How to Measure AI Visibility and GEO Results

Author: Lucky Oleg | Published Updated

AI visibility is not one metric, so measuring AI visibility means tracking several outcomes separately. A platform may mention your brand without linking, cite a page without sending a click, or send a small number of visits that convert unusually well. Combining all of those into one number hides more than it reveals.

A useful GEO measurement system keeps four outcomes separate:

  1. appearances in platform-provided reports;
  2. citations or mentions observed in a repeatable prompt sample;
  3. referral sessions from AI platforms;
  4. business outcomes from those sessions.

What each source can prove

SourceWhat it measuresWhat it cannot prove
Search Console generative AI performance reportsImpressions in AI Overviews and AI Mode, and in Discover’s generative AI features, with the pages, countries, dates, and (for Search) devices involvedClicks, CTR, position, or the queries behind those impressions; visibility in ChatGPT, Perplexity, or other platforms
Standard Search Console Performance reportOverall Google Search clicks, impressions, queries, and pages; Google says AI Overviews and AI Mode activity is included in the Web search typeWhich of those clicks, queries, or impressions came from AI features, because the report has no AI-feature filter
Web analyticsVisits and conversions where a referrer or campaign parameter is preservedMentions and citations that produced no click
Manual prompt panelWhat selected prompts returned under recorded conditionsPopulation-wide share of voice or a stable ranking
Third-party monitoring toolThe vendor’s sampled prompts, locations, accounts, and matching logicComplete platform coverage unless the methodology demonstrates it

Google Search Console in 2026

Google announced its Search generative AI performance reports on 3 June 2026 and first rolled them out to a subset of sites. A note on that announcement, repeated on the Search Console Help page, says that as of 31 August 2026 the reports have rolled out to all websites worldwide.

There are two reports: one for Search, covering AI Overviews and AI Mode, and one for generative AI features in Discover. They show:

  • impressions, meaning how many times links to your site were shown in a generative AI feature;
  • the pages shown;
  • countries;
  • dates, so you can follow the trend over time;
  • devices (Search report only).

They do not show clicks, CTR, average position, or the queries that triggered the feature, and Google excludes Search Labs experiments. Google’s Search Analytics API reference lists no generative AI search type, so as of this review the data comes from the Search Console interface and its export button rather than the API.

Google’s AI features guide also says activity from AI Overviews and AI Mode stays included in the overall Performance report under the Web search type. A click on a link in an AI Overview or AI Mode counts as a click there, and every link in an AI Overview shares the overview’s single position, but that report has no filter that separates AI-feature traffic. Therefore:

  • use the generative AI report for Google-specific AI visibility trends;
  • if the report does not appear for your property, record that rather than estimating: Google’s Help page lists the gradual rollout, too few impressions in these features, or the site being excluded from them as possible reasons;
  • do not describe ordinary Web impressions or clicks as AI Overview impressions or clicks;
  • preserve screenshots or exports with the report name, property, filters, and date range.

An impression is not necessarily a visit, and a visit is not necessarily a lead. Pair the report with analytics and conversion data.

Track AI referral traffic

Start with GA4’s AI Assistant channel

Google introduced a native AI Assistant channel on 13 May 2026. For recognized assistant referrers, it assigns medium ai-assistant and campaign (ai-assistant). You do not need to build a custom channel just to see this traffic.

Use this starting view, following Google’s Traffic acquisition documentation:

  1. Open Reports → Acquisition → Traffic acquisition. Navigation may differ if the report collection has been customized.
  2. Select Session default channel group and inspect AI Assistant for a fixed date range.
  3. Switch to Session source / medium to inspect the underlying sources. Keep session-scoped dimensions when comparing sessions.
  4. Review landing pages and your chosen lead or purchase key event for the same session population. Record sessions, sessions with that key event, and the resulting session key-event rate.

Google’s channel definitions explicitly keep AI Overviews and AI Mode in Organic Search. The AI Assistant row therefore cannot measure all Google AI traffic. A source of google / organic does not tell you whether the visitor clicked an AI feature or an ordinary search result.

Keep session source and page referrer separate

Session source describes acquisition for the session. Page referrer is the preceding URL attached to an event, which can be another page on your own site. A campaign source can survive when a referrer is absent, and a later page view’s internal referrer is not a new acquisition source. See Google’s traffic-source dimensions and event parameter definitions.

Inspect actual values before adding rules. These deliberately narrow examples cover selected assistant hosts, not Google’s full recognition list. The source expression expects a source value, without a protocol or path:

(?i)^(chatgpt\.com|chat\.openai\.com|perplexity\.ai|www\.perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai)$

For Page referrer, a URL-shaped field, use a separate expression:

(?i)^https?://(chatgpt\.com|chat\.openai\.com|perplexity\.ai|www\.perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|claude\.ai)([/?#].*)?$

Use matches regex where the report or exploration supports it. Google Analytics uses full matches by default, so the referrer expression includes the rest of the URL before its end anchor. The (?i) modifier ignores case. Escaped dots require literal domain dots, and the host boundary rejects lookalike domains. Do not paste a path such as bing.com/chat into a source-host expression or include all of google.com, microsoft.com or openai.com as assistant traffic.

Fixture inputField testedSelected-host match?What it establishes
chatgpt.comSession sourceYesThe recorded source belongs in this selected-source view
https://chatgpt.com/c/examplePage referrerYesAn event has an assistant-host referrer; it is not a count of new sessions
chatgptXcomSession sourceNoAn unescaped dot would incorrectly accept this value
chatgpt.com.example.orgSession sourceNoThe source is outside the selected hosts
https://example.org/?from=chatgpt.comPage referrerNoA hostname appearing in a query string does not make it the referrer host
google or https://www.google.com/Corresponding fieldNoBroad Google traffic stays outside this assistant-host subset
(direct) or a blank referrerCorresponding fieldNoThere is no assistant-source evidence in that field

The downloadable classification fixtures and offline check script test these string expressions against constructed examples. They do not simulate GA4 channel assignment. Review your property’s recorded values before adopting an additional rule.

If you need an additional source grouping, use a separate custom group and compare it with the native channel over the same dates. Google’s custom-channel documentation says report groupings can be applied retroactively. Changing the property’s primary channel definition is different: that dimension uses the new criteria from that point forward. Put a specific assistant rule before a broad Referral rule, preserve other channels, and inspect which existing rows moved. Do not add the custom total to the native total because the populations may overlap.

Treat attribution gaps as unknown

When a landing URL contains utm_source=chatgpt.com, inspect the resulting campaign-source value as well as any referrer. Our September review cited OpenAI’s publisher FAQ for automatic tagging; on 28 October 2026 that FAQ URL redirected to its Atlas retirement page, which no longer documents the tag. Verify current links you actually receive instead of assuming every ChatGPT click is tagged.

For every source, monitor:

  • sessions and engaged sessions;
  • landing page;
  • sessions with a chosen enquiry, signup, purchase, or other meaningful key event;
  • revenue where it can be attributed responsibly;
  • changes in referrer or campaign naming over time.

Referrer data is incomplete. Apps, redirects and browser behavior can remove attribution. Direct traffic with neither a source tag nor a recognized referrer remains unknown; no regex can reconstruct its origin. A tagged URL alone also does not prove an AI recommendation because tags can be copied or added by anyone.

Consent affects the observation too. Google’s consent-mode guide distinguishes blocked tags in basic mode from cookieless pings in advanced mode. Behavioral modeling is conditional on eligibility. Record your consent setup, reporting identity and data-quality notices rather than assuming all visitors are observable or every missing session was modeled.

In a test property, check your site’s consent states and a tagged landing page. Confirm that tags follow the consent choice, source evidence survives permitted redirects, and the lead event follows a successful submission. Reconcile accepted enquiries with your own records.

Build a repeatable prompt panel

Platform reports do not cover every system, so a controlled prompt sample is useful. It becomes evidence only when the method is recorded.

Create a sheet with these columns:

FieldExample
Query IDlocal-service-01
Exact promptBest way to evaluate a web design agency
IntentCommercial research
Platform and modePerplexity web search
Account stateSigned out
Location/languageUS / English
Run date2026-08-20
Brand mentionedYes/No
Linked citationURL or blank
Position in source listNumber or blank
Screenshot/evidenceStored reference

Download the prompt observation log CSV. It includes two example questions with unfilled result fields and a not run status. Add one row per prompt, platform and run. Record API versus consumer interface, search mode, model when disclosed, failures, and cached answers. An unavailable answer is unknown, and a cached repeat is not a new independent sample.

Use a fixed set of representative prompts. Run them on a consistent schedule. To track AI visibility in more than one country, run the panel once per country with its own location and language, and keep each country in its own rows. Do not silently replace weak prompts with easier ones, and do not count your own domain when it appears only because the prompt names the brand.

Because generated answers can vary between runs, report a range or repeated-sample rate when practical. A single favorable screenshot is an example, not a trend.

Keep the metrics separate

A compact monthly dashboard can contain:

  • Google generative AI impressions (Search, plus Discover where it matters);
  • linked-citation rate across the fixed prompt panel;
  • unlinked brand-mention rate;
  • AI-referred sessions;
  • AI-referred conversions and conversion rate;
  • top cited and top converting landing pages;
  • implementation notes for material content or technical changes.

Use the monthly measurement CSV and template instructions to record periods, numerators, denominators, evidence files and limits. Calculate a session lead rate as sessions with the selected lead key event ÷ matching sessions. Event count divided by sessions is a different metric because one session can trigger the same event more than once. Keep CRM-qualified enquiries separate from analytics events.

If you create a composite “AI visibility score,” publish the formula and keep every component visible. The score is useful for internal trend comparison only; it is not an industry-standard market share figure.

Avoid common measurement errors

  • Do not treat an AI-readiness audit as citation tracking. A crawler-access or schema check shows implementation conditions, not whether platforms cite the site.
  • Do not attribute every Google impression to AI. Use the generative AI report for AI-feature impressions, and remember that it has no click or query data.
  • Do not mix prompts, locations, and account states without labels. Personalization and availability can change outputs.
  • Do not promise a fixed time to results. No official platform publishes a universal GEO timeline.
  • Do not confuse training controls with search visibility. For OpenAI, GPTBot and OAI-SearchBot serve different documented purposes.
  • Do not report a mention as a referral. Visibility, traffic, and conversion are separate stages.

A practical baseline

Start with the data you can defend today:

  1. export the Search Console generative AI performance report (Search, plus Discover where relevant), or record why it is not available for the property;
  2. inspect Session default channel group = AI Assistant in GA4, then review session sources and any additional-source segment separately;
  3. define a small, fixed prompt panel around real customer questions;
  4. record the site, content, and crawler configuration at baseline;
  5. note substantive changes and compare the same measures over time.

Our AI Search Visibility Checker is a technical readiness screen; it can help identify crawl access, schema, and page-structure issues, but it is not evidence that a platform currently cites the site. For implementation priorities, use the complete GEO guide. If you need a managed measurement program, see the scope and limits of our GEO services.

Recommended tools

Recommended tools for this guide

Use these free tools to apply the ideas from this guide to your own website.

Useful info? Spread the Aloha:

Lucky Oleg

Lucky Oleg is the founder of Web Aloha, a web design & SEO agency helping businesses ride the digital wave. With years of experience in WordPress, technical SEO, and web performance, he writes about what actually works in the real world.