Google's Generative AI Performance Report: Compare the New Search Console Data Without Inventing Outcomes

On June 3, 2026, Google announced dedicated generative AI performance reports for Search and Discover. The Search report covers supported generative AI capabilities in Google Search, currently AI Overviews and AI Mode; Search Labs experiments are excluded. Google documents three reasons a property may not show the report: access has not reached the property during the limited rollout, the property lacks enough AI-feature impressions for the report, or the property is excluded from supported generative AI features. A missing report therefore does not prove that a site has never appeared in an AI feature.

The dedicated Search report is impression-focused. Its chart shows total impressions over time. Its table can group those impressions by page, country, device, or date. The newest values can be preliminary, and the normal Search Console limitations—including row limits and aggregation differences—also apply. The report can be exported, but unavailable values represented in the interface may become zeros in the download, so an analyst must not silently treat every exported zero as a directly observed zero.

Search Console already had rules for clicks, impressions, and position in AI Overviews and AI Mode. The dedicated report does not replace those rules; it creates a filtered view of AI-feature impressions. In an AI Overview, a link generally must be scrolled or expanded into view before its impression is counted. All links in the AI Overview share the position assigned to that result element. In AI Mode, standard impression rules apply, and a follow-up question is treated as a new query for reporting purposes.

Property-level and page-level totals can differ without either being wrong. If two URLs from the same property appear in one AI result, the property chart can count one impression while the page table can show an impression for each page. This is why summing table rows and comparing the result directly with the chart can produce a discrepancy. The report documentation explicitly tells users to expect differences caused by aggregation.

A page filter also changes the chart to URL aggregation, so the filtered chart is not the same population as the unfiltered property chart. The interface and export can expose no more than 1,000 table rows. Preserve the aggregation mode, filter, row count, and truncation state with every export; never present a page-row sum as an exhaustive property total.

Canonicalization creates another important boundary. Most performance data is credited to the canonical URL Google selects, not necessarily the URL a team expected to measure. A campaign parameter, alternate device URL, or near-duplicate page may therefore disappear into a canonical page's total. Before diagnosing an AI visibility change, inspect Google's selected canonical and the page grouping rather than assuming the export preserves every served URL.

The dedicated report does not currently provide a query dimension. A page may accumulate impressions from many different needs: broad research, a comparison, a local service question, or a branded lookup. Without query data in this view, the page topic can suggest possible intent, but it cannot prove the prompts that caused the impressions. Grounding a content decision in an imagined prompt would turn an observed page impression into an unsupported story.

The report also does not display dedicated click or click-through-rate fields. Google states that external link clicks from AI Mode and AI Overviews count as clicks in the broader Search performance methodology, but the new generative AI report is designed around impressions. Analysts must resist manufacturing an AI click-through rate by dividing a separately filtered or differently aggregated click total by these impressions unless Google provides compatible dimensions and methodology.

Nothing in the report identifies the wording of an answer, the claim associated with a link, the order of sources, or the degree to which a page influenced the generated response. It is possible for the same page to appear as a supporting link in different contexts. An answer-level citation audit requires captured responses, prompts, dates, locations, model surfaces, and repeated observations outside Search Console.

The dedicated report exposes AI-feature impressions, while the standard Search performance report includes clicks and impressions from Search features under its broader methodology. Google does not currently provide a compatible dedicated-click field or query dimension in the generative AI view. That means the two exports can be placed beside one another for context, but they cannot be subtracted or divided to manufacture an AI click total or click-through rate.

The first export should be treated as a baseline, not a verdict. Google added AI Mode to standard Search performance totals on June 16, 2025, while the dedicated generative AI report launched to a subset of properties in June 2026. Treat the former as a standard-report methodology milestone and the latter as a product-access milestone unless Google documents a data-definition change. Record both beside site changes.

Choose comparisons that reduce avoidable noise. Compare the same logical pages, countries, and devices over equivalent periods. Use weekly or monthly granularity when daily impressions are sparse. Keep newly published pages separate from established pages. If a canonical changed, preserve both the old and new URL mapping. Record report access separately from observed data availability. Start the baseline at the first period whose visible rows, filters, and completeness support a like-for-like comparison; do not infer inaccessible history or fill it with assumed zeros.

Page-level impressions can prioritize investigation. If a detailed service guide gains visibility while the corresponding commercial page does not, review whether the guide answers an exploratory need and whether its internal path to the service is useful. If impressions concentrate on obsolete pages, fix content freshness and canonical signals. A rise in mobile AI-feature impressions and a separate decline in mobile conversion-event health should be investigated as two aggregate observations; the report cannot show that the same impression produced the affected session or conversion.

Country and device dimensions can also expose operational questions. Unexpected country exposure may indicate broad informational relevance, not a new market. A device shift may change page experience requirements without changing the underlying topic. These dimensions help teams decide where to inspect, but they do not reveal identity, precise location, or intent.

The best content action remains evidence-led improvement: correct stale facts, make important information available in text, provide first-hand detail, maintain clear internal links, and ensure structured data matches visible content. Google explicitly warns against unsupported AEO or GEO hacks. The report is a feedback surface for useful pages, not permission to create hundreds of thin fan-out pages.

A concise summary should stay inside the report's contract: which canonical pages received supported AI-feature impressions, how the page, country, or device mix changed, whether the property had complete access for the period, and which page investigation follows. Referral and outcome reporting belongs in a linked measurement section with its own definitions, not inside the AI-impression total.