How to measure AI search visibility: mentions, citations and customers

The useful change is better measurement

If an owner asks whether AI search is bringing the right people to their business, a screenshot of one favourable answer does not settle it. You need to know what appeared, where it appeared, whether someone visited, and whether the enquiry was relevant. Those are different questions, and they need different evidence.

Google documents a dedicated Generative AI performance report for Search, with worldwide rollout completed by 31 August 2026. It reports impressions for links shown in AI Overviews and AI Mode, with page, country, device and date dimensions. It may not appear if your site has too few impressions. Use it to observe exposure, not as a sales report.

Define the event before counting it

• Correct mention: the answer identifies your business accurately, even if it does not recommend it.
• Recommendation: your business is presented as a suitable provider for the stated need.
• Citation: the answer links to a particular source, which may be your page or another site's profile.
• Referral visit: your site records a visit attributed to an AI source; some journeys will not preserve that referrer.
• Qualified enquiry or customer: your business records a relevant conversation or sale using a consistent definition.

A provider can be mentioned without its site being cited. A citation can appear without a visitor clicking it. Report these events in separate columns and describe missing attribution as unknown rather than zero. Avoid dividing unrelated counts from different periods or samples to create an apparent conversion rate.

ChatGPT search access is a separate check

OpenAI identifies OAI-SearchBot as its search crawler. GPTBot relates to potential training use, while ChatGPT-User handles certain user-triggered visits. These roles are distinct. Allowing search access does not promise a recommendation, and allowing training is not a condition for allowing search.

My practical starting point is to inspect the pages a buyer would need: the service or product, who provides it, who it suits, what it costs or how it is quoted, and how to enquire. Check what actually loads publicly, including translated versions. A description hidden behind an account or an unfinished route cannot help someone evaluate the offer.

Write a test that reflects the customer

Choose scenarios with a real buying decision: the service needed, useful constraints, location if relevant and language. Include UK, US, Canada and UAE scenarios only when those markets matter to the business. Keep English and Persian runs separate. Do not insert your brand into an unbranded recommendation question, because that changes what the test measures.

Record each prompt before testing and use it unchanged for the comparison. Note signed-in state and any relevant personalisation you can observe. Use authorised accounts and avoid customer details. An occasional manual check is a sample of that tool in those conditions, not a census of every answer your audience receives.

Run a small, honest visibility check

Choose a few real customer scenarios before testing. Keep brand-identity questions separate from unbranded provider recommendations. Use fresh conversations, record the platform and search mode, and save the answers and source links. Report the raw sample, such as two recommendations from five completed runs, instead of claiming a universal position in AI.

Then connect the observations to your own enquiries. A citation without a click can still be exposure, but it is not a visit. A visit without a qualified conversation is not a new customer. Ask how people heard about you and compare consistent periods, allowing for the time it takes them to decide.

Turn the monthly review into a decision

Review which questions produce accurate mentions, which pages are cited, whether referral visits reach useful destinations, and what qualified enquiries say about discovery. Choose one supported improvement, record its release date and compare the next period without claiming the change caused every movement. For the underlying page work, read the GEO foundations guide. An AI search visibility review can connect the evidence to priorities.