For two years Google folded AI Overview impressions into your Web search totals without telling you which was which. That quietly ended on June 3, 2026, when the Generative AI performance report landed in Search Console. It's the first official window into GSC AI search data, and most of the commentary about it is wrong about what the window actually shows. This field guide covers every field in the report, the counting rules that make the numbers smaller than reality, and the gaps you still have to fill somewhere else.
Where the AI Search Data Lives in Search Console
The report is called the Generative AI performance report (Search), currently in Beta. You'll find it under Performance reports, or you can jump straight there by appending /ai to your existing performance report URL. A parallel report exists for Discover.
Per Google's Search Central announcement, it covers AI Overviews, AI Mode, and AI features in Discover, with Search Labs experiments excluded. Four things to know before you open it:
- Data collection starts May 18, 2026. Nothing before that date exists, no matter how long AI Overviews have been citing you.
- The rollout began with UK properties, a sequencing shaped by the UK competition regulator, and is expanding since.
- The report only appears once your site clears a minimum impression threshold in AI features. Small or new properties may see nothing at all yet.
- The newest data shows as a dotted preliminary line that can update within hours.
Worth noting: Google also shipped an opt-out toggle under Settings on June 17. It defaults to including your content in AI features, and for anyone reading this article, that default is exactly where you want it.
The Fields: What Each Metric Actually Means
The metric list is short. One metric, in fact: impressions. Google defines an AI impression as the number of times links to your site were shown to a user in a generative AI feature on Google Search. You can slice it by page, country, device, and date, with the standard 1,000-row table limit and CSV export.
That's the entire report. No clicks. No CTR. No average position. No query or prompt text. And no API: the Search Analytics API still accepts only web, image, video, news, discover, and googleNews as search types, so none of this data reaches BigQuery, Looker Studio, dashboards, or any automated pipeline. If your dashboard vendor claims to pull GSC AI data programmatically today, ask them how.
How big can the numbers get? Practitioner analysis of one large early-access property showed 1.79 million AI impressions in 18 days, roughly 10% of its total Search impressions. AI surfaces are already a meaningful slice of visibility for established sites.
15+ content types. Published on your domain. Matched to your brand.
Guides, comparisons, listicles, case studies, and more. RankControl generates content that gets cited by ChatGPT, Claude, Perplexity, and more.
Start free→
The Counting Rules That Shrink Your Numbers
Backing up a step, because this is where most readings of the report go wrong. Two rules control what actually gets counted, and both push the numbers down.
First, the blended-count rule. AI Overview and AI Mode impressions were already inside your Web totals in the standard Performance report. The new report reclassifies an existing slice; it adds nothing. If the same URL appears in an AI Overview and as a blue link for one query, that's one impression, not two. So don't expect your aggregate numbers to move, and don't read the new report as newly discovered traffic.
Second, the visible-link rule. An impression fires only when an actual link to your page is shown. An unlinked brand mention counts nothing, and neither does a favicon appearance. A link sitting collapsed behind a show-more control registers only after a user expands it. Your brand exposure inside AI answers is structurally higher than your impression count, sometimes dramatically so.
Put those together and you get the report's core paradox: it undercounts your AI visibility while simultaneously revealing that AI visibility was hiding inside numbers you thought you understood.
How to Read GSC AI Search Data Without Fooling Yourself
Treat the report as directional, and build a small routine around it:
- Set a baseline this week. Export the report, segment pages by brand versus non-brand and by content type, and file it. The report's value compounds with time; the first export is worth little on its own.
- Overlay clicks from the main report. The pattern to watch is AI impressions rising while Web clicks stay flat or sink. That's AI Overviews absorbing zero-click demand, the dynamic we broke down in the zero-click crisis. GSC will never flag it as a problem, because every health indicator stays green while it happens.
- Layer conversion data before panicking. Several practitioners report the same counterintuitive pattern: informational traffic falls after AI Overviews expand, but the visitors who still click convert noticeably better. The AI answer filters out shallow intent before it reaches you.
- Re-pull on a schedule. Industry tracking puts citation rotation in AI Overviews at roughly 70% over a two-to-three-month window. Cited today means nothing about next quarter. A one-time audit is a snapshot of a moving target.
Is any of that a complete measurement system? No. It's the Google-shaped corner of one.
Your competitors are getting cited by AI. You're not.
Every day without citation tracking is a day your competitors pull ahead in ChatGPT, Perplexity, and Claude.
See what you're missing→Four Ways Teams Misread the Report
The report launched five weeks ago and the misreadings are already settling into conventional wisdom. Four worth killing early:
"Search Console is green, so traffic is healthy." GSC health indicators measure crawling and indexing, and they stay green while AI Overviews strip clicks from a technically perfect site. The deeper issue is that GSC samples performance data for privacy and simplifies it for the UI. Practitioners who cross-check against server logs describe GSC-only measurement as flying half blind, and that was before an impressions-only AI report got added to the mix. Directional tool, never a ledger.
"The report goes back to May 18, so that's my history." May 18 is when collection started, and that distinction matters. A site cited in AI Overviews since 2024 has zero GSC record of any of it. Whatever trend you see in the report describes weeks, and drawing year-over-year conclusions from it is fiction.
"It hasn't appeared for my site yet, but any day now." Maybe not. Early rollout observations show the report appearing on older, established properties while newer sites with real AI citations see nothing, because they haven't cleared the impression threshold. Absence of the report doesn't mean absence of AI visibility. Your brand can be cited daily while GSC shows you nothing.
"Position must work like it does in web search." The report exposes no position metric at all, and the underlying mechanics wouldn't map anyway. Search industry analysis of AI Mode found each component of a response carries its own position per element type, so a single answer contains multiple positions simultaneously. Any third-party tool selling you a clean "AI rank" number is approximating something Google itself doesn't report.
What This Report Will Never Tell You
Here's the thing: the Generative AI performance report is a Google product reporting on Google surfaces. The rest of the AI search world is invisible to it.
ChatGPT citations of your brand? Not in Search Console. Perplexity, Claude, standalone Gemini? Nothing. The prompts buyers typed that triggered your citations? Not exposed, and the report doesn't even split AI Mode from AI Overviews, so you can't tell which Google surface is doing the work. Meanwhile clicks from external AI engines land in GA4 mislabeled as direct or referral traffic, which leaves your attribution scattered across tools that don't agree with each other.
Filling the query gap takes real work. You have to mine your conversational GSC queries and test them across engines yourself, the process we documented step by step in how to find the exact prompts your buyers ask ChatGPT. Budget honestly for the routine: the baseline export and segmentation take about an hour, the cross-engine prompt testing runs 3 to 4 hours per cycle, and re-running everything every two weeks adds 2 to 3 hours each time. Call it 5 or 6 hours a month, indefinitely, per brand.
Or let the machine do the recurring part. RankControl runs citation tracking across ChatGPT, Perplexity, Claude, Gemini, and Google's AI surfaces continuously and unifies it with your Google data, so the GSC report becomes one input in a full picture instead of the whole picture. Across our customer base, Google's AI surfaces are consistently just one slice of total AI citations, and rarely the fastest-growing one. Plans start at $499/mo.
200+ SaaS teams already track their AI citations.
They know exactly when ChatGPT mentions their brand, and when it stops. Do you?
See plans→
The Field Guide, Condensed
Save this checklist:
- Find the report under Performance, or add
/aito your performance URL. Beta, threshold-gated, data from May 18, 2026 onward. - One metric: impressions. Four dimensions: page, country, device, date. CSV only, no API.
- Impressions require a visible link. Your real exposure is higher than the number.
- The data was already in your Web totals. Reclassified, not new.
- Watch impressions against clicks for zero-click erosion, and layer conversions before judging.
- Re-audit on a cadence; AI citations churn fast enough to invalidate any single snapshot.
- For everything outside Google, pair the report with cross-LLM citation tracking.
Google took two years to show you this much. The engines it doesn't own will never send you a report at all, and the teams who figure that out first are the ones who'll be cited everywhere their buyers actually ask.
