Measurement guide

How to measure AI search visibility

A practical measurement plan for teams that need to know where they appear, what gets cited, and which visits turn into leads.

Last reviewed 24 September 2026 · 9 min read

The short answer

Use four separate evidence streams: Google's AI-feature impressions in Search Console, dated samples of actual AI search answers, identifiable AI referral sessions in GA4, and tracked leads. Keep their counts separate. An impression is not a mention, a citation is not a visit, and a visit is not a lead.

3D magnifying glass, search result card, chart, and linked source representing AI search measurement

Key takeaways

  • 1

    Start with a fixed set of buyer questions and a consistent reporting window; changing the sample every week makes trends hard to interpret.

  • 2

    Search Console's Generative AI report measures impressions of links in supported Google Search AI features. It does not provide a prompt-by-prompt citation list or a separate AI click total.

  • 3

    Save actual answer observations with the question, date, region, search mode, brand mention, and linked source. A model API response is not a substitute for a consumer search answer.

  • 4

    Use GA4 for identifiable referral sessions and configured key events. Report these beside answer samples and Google impressions, never as one combined percentage.

Before opening a dashboard, decide the questions, pages, region, language, and date window you will revisit. Give each source its own tab or section in the report. This makes a later change traceable to the source that actually recorded it. Write down who collects each stream and when. For example, a content lead may save answer samples, an analyst may export Search Console and GA4, and a sales owner may verify qualified leads in the CRM. Agree on the meaning of a mention, recommendation, and linked citation before anyone starts counting. If a search experience changes its mode or interface, mark the break in the series instead of presenting the next sample as directly comparable.

Four signals, four different questions

Start with a question you can answer with a real source. Each signal below observes a different part of the buyer journey, so none is a universal AI visibility score.

Four separate 3D panels for Search impressions, saved AI answer observations, referral visits, and tracked leads
Four separate evidence streams answer different questions. Report their methods and units separately; this conceptual illustration shows no measured results.
01

Google AI-feature impressions

Search Console's Generative AI performance report, when available

Shows: A link to your site was shown in a supported Google Search AI feature.

Does not show: It does not identify the exact buyer question, count other assistants, or show a separate AI click total.

02

Answer mentions and source links

Dated observations in the actual search experience

Shows: For a recorded question and search mode, the answer named your brand or linked a specific page.

Does not show: One answer is a sample, not a fixed ranking or a count of everyone who saw it.

03

Identifiable AI referrals

GA4 Traffic acquisition, using AI Assistant and Session source / medium

Shows: An attributed session arrived with a recognizable source or campaign value.

Does not show: Missing referrers, redirects, privacy settings, and later direct visits can hide or change attribution.

04

Leads and other key events

GA4 key events or a reconciled CRM record

Shows: A tracked business action followed a visit under the chosen attribution rules.

Does not show: It does not prove that a particular AI answer caused the action.

Build a repeatable measurement routine

STEP 01

Freeze a small set of buyer questions

Choose questions from sales calls, Search Console queries, and the pages buyers compare before contacting you. Record exact wording, intent, target page, language, region, and the search experiences you will inspect. Keep the set stable for a few reporting cycles; add new questions in a separate cohort.

Do this: Example: a B2B SaaS team starts with 12 questions covering problem, comparison, and vendor selection stages.

STEP 02

Read Google's two reports for different jobs

Use the ordinary Search results Performance report for queries, pages, clicks, and impressions. If your property has the Generative AI performance report, use it for impressions of links shown in AI Overviews and AI Mode, grouped by page, date, country, or device. The AI report does not provide a question-by-question citation log or a separate click figure.

Do this: Export the same date range each week and record which report, dimension, and timezone each number came from.

Two reports, two jobs

Illustrative report map

1

Search results Performance

Search Console → Performance → Search results

Read: Queries and pages with standard clicks and impressions

Use it to: Find the buyer questions and pages worth checking in AI answers.

2

Generative AI performance

Open the Generative AI performance report, if available

Read: Impressions for linked pages in supported AI Overviews and AI Mode

Use it to: See which pages appeared as links in those Google Search features.

This map is an original explanation of the reports, not a Search Console screenshot or TopFound result. Read Google's Generative AI report documentation.

Check before comparing exports: Google groups the AI report's chart by property but its Pages table by page, so totals can differ. Its dates use Pacific Time; check your GA4 property timezone before lining up daily rows. An unavailable AI report is 'not available,' not zero. Google also notes that exported ~ or - values become zeros, so keep the on-screen context with the export. Save the applied filters and whether the number is for a property or a page. If two pages from the same site appear in one result, a property-level chart can count that result differently from page-level rows. Compare like with like across weeks; a change in grouping or filter can look like growth even when the underlying visibility did not change.

STEP 03

Save answer observations with their conditions

Run the fixed questions in the consumer search experiences you care about, where access and terms permit. Save the answer text or screenshot, linked source URLs, date, region, language, and whether search was active. Mark a brand mention only when the brand is named; mark a citation only when the answer links to a source page. Count recommendations separately. Do not substitute a general model API response for a product's search answer.

Do this: Repeat important questions. If an answer changes, inspect the saved responses before calling it a trend.

Evidence log template

One answer observation, recorded properly

Blank example
Exact buyer question
Paste the question verbatim
Experience
Platform and search mode
Conditions
Date, region, language, account context
Observed answer
Saved text or screenshot link
Brand outcome
Mentioned, recommended, or absent
Source evidence
Linked URL, if one was cited
Blank log layout — fill it with the answer you actually saw. Keep the saved answer and linked source URL together so another reviewer can verify the observation.

STEP 04

Find identifiable referrals and leads in GA4

Open Reports → Acquisition → Traffic acquisition. GA4's AI Assistant default channel groups recognized assistant referrals, while Session source / medium lets you inspect individual sources such as ChatGPT. Google's AI Overviews and AI Mode are counted in Organic Search, not the AI Assistant channel. OpenAI says ChatGPT search referral URLs include utm_source=chatgpt.com; confirm your property's values and configured key events.

Do this: Keep unknown or direct traffic outside the AI-referral total. Do not estimate it as AI traffic without independent evidence.

Google documents the Traffic acquisition report and the AI Assistant channel. OpenAI documents the ChatGPT search referral parameter. Check the values in your own GA4 property.

STEP 05

Turn the four streams into one decision log

Report the streams side by side, with their own denominators and date ranges. For each important page, record what you changed, what was observed afterward, and what you still cannot tell. Prioritize a page when buyer questions are relevant but answers lack a useful source, or when identifiable visits arrive but the page fails to convert.

Do this: Treat changes as hypotheses to test. A before-and-after chart alone does not establish causation.

What to investigate next

If you seeA comparison question names competitors but links no useful page from your site.

Then checkReview the comparison page for missing facts, product evidence, and a clear answer to that question; then repeat the same answer sample.

If you seeIdentifiable referrals reach a relevant page but no key event appears.

Then checkConfirm the event fires, then review whether that landing page answers the visitor's likely next question and has a clear action.

If you seeThe Generative AI report is absent or a table export shows zeros where the UI showed ~ or -.

Then checkRecord the state as unavailable or suppressed rather than treating it as proven zero visibility.

A worked example

Fictional worked example — not a TopFound or client result

A SaaS team checks 12 fixed buyer questions in two search experiences during one week. Its Search Console and GA4 reports use the same calendar window, but the rows below still have different denominators.

Saved answer observations
24 answers; 5 name the brand; 2 link its pages
Google AI-feature impressions
84 impressions in Search Console
Identifiable AI referrals
4 sessions in GA4
Tracked demo requests
1 key event from an identifiable referral session

What the team does next: The two linked answers both cite the product page, while comparison questions cite other sources. The team reviews its comparison page for missing product details and evidence, then repeats the same questions. It does not add 84 + 24 + 4, or claim that the one demo came from a particular answer. In its decision log, the team records the exact comparison questions, the existing page section that should answer them, the facts the product owner must verify, and the date of the revision. At the next review, it checks the same question set and the same Search Console and GA4 definitions. A different citation pattern would be a reason to keep investigating, not proof that the page edit alone caused the change.

Fields for a weekly reporting sheet

One row per answer observation keeps the evidence auditable. Keep aggregate Search Console and GA4 totals in separate rows or tabs so their denominators stay clear.

  • Question and buyer stage
  • Platform, search mode, region, language, and observation date
  • Brand mentioned, recommended, or absent
  • Linked source URL and saved answer evidence
  • Google Search results clicks and impressions, labeled separately from AI-feature impressions
  • GA4 AI Assistant channel, Session source / medium, sessions, and configured key events
  • Page changed, change date, next check, and unresolved uncertainty
Download the blank answer log (.csv)

This file covers answer observations only. Keep Search Console and GA4 exports in separate tabs with their own sources and date ranges.

What this method cannot prove

  • Search answers vary with wording, mode, location, time, and sometimes account context. A sample shows what happened under recorded conditions.
  • Search Console's generative AI report covers supported Google Search features, not ChatGPT, Claude, Gemini apps, or every AI-generated answer.
  • GA4 only sees measurable site visits. A citation without a click and an answer that mentions a brand without linking it do not create a referral session.
  • Direct or unassigned traffic is not automatically AI traffic; attribution can be incomplete even when the visit was real.

Common measurement questions

Can GA4 show every visit from ChatGPT or another AI answer?

No. It can show sessions that arrive with source information it can attribute. Missing referrers, privacy settings, redirects, and later direct visits can leave real activity unattributed. Keep direct and unassigned traffic separate.

Does Search Console show the exact AI question that cited my page?

The Generative AI performance report groups link impressions by page, date, country, or device. It does not provide a question-by-question citation log. Save answer observations separately for the exact questions you test.

Should I turn these four signals into one AI visibility score?

Not by simply adding them. An impression, an answer observation, a referral session, and a lead have different units and denominators. Show each trend with its source and sample conditions, then make a decision from the pattern across them.

How often should I sample AI search answers?

Pick a manageable set of buyer questions and repeat it on a fixed schedule, such as weekly. Record the platform, search mode, location, date, prompt, and cited URLs so changes in the sample are distinguishable from changes in visibility.

Primary sources

Product reports and definitions can change. These official sources were checked for this guide on 24 September 2026.

Want a measurement plan for your site?

TopFound can define the buyer questions, document answer observations, and connect them to search, referral, and lead reporting.