurrence	Observations with an AI answer ÷ valid observations	How often the surface provides an answer
All-valid citation rate	Answers citing your domain ÷ valid observations	Citation frequency in the panel

Show how the denominator changes the result

Suppose you attempt 60 observations. Six captures fail. Of the 54 valid observations, 14 have no AI answer, leaving 40 answers. Eighteen name your brand and nine cite your domain.

The all-valid mention rate is 18/54 = 33.3%. Among AI answers it is 18/40 = 45%. AI-answer occurrence is 40/54 = 74.1%. The citation rate is 9/54 = 16.7%. These fictional numbers answer different questions; none is market share.

Exclude failed captures from the outcome denominator and disclose them separately. Keep valid observations without an AI answer in the all-valid rate. When a denominator is zero, report the result as undefined rather than 0%.

Unequal runs also change weighting. If question A receives 9 mentions in 10 runs and B receives 1 in 2, pooling gives 10/12 = 83.3%. Weighting the two questions equally gives (90% + 50%)/2 = 70%. Choose and document the weighting before comparing periods.

Add platform and website evidence

Search Console: Google's AI-feature traffic is included in overall Web search reporting. That aggregate cannot isolate AI-answer visits for this panel. Google reporting guidance.

Bing Webmaster Tools: Microsoft's AI Performance public preview reports citations, cited pages and sampled grounding queries across supported experiences. Those measures are not clicks or ranking positions. Microsoft's announcement.

ChatGPT referrals: OpenAI documents the parameter utm_source=chatgpt.com on referral URLs. Preserve it through redirects and track landing pages and qualified outcomes. This measures identifiable visits, not everyone who saw the answer. OpenAI publisher FAQ.

Keep self-reported discovery separate from click attribution. A later branded search may follow an AI answer, but that possibility is not enough to assign the conversion to AI.

Compare periods honestly

Use the same questions, surfaces and conditions. Show the numerator and denominator next to each rate. Moving from 2/10 to 3/10 is one additional observation, even though the relative increase is 50%.

Repeated answers can be correlated. Avoid a conventional survey margin of error for a hand-selected prompt panel. Keep raw responses and a log of content changes, model updates, capture failures and other campaigns. Before-and-after improvement alone does not establish incremental revenue.

Choose the next action from the evidence. If facts are wrong, inspect their sources. If referrals arrive but enquiries are unsuitable, review the page and offer. If useful citations bring few visits, check whether the answer resolves the question without a click.

Sources checked October 4, 2026. All numerical examples are illustrative. Use the calculator below to test the denominators in your own report.