AI Measurement10 min readJuly 17, 2026

Google Generative AI Performance Report: Where to Find It and What to Measure

Find the Search Console generative AI report, export AI Overview and AI Mode impressions, troubleshoot missing data, and build a comparable reporting baseline.

AI feature report / Search
PagesCountriesDevicesDates
Supported visibility, kept inside its evidence boundaryImpressions · pages · countries · devices

Google’s generative AI performance report is a Search Console report for impressions in AI Overviews and AI Mode. Google announced it on June 3, 2026 and completed its worldwide rollout on August 31. Instructions that still tell every missing-report user to wait for a limited rollout are out of date.

Start with the official report link in Google’s help page, select the property you manage, and check whether it has reportable data. Then save an export before interpreting the chart. This guide explains what to capture, how to compare it with ordinary Search performance, and what to investigate when the numbers fall. The working examples are hypothetical, not results from a client account.

Section 01

Find the report and check access

Use the official help page linked in Sources and its Open the Generative AI performance report action. Choose the correct Search Console property when prompted. Keep the Search report separate from the Discover version: they describe different surfaces. The supported Search features are AI Overviews and AI Mode; Search Labs experiments are excluded.

First export workflow

  1. Confirm the property identifier. A domain property and a URL-prefix property can represent different sets of addresses; do not switch between them during a comparison.
  2. Choose a completed period. Save the start and end dates rather than recording only a moving label such as last month.
  3. Record the page, country, and device filters before exporting. Keep a screenshot of the selected view alongside the file.
  4. Export the chart and selected table. Save the date trend and page breakdown as distinct datasets rather than pretending they form a page-by-day table.
  5. Record the export time, reporting time zone, visible row count, and whether the interface shows preliminary or unavailable values.

If the report is missing, check account access, property selection, and Google’s stated impression threshold and eligibility conditions. Google’s dated August 31 announcement takes precedence over older rollout language still present lower on the help page. Do not buy an enabling plugin or add speculative markup to make a reporting menu appear.

Section 02

Read impressions without double-counting

Google applies its standard impression rules to AI Overviews and AI Mode. An AI Overview occupies one result position, shared by its links. A follow-up question in AI Mode is a new query for reporting. Those definitions describe search-result exposure; they do not measure how persuasive an answer was or how much of your page it used.

Choose the view that answers the question
ViewUse it forKeep separate
Property chartThe overall impression trend for the chosen propertyThe sum of individual page rows
PagesThe canonical pages receiving reported exposureUnverified mappings to duplicate URLs
Countries and devicesBroad changes in the reported audience mixPrecise location or individual identity
DatesComparable periods in Pacific TimeA different analytics property’s daily boundary

Consider a hypothetical result with two links to one company’s site. A property-level view can count one impression while a page-level table credits both pages. Neither number needs to be faulty. They answer different questions. Label the aggregation beside the figure before asking a developer to fix a discrepancy.

For an internal spreadsheet, keep separate columns for property impressions, page impressions, and the relevant export identifier. Never overwrite the property figure with the sum of downloaded rows. If a page filter changes the aggregation, save that as a new view. This small bookkeeping step prevents a filtered report from quietly replacing the baseline in a monthly deck.

The dedicated report is a subset of Web Search performance. A dashboard that adds AI impressions to the Web total counts overlapping exposure twice. Display the two series beside one another, with their names and definitions intact. Do not relabel the broader series traditional search unless the data actually supports that separation.

Section 03

Know which questions need another dataset

The dedicated view does not expose a query dimension or dedicated click and CTR fields. External link clicks from AI features count in the broader Search methodology, but that does not create a compatible AI-only numerator. Dividing Web Search clicks by AI impressions would produce a number without the meaning implied by its label.

A practical measurement map
QuestionEvidence to inspectDecision it supports
Did a page receive AI-feature exposure?Dedicated impression reportChoose a page for review
Which search terms reach the site?Available ordinary Search query rowsReview existing search intent, not hidden AI prompts
Did visitors arrive and use the page?Site analytics and landing-page eventsInspect experience and instrumentation
Did an inquiry become suitable work?Accepted lead records and qualification statusEvaluate commercial usefulness

A service owner may ask whether a rise in AI impressions caused new bookings. Start by defining the booking event and the source evidence available for it. A page-level coincidence is a reason to investigate, not a person-level attribution join. Keep accepted forms distinct from button clicks and scheduled consultations distinct from unqualified inquiries.

For citation quality, retain a separate, reproducible observation: the question asked, date, product, response, linked page, and the specific statement the page supports. Repeat important tests rather than presenting one favorable answer as a share-of-voice score. These observations help editorial review, but they are not a replacement for first-party traffic reporting.

Section 04

Build a comparison you can reproduce

Keep one row per dataset and export, with a stable identifier. Record the property, surface, start date, end date, filters, aggregation, export time, and completeness notes. Store the original download unchanged. Transform a copy so another analyst can trace a reported number back to the original file rather than reconstructing last month’s interface settings.

Before comparing two periods

  • Both periods cover the same number of completed days and comparable weekdays.
  • The property, search surface, countries, devices, and page filters match.
  • The page set is defined before looking for winners; new pages have a separate cohort.
  • The export’s row limit and unavailable-value handling are documented.
  • Any canonical change is recorded explicitly instead of silently rewriting historical URLs.
  • The comparison distinguishes observed change from the suspected explanation.

Suppose a hypothetical property records 1,000 impressions in one period and 800 in the next. Two downloaded page tables do not automatically explain the 200-impression loss. Their aggregation may differ from the property chart, and the download may omit rows. Calculate changes for matched pages, report one-sided rows as unresolved, and keep the property change as a separate result.

Search Console exports are limited, and unavailable display values can become zeros in the download. Preserve the interface’s meaning before making arithmetic assumptions. For ordinary Web Search API exports, pagination reduces avoidable truncation but does not remove Google’s underlying top-row limitations. Use the supported API contract; do not invent a generative-AI search type.

Keep a comparison ledger with the question, evidence files, finding, proposed action, and review date. This is especially useful when several people edit a site. A second analyst should be able to see that a title change was prompted by a particular query mismatch rather than by an unexplained red arrow in a dashboard.

Section 05

Investigate missing data before changing pages

Check Google’s data-anomalies page against the exact dates and report involved. Google recorded a logging error affecting generative AI Search impressions for August 13–17, 2026, and states that the missing data was restored on August 21. An older download may therefore differ from a new export of that period. Retain both versions and explain the revision.

That resolved August incident is not evidence that a September traffic decline is a logging error. Match the product, affected dates, and repair status before using an incident to explain a chart. Separately check your own changes: tracking releases, access settings, redirects, canonical selection, and publication dates. A release calendar is more useful than a list of unrelated industry headlines.

Triage a drop

  1. Re-export a completed period and compare it with the saved version. Determine whether the source data changed before attributing a change to the website.
  2. Identify whether the loss appears in AI impressions, broader Search clicks, recorded sessions, or several measures. Preserve those distinctions.
  3. Inspect the pages with supported losses and check their live content, response status, indexing directives, and recorded Google crawl dates.
  4. Compare stable page and query cohorts where data exists. Keep disappearance from a limited export separate from a measured zero.
  5. Fix a reproduced defect first, then retain the released version and a later measurement window. Avoid simultaneous rewrites that erase the baseline.

A useful incident note can be brief: what changed, when it changed, which dataset shows it, what test was run, and what remains unexplained. Assign the next check to a named owner. Do not keep changing a working page simply because Google has not yet crawled the last release.

Section 06

Turn the evidence into a page decision

Use exposure to choose where to look, then read the page as a buyer would. Does its first screen answer the question? Can a visitor find the relevant service, a concrete example, and a sensible next action? Improving an existing useful page is a different project from creating another page that repeats the same introduction with a different keyword.

When ordinary Search data shows meaningful exposure near the first page but few clicks, inspect the actual query mix before editing a title. A broad average can combine branded searches, informational questions, and distant service queries. Choose wording that accurately describes the answer. Save the old title and the release date, then compare equivalent cohorts after sufficient observations accumulate.

Do not read a better sitewide average position as proof that no rankings fell. Hypothetically, a page can retain ten impressions at position five while losing ninety impressions at position fifty. Its reported average improves as exposure disappears. A fixed query-page comparison is needed to distinguish a ranking change from a shift in which searches remain visible.

For content investment, choose a small set of pages with distinct tasks and credible commercial relevance. Give each one a useful example, a clear answer, and relevant internal links. Preserve successful URLs. Consolidate only when the pages really duplicate the same task and the destination serves their users; a lower indexed-page count is not itself a business result.

Section 07

Report the result and the next decision

A useful monthly report starts with a decision, not a composite score. State whether to maintain, investigate, or improve the named page group. Support that decision with the observed trend, the comparison settings, and the unresolved limits. Keep implementation completed, Google recrawling, search exposure, visits, and qualified work as separate milestones.

A reusable reporting structure
FieldWhat to write
ScopeProperty, Search AI report, fixed dates, filters, and aggregation
ObservationThe measured change, with absolute counts beside percentages
CoverageUnavailable rows, preliminary values, limits, and known data revisions
ActionThe exact page and the defect or reader task being addressed
Follow-upRelease date, unchanged control, and completed comparison window

For a small site, show four clicks versus six before presenting a percentage. For a larger site, separate stable pages from new launches and report the main losing groups instead of averaging them away. In either case, the report should make the next investigation narrower and more testable, not provide an excuse for unlimited publishing.

Primary and authoritative references

Source ledger

These sources support the operating guidance above. Platform behavior and documentation can change, so volatile implementation details should be rechecked before a rollout.

  1. Generative AI reports: launch and August 31 rollout update — Google Search Central. Launch timing, worldwide availability, and the relationship to broader reports.
  2. Generative AI performance report for Search — Google Search Console Help. Report access link, fields, aggregation, export limits, and unavailable values.
  3. Impressions, position, and clicks — Google Search Console Help. Search metric definitions and AI-result counting rules.
  4. Data anomalies in Search Console — Google Search Console Help. August 13–17 AI Search logging issue and August 21 restoration.
  5. Using Search Console and Google Analytics data for SEO — Google Search Central. Use search and on-site measurement as complementary evidence.
  6. Search Analytics query API — Google for Developers. Supported ordinary Search API requests, final data, pagination, and top-row limits.
Implementation

The practical next step

Save one completed-period export and its comparison settings. Choose one existing page whose answer can be improved, record the specific defect, and keep its URL stable. Review the released page before checking later search results. Use a separate lead-measurement record for inquiries rather than treating an AI impression as a new customer.