Measure AI visibility without inventing demand

Measure AI visibility with separate records for what an answer says, whether someone visits, and whether a suitable buyer inquires. For a B2B SaaS founder, those records support different decisions. A citation may justify improving a linked page; a qualified inquiry may justify a sales conversation.

Start with buyer questions you can trace to sales notes, search data, or product research. Then keep the question set fixed long enough to compare the same work. A dashboard percentage means little if its questions, settings, or denominator changed between reports.

What should count as visibility?

Count mentions, owned citations, and recommendations independently. An answer can name your product without linking to it, or cite your documentation without recommending your software.

Observation	What qualifies	What it doesn’t establish
Brand mention	The answer identifies your company, after checking ambiguous names	An endorsement or accurate description
Owned citation	A displayed source link points to your configured domain	A visit or recommendation
Recommendation	A reviewed passage suggests or shortlists your product for the stated need	Fit for every buyer
Referral visit	Analytics records a visit with an identifiable source	Exposure to everyone who saw the answer
Qualified inquiry	An inquiry meets your recorded buyer-fit criteria	A sale or a known discovery path

Read the surrounding answer. A product can appear as an example to avoid. Save the passage that establishes its role, and check material claims against current product facts. An unreviewed recommendation is unknown, even when a name detector finds the company.

Keep third-party citations separate from links to your own site. If a cited comparison page mentions your product, save the relevant page passage too. The search and SEO guide explains how buyer questions connect to useful owned pages; this article adds the observation and comparison rules.

What goes in the measurement sheet?

Use one row per attempted answer check, including failed attempts. Give each buyer question a stable ID, keep its exact wording, and record where it came from. Remove customer names and other personal details from the shared sheet.

Separate branded questions, which name your company, from unbranded questions. “What does [company] do?” can test description accuracy when the prompt names it. Discovery questions should describe the buyer’s task without supplying your company as the answer. The missing-from-ChatGPT diagnosis uses that distinction before deciding which page needs work.

Copy these fields into a sheet:

Question: identifier, exact text, buyer intent, source, branded/unbranded, version
Check: identifier, timestamp, platform, interface, model if shown, search setting
Context: language, approximate geography, session/memory state if known
Evidence: saved answer, displayed citation URLs, capture completeness
Status: completed, failed, or missing; error or missing-data reason
Review: mention, owned citation, recommendation, material error
Review evidence: supporting passage, fact source, reviewer, unknown reason

Use yes, no, or unknown for each reviewed outcome. A partial citation capture with no visible owned link doesn’t prove absence. Keep the answer evidence in an approved location, and share only the fields needed to inspect the finding.

Consumer and API checks need separate rows and summaries. ChatGPT’s search documentation explains that memory and location can influence searches. Record the session context you can observe, and mark hidden settings unknown.

If you use an API, record its configured model, retrieval tool, and location settings. The API web-search documentation distinguishes displayed inline citations from the fuller list of consulted sources. Count the displayed citation for that metric, rather than treating every retrieved URL as a citation a buyer saw. An API answer remains evidence about that configuration, not a substitute for a consumer answer.

Which denominator belongs beside the rate?

Use the number of answers assessed for that specific outcome. Failed attempts aren’t observed non-mentions, and a completed answer with an unreviewed recommendation isn’t an assessed recommendation.

Illustrative calculation, not a business result. Suppose you plan 12 checks of one question on one recorded surface. Ten complete, one fails, and one is missing. Six completed answers mention the company. The mention rate is 6/10, with the failed and missing attempts reported separately.

Suppose only eight completed answers have usable citation capture, and three cite your domain. The owned-citation rate is 3/8. If seven answers receive recommendation review and two suggest the product, report 2/7, with three completed answers unreviewed. These denominators differ for a reason. Adding the outcomes into one “visibility score” would conceal it.

Choose the repeat plan before collection, within your access and budget. Record a rerun as a new attempt, linked to the failed check; preserve the failure. Two disagreeing answers show variation in those answers. Agreement doesn’t prove that buyers will always see the same response.

For a before-and-after comparison, keep exact questions, surfaces, settings, and planned coverage matched. If a panel is incomplete, finish matched usable checks before reporting a comparable rate. If the model or question set changes, establish a new baseline. Label a changed count as an observation, and record other changes that could explain it.

What do native reports add?

Native reports describe activity you didn’t create with your own question panel. Keep their metrics alongside the panel, with the reporting window and documented coverage attached.

Google’s Generative AI performance report covers impressions from AI Overviews and AI Mode. Its Help page says worldwide rollout completed on August 31, 2026, but a property may lack the report if it hasn’t received enough impressions. The main Web performance report also includes AI-feature traffic. Don’t label its total clicks as AI-only clicks.

Before exporting, record the selected dates, dimension, and any unavailable cells. Values shown as ~ or - become zeros in the export. Keep a screenshot or companion note, and mark those entries unavailable in your working sheet. A missing report or unavailable value isn’t evidence of no AI exposure.

Bing’s AI Performance public-preview documentation covers citations across supported AI surfaces, including Copilot and AI summaries. It explicitly separates citation counts from placement and ranking. Grounding queries are a sample of citation activity, not a complete record of buyer prompts. Check access in your property before promising this data in a report.

How do visits connect to real inquiries?

Inspect identifiable referral visits by source, landing page, and date window. Analytics guidance on traffic-source dimensions distinguishes session source from first-user source. Use session scope for the visit you’re examining, and keep the chosen definition consistent across periods.

Check your collection setup before reading the counts. Known visits from quality-assurance checks and form tests belong in a test record. Internal-traffic filtering guidance warns that excluded data cannot be recovered through Analytics; test filter behavior before activating it. For historical contamination, document the known test window and the limits of any retrospective exclusion.

Keep attribution gaps visible. A buyer may return directly, use another device, or decline a discovery question. Ask how they found you where appropriate, retain the answer privately, and publish aggregate counts. Record unknown discovery as unknown. Don’t relabel all direct traffic as AI traffic or infer an individual visitor’s identity from an anonymous session.

Define qualification before counting inquiries. Track inquiries, booked calls, held calls, and opportunities separately. A tested booking event establishes that tracking works; it doesn’t establish a buyer inquiry.

What should you do with the baseline?

Choose a next action from the evidence. Correct a material product error, investigate a cited page that misses the buyer’s decision, or repair a broken inquiry path. If coverage is incomplete, collect the missing checks before announcing a change.

Start with one buyer question and one recorded surface. Save the answer, classify its outcomes, and test how the sheet handles a failed attempt. The AI visibility baseline kit provides the companion worksheets. The JAC search experiment shows the observation dates and reporting limits behind a separate public baseline.

The audit handoff guide explains what belongs in the agreed work and completion checks. If the baseline exposes page or technical work you want implemented, AI Search & SEO handles it with monthly reporting. The optional AI Audit can scope a one-time set of fixes. Finish the first inspectable record before deciding whether outside help is needed. The AI search overview connects measurement to ongoing page work. Book a call if you want to discuss the findings and scope.