Google Search Console's new AI performance reports, explained

Since AI Overviews launched, site owners have lived with a strange asymmetry: Google reshaped half the results page with generated answers and reported none of it. AI Overview impressions were folded invisibly into regular Search performance data, AI Mode sessions left no trace at all, and the industry resorted to third-party SERP scraping to estimate what Google would not say. By March 2026, with Ahrefs measuring AI Overviews on 48 percent of queries and AI Mode past a billion monthly users, the missing data had become the biggest hole in every SEO report.

On June 3, 2026, Google closed part of the hole. The Search Central blog announced Search Generative AI performance reports in Search Console: dedicated views of the impressions your site earns inside generative AI features on Search, plus a parallel report for generative AI features in Discover. The announcement arrived the same day as a related control for opting content out of AI features, which we analyze separately in Google's AI Overviews opt-out toggle, and both trace back to regulatory pressure from the UK's competition authority.

It is a genuine milestone: the first official, Google-supplied measurement of the AI answer layer. It is also, read carefully, a narrower instrument than the headlines suggested. This article covers what the reports contain, what they deliberately omit, how to use the data well, and how to pair it with the answer-level tracking the reports cannot do.

The Search report counts impressions your pages earn inside generative AI features on Search, which Google defines to include AI Overviews and AI Mode together. Three breakdowns are available: by page, so you can see which URLs surface most in AI responses; by country, so you can see where the AI footprint concentrates; and by device. The headline chart trends total AI-feature impressions over time, which is the single most useful pixel in the release, because direction is the thing nobody could measure before.

The Discover report does the same for generative AI features inside Discover, a surface most brands ignore and news publishers live on. For both reports, an impression means a link to your page appeared within an AI response shown to a user. Google's help documentation frames the intended uses plainly: watch how AI-feature impressions change over time, find your highest and lowest performing pages in AI surfaces, and see where the impressions originate geographically.

Access is the final practical detail. The rollout began with a subset of site owners, starting in the UK, with global expansion promised but undated. If your property does not show the report yet, that is the rollout, and there is no setting to flip. The wait is worth planning around rather than fuming over, because the useful work, building the question-level layer underneath, does not depend on it.

No queries. The report will tell you a page earned 4,000 AI impressions and refuse to say for which questions. You cannot connect an impression to a buying moment, cannot tell an informational query from a comparison, and cannot build a work queue from the data. Query-level insight, the core of classic Search performance reporting, is absent from the AI report at launch.

No clicks or CTR. Impressions without clicks measure exposure without payoff. The omission stings because the independent click data is dramatic: Pew Research found users click a traditional result on 8 percent of searches with an AI summary versus 15 percent without, and click links inside summaries on about 1 percent of visits. Whether your own AI impressions convert to anything is exactly the question the report cannot answer.

No surface split. AI Overviews and AI Mode are merged into one bucket, though they are different products with different stakes, a distinction we map in AI Mode vs AI Overviews. A thousand impressions could be a thousand glancing Overview renders or a thousand deep AI Mode research sessions, and the report cannot tell you which.

No answer content.The deepest gap. An impression records that your link appeared; it records nothing about the sentence next to it. The answer may have recommended you, dismissed you, or cited your data while recommending a competitor. Brand mentions without links, common in AI answers, register nowhere. The report measures your pages' presence in the scaffolding of answers, never your brand's presence in the verdicts.

Set the baseline immediately.The first thing to do when the report reaches your property is nothing: record four to eight weeks of AI-feature impressions untouched, per page, so every later content change has a before. Trend data is the report's entire strength, and a trend needs a starting point you did not contaminate.

Read the page mix, and the ratio. Sort pages by AI impressions and study what surfaces: in most accounts the AI-heavy pages are the answer-shaped ones, definitions, comparisons, how-tos, while product pages lag. That mix tells you which of your formats the answer layer already trusts, which is the format to publish more of. Then compare each page's AI impressions to its classic Search impressions. Pages with high AI share are your answer-layer assets; protect and update them. Pages with none, despite ranking, are candidates for restructuring along the lines the Princeton GEO research validated: statistics, quotations, cited sources, the GEO playbook.

Watch for cliffs, and annotate.A sudden drop in AI impressions on a stable page usually means the answer moved: a competitor's page displaced yours in the citations, or the engine changed its sourcing pattern. The report will show you the cliff without the cause. Keep a change log of your own publishes and fixes against the impression timeline, because Google gives you no annotations, and causeless charts breed bad theories.

The report's gaps have a common shape: everything Google omitted is question-level or answer-level. Which questions trigger AI answers in your category, whether those answers name your brand, which competitors they name, and which sources they cite. None of that is in Search Console, and all of it is observable from the outside, by asking the engines your buyers' questions and reading what comes back.

Doing that observably well takes discipline, because generative answers are probabilistic. SparkToro measured under a 1 percent chance that two identical runs return the same brand list, so single checks mislead and repeated sampling is mandatory. The workflow that holds up: a frozen list of real buying questions, scheduled runs across engines, mention rates instead of binary presence, branded questions excluded from the headline score, and every raw answer stored as an auditable receipt. The full method is written up in how to track your brand in Google AI Mode.

Reachroller runs that workflow as a product and adds the step no report offers: the fix. When a tracked question comes back without your brand, it generates a publish-ready, answer-shaped page, slug, title tag, meta description and schema included, then rechecks whether the answer flipped. Starter is $29 per month for 400 credits and 25 tracked questions. The honest caveats: it is a young product, ChatGPT tracking is live today with Gemini, Claude, Perplexity and Grok rolling out, and it reads answers rather than Google's private impression logs. That is precisely why the pairing works: Search Console tells you the footprint, Reachroller tells you the verdicts, and together they turn the AI layer from a rumor into a managed channel. Where it sits among alternatives is covered in the best AI visibility tools.