Search Console Generative AI Report: What It Leaves Out

Google's Performance → Generative AI report tells you one thing precisely: how many times links to your pages were shown inside AI Overviews and AI Mode. That is it. There are no clicks, no CTR, no position, no queries, and no way to split the two surfaces — Google's own help documentation lists impressions as the only metric and Pages, Countries, Dates and Devices as the only dimensions. So when leadership asks "are we winning in AI search?", the honest answer is that this report can confirm you were shown, and cannot confirm you were chosen.

That gap matters more than it sounds — as the first published numbers from a live property show.

The first real numbers, and what they couldn't say

Brodie Clark got early access and documented a large UK site with heavy branded traffic: 1.79 million AI impressions in roughly 18 days, about 120,000 a day, 813,000 in the trailing seven days — around 10% of the 8.4 million total Search impressions in that window. His verdict on the data was that it is "nowhere near actionable enough to be used in a practical sense," precisely because query and click data are missing.

Put that in a board deck and you have a number that grows, a chart that trends, and no way to answer the first follow-up question anyone will ask: did any of that turn into a visit? A more ordinary property makes the point again. When aioseo.fr ran the same analysis on its own site in July 2026, it found 4,820 AI impressions against 66,200 Web impressions — a 7.3% AI impression share — with only 40 of its 126 classically visible URLs appearing in the AI report at all. Its top 10 URLs held 94.7% of AI impressions versus 50.8% of classic impressions. AI visibility, on that property, was roughly three times more concentrated than organic visibility.

That concentration finding is genuinely actionable. The rest needs careful handling — so here is what is in the box, what the numbers mean, why the missing metrics have a scheduled arrival date, and what to say on Monday.

What's actually in the report

The inventory is short, which is a feature of a v1, not a scandal:

• One metric: impressions. Defined as how many times links to your site were shown to a user in a generative AI feature on Google Search.
• Four dimensions: Pages, Countries, Dates, Devices. No Queries dimension exists.
• Two surfaces, merged: AI Overviews and AI Mode. There is no toggle to separate them. If your AI impressions jump, you cannot tell which surface moved.
• Search Labs experiments are excluded, as are non-Web search types. Discover has its own separate report, which counts one impression per result per session and has no Devices dimension at all.
• Standard Search Console limits apply: the 1,000-row table cap, preliminary data that can shift within hours, and values shown as ~ or - landing in your CSV as zeros.
• Export is a button. There is no documented API or BigQuery bulk export for this report, so recurring client reporting means a recurring manual download.

Two practical notes before you start reading charts. The report lives under Performance → Generative AI, beneath "Search results"; if it hasn't appeared in your navigation yet, appending /ai to your Performance report URL reaches it directly. And history starts on 18 May 2026 — nothing exists before that date, which means your first genuine year-over-year comparison is not possible until May 2027. Every "growth" number you produce this year is growth from a standing start.

If your report shows zero or near-zero, resist the conclusion that you have no AI visibility. Check Settings → AI controls → Search generative AI first, because the exclusion setting can be inherited from a parent property. Google is also still rolling the report out to a subset of properties, and a site with few impressions may simply not qualify yet. Rule out the plumbing before you diagnose the content.

Read the impression definition before you read the chart

An impression in this report is not the same shape as an impression you're used to, and the differences change what the number means.

AI Overviews links must be scrolled or expanded into view to count. Google's impressions documentation applies standard impression rules to AI Overviews, which means a citation buried behind a "show more" expansion only registers once a user has already leaned in. That cuts both ways. Your impression count understates how often you appeared in the raw HTML, and any homemade CTR you build from these impressions will look flatteringly high, because the denominator quietly excludes everyone who never expanded the answer.

All links in an AI Overview share one position. Google assigns the AI Overview a single position and gives every link inside it that same value. This is why average position in your main Performance report can behave strangely on AI-heavy queries — you are averaging a genuine rank against a structural constant.

Property-level and page-level totals will not reconcile. As Search Engine Journal explains, two of your URLs appearing in one AI response can count as a single impression at property level while each URL gets its own at page level. Summing the page table and comparing it to the property chart produces a discrepancy that looks like a bug and is documented behaviour.

These are not new impressions. This is the point most easily missed in a stakeholder conversation. Generative AI impression data was already inside your overall Performance report; the June release added a dedicated view of a subset you were already counting, as ppc.land noted in its coverage of the announcement. Nobody's traffic went up because the report shipped.

That makes the one derived metric worth computing a ratio rather than a total. AI impression share = generative AI impressions ÷ Web impressions × 100, over a strictly identical window. Three ways teams get it wrong, per the aioseo.fr method: comparing a shorter AI history against a longer Web history, which mechanically deflates the ratio; leaving the Performance report unfiltered so Discover and News inflate the denominator; and ignoring canonicalisation, since pages are grouped by Google's selected canonical after redirects, not by the URL you happened to look up.

Get the ratio right and you have a trend line. What you still don't have is causation — and that is where the missing metrics start to bite.

Why clicks are missing — and why they have a due date

Here is the part most launch coverage skipped, and it reframes the whole "why won't Google give us clicks?" conversation.

The report did not appear because Google was feeling generous about transparency. It appeared on 3 June 2026 — the same day the UK's Competition and Markets Authority imposed its first binding Publisher Conduct Requirement on Google under the Digital Markets, Competition and Consumers Act 2024. CMA Chief Executive Sarah Cardell described it as a "world-first requirement on Google's search services in the UK". The AI opt-out toggle that now sits in Search Console Settings landed the same day.

And the order does not stop at impressions. Paragraph 7 of the conduct requirement obliges Google to give publishers clear, detailed metrics on user engagement with their content in search generative AI features: impressions, engagement clicks through to publisher sites, and click-through rate — disaggregated per publisher, separated from organic search results, and delivered through an accessible platform, with Search Console named as an example.

The dates are the useful part for planning:

Milestone	Date
Notice issued; report and opt-out ship	3 June 2026
Main control obligations in force	3 December 2026
First compliance report due	3 January 2027
Page-level grounding controls in force	3 March 2027
Designation period expires	10 October 2030

So the missing clicks are not a permanent design choice. They are a gap with a published deadline, and you can put that deadline in front of leadership instead of shrugging. The order also carries an anti-retaliation clause: Google "shall not seek to prevent publishers from, or penalise them for, using the controls… for example by downranking their Search Content."

Two caveats keep this honest. First, the CMA order binds Google's search services in the UK only. Building a click pipeline once and then geofencing it would be strange, but nothing published commits Google to a global rollout — treat it as likely rather than promised. Second, the order covers impressions, clicks, CTR and click quality. It does not appear to require query or prompt disclosure at all. Google's stated reasons for withholding prompt data are privacy and the difficulty of mapping a synthesised answer back to one query, since AI Mode and AI Overviews use query fan-out to run several sub-searches per prompt. Regulatory silence plus a technical argument makes the planning assumption clear: clicks are coming, prompts are not.

The three questions the report structurally cannot answer

1. Did any of it earn anything?

Clicks from AI Overviews and AI Mode do exist in Search Console — Google's documentation confirms a click on an external link inside either surface counts. They are simply blended into your Web totals with no way to isolate them.

You cannot rescue this in analytics either. GA4's AI Assistant default channel explicitly excludes Google's AI Overviews and AI Mode — it is defined as the channel for arrivals from sources like ChatGPT, Gemini, Deepseek, Copilot and Grok, while Google's own AI surfaces are classified as Organic Search. The technical reason is unglamorous: both surfaces send the same google / organic referrer as a blue link, because the referer header is google.com/search regardless of which layout rendered the result. No tool can separate them from that signal alone.

Net position for the growth lead: Search Console gives you Google AI impressions with no clicks; GA4 gives you AI clicks only for the assistants that aren't Google's. There is no supported join between them — a documented boundary between two Google products, not something you configured wrong. Say that out loud before someone asks you to "just connect them." For the analytics side of the problem, our guide to connecting an AI visibility score to GA4 signups covers the attribution layers involved.

And the external evidence suggests the missing clicks are not a rounding error. Pew Research Center tracked 900 US adults across 68,879 Google searches in March 2025, 12,593 of which produced an AI summary. Users clicked a traditional result on 8% of visits with a summary versus 15% without, and clicked a link inside the summary on just 1% of visits. Sessions ended outright on 26% of pages with a summary versus 16% without. Ahrefs, analysing 300,000 keywords with aggregated Search Console data, measured a 58% CTR reduction for the top-ranking result on AI Overview keywords between December 2023 and December 2025, with the effect shrinking down the page to −19.4% at position 10.

In fairness, Google disputes the pessimistic reading. Liz Reid, who runs Search, has argued that total organic click volume to websites is "relatively stable year-over-year" and that click quality has risen — though Google published no data with that claim. Both sides are arguing about a number neither can produce from first-party reporting. That is exactly what the December 2026 obligation is meant to fix.

2. What did people actually ask?

There is no Queries dimension, and query fan-out means a single user prompt can generate impressions against several of your URLs from sub-searches you will never see. Worth knowing: AI Mode queries do appear in Search Console's general Performance report — as reported when John Mueller confirmed it — they are folded into your normal query data rather than isolated in the generative report. Since AI Mode queries average around three times longer than traditional searches, unusually long query strings in your standard report are a rough proxy, not a substitute.

3. Were we recommended, or just footnoted?

This is the question the report was never built to answer, and the one your CEO is actually asking. An impression logs identically whether your page is the recommended answer or the fourth footnote beneath a competitor's recommendation. There is no sentiment, no prominence, no share of voice — Search Console is first-party only, so it will never tell you what fraction of the answer set you own.

More fundamentally, the report counts links. Semrush's "ghost citations" study with Kevin Indig, published 9 June 2026 across 3,981 domain appearances from 115 prompts in 14 countries on ChatGPT, Gemini, Google AI Overviews and Google AI Mode, found that only 13.2% of appearances were both cited and mentioned. 61.7% were ghost citations — linked as a source while the brand went unnamed — and 25.1% were brand mentions carrying no citation at all. That last quarter is completely invisible to any link-counting system: your brand is named in the answer a buyer reads, and Search Console records zero impressions.

The per-platform split shows why one report can't stand in for AI visibility. Gemini named brands in 83.7% of appearances but cited them in only 21.4%; ChatGPT was almost the mirror image at 20.7% mentions and 87% citations. If your measurement counts links, you will systematically under-read Gemini and over-read ChatGPT.

That asymmetry is why Geoptimizer scores mention rate and citation rate as separate components — 35% and 25% of the visibility score respectively — rather than collapsing them into one "appeared in AI" number. Search Console can only ever measure the citation half, because links are all it can see.

And the fourth gap: the engines that aren't Google

Search Console is the definitive record of two very large surfaces and completely silent on the rest of the category. That silence covers a lot of ground. Similarweb's worldwide generative-AI web traffic data puts ChatGPT at roughly 53% of gen-AI web traffic as of May 2026, down from about 76% a year earlier, with Gemini up from under 9% to around 27–28% and Claude up from about 2% to nearly 9%. Google's own scale is real — Sundar Pichai told Alphabet's Q2 2026 earnings call that AI Mode has surpassed 1 billion monthly active users, with the Gemini app at 950 million — but ChatGPT's share of assistant traffic is not something a Google-first-party report will ever reflect.

The divergence runs deeper than traffic share. Semrush and Kevin Indig found that ChatGPT's minimal-reasoning and high-reasoning modes share only 25.6% of cited domains for identical prompts, with citation rate climbing from 50% to 68%. As Indig put it, the brand that wins under minimal reasoning is not the brand that wins under high reasoning: "These are two different systems." One engine, two configurations, two winner lists — none of it visible in a Google report.

Even within Google, the surfaces behave differently enough that merging them costs you information, which is why we argue for treating Gemini, AI Overviews and AI Mode as three separate tracked surfaces rather than one Google number. The generative AI report merges two of the three by design.

What to actually tell leadership on Monday

Four lines, and two things not to say.

Report these:

1. AI impression share as a trend, not an absolute. Week-over-week movement in generative AI impressions ÷ Web impressions, on matched windows with search type filtered to Web. Absolute impression counts invite the wrong follow-up question.
2. Distinct URLs appearing in AI. If 40 of your 126 visible URLs show up, that ratio is your AI coverage — and improving it is a content-shape project you can staff.
3. Top-3 URL concentration versus organic concentration. When 95% of AI impressions ride on ten pages, your exposure to a single ranking change is enormous and worth naming as a risk. Apply a floor — ignore URLs under roughly 500 Web impressions — before you read anything into a movement.
4. The gap, named and dated. "Click data is not in this report. Under the CMA's conduct requirement, Google's main obligations — including click-through data and CTR — come into force on 3 December 2026, for UK search services." A dated gap is a plan; an unexplained gap is a credibility problem.

Don't do these two things. Don't blend generative and traditional impressions into a single composite CTR — Search Engine Journal's guidance is explicit about it, and the scroll-or-expand rule means an AI-derived CTR isn't comparable to organic CTR anyway. And don't present impression growth as traffic growth. The cleanest phrasing of that trap comes from Pragma-Code's write-up: "A 200% surge in AI impressions does not imply a 200% increase in website traffic."

One more decision this report will tempt you into: opting out. The Search generative AI control under Settings → AI controls excludes the property from AI Overviews, AI Mode and generative features in Discover, with changes generally taking effect within one to two days. It does not affect conventional ranking, does not govern model training (still Google-Extended in robots.txt), and is not a removal from Search. But exclusion carries non-obvious costs — Top Stories carousels now nest inside AI Overviews on roughly one in six US trending news queries — so opting out can quietly forfeit placements you thought were conventional. You cannot make that trade rationally on impressions alone, which is a reason to revisit it after December.

The instrument panel, not the instrument

It would be easy to read all of this as a complaint about Search Console. It isn't. The report has one property no third-party tool can match: it is a census. It reports what actually happened to real users at population scale, with no sampling error and no prompt-selection bias. Every AI visibility platform, this one included, samples — it runs a chosen prompt set on a schedule and observes what comes back.

And sampling has a real cost, which is worth stating plainly. Parse analysed 693,509 answers from 16,143 ChatGPT prompts and 15,805 AI Overviews prompts between 26 March and 25 April 2026, and found that repeat runs of the same prompt shared only 21.2% of cited domains on ChatGPT and 31.5% on AI Overviews. Roughly four in five ChatGPT sources churn between identical runs. A single scan is an anecdote — which is why a defensible score reports a rolling window with a confidence band rather than a screenshot, and why distinguishing a real shift from measurement noise after a model release is its own discipline.

So: different instruments, different questions. Search Console answers "how often were we shown on Google's AI surfaces?" with census-grade authority and link-only vision. A prompt-level tracker answers "when buyers ask, do the engines name us, cite us, and how do we compare to the competition?" across engines Google will never report on. Neither replaces the other, and a team using only one will confidently misread its position.

FAQ

Does the Search Console generative AI report show clicks or CTR? No. Impressions are the only metric, and Pages, Countries, Dates and Devices are the only dimensions. Clicks from AI Overviews and AI Mode do count in your main Performance report, but they are blended into Web totals with no way to isolate them. Under the CMA's conduct requirement, Google's main obligations — covering clicks and CTR for publishers — come into force on 3 December 2026 for UK search services.

Can I separate AI Overviews from AI Mode in the report? Not currently. Both surfaces are merged into one impressions curve with no filter to split them. If you need surface-level detail, you have to infer it from other signals or track the surfaces independently outside Search Console.

Why does my generative AI report show zero impressions? Check three things in order. Is the property excluded under Settings → AI controls, possibly inherited from a parent property? Has the report rolled out to your property yet — Google is still releasing it to a subset of owners. And do you have enough AI impressions to register at all? Only after ruling those out is a content or crawler-access diagnosis worth running.

How do I calculate AI impression share correctly? Divide generative AI impressions by Web impressions over a strictly identical date window, with the Performance report filtered to search type = Web. Remember that generative data starts on 18 May 2026, so any window that reaches further back will understate the ratio.

Is Search Console enough to measure AI visibility on its own? It is enough to measure how often Google's AI surfaces showed links to your site. It cannot tell you what people asked, whether anyone clicked, whether a competitor was recommended above you, or whether ChatGPT, Claude or Grok mention you at all — and on the ghost-citations data, about a quarter of brand appearances in AI answers carry no link for it to count.

Where to go from here

Read the report for what it genuinely gives you: a census-grade trend line on Google's AI surfaces, a coverage ratio you can improve, and a concentration risk you should name. Then be explicit with leadership about the four things it cannot say — no clicks until December at the earliest, no prompts probably ever, no recommendation signal, and no engines beyond Google.

For the half Search Console can't see, Geoptimizer tracks ChatGPT, Gemini, Claude and Grok in every plan including the free tier, scoring mention rate and citation rate separately so you can tell "they linked us" apart from "they recommended us." Run a free check on your domain, put the two numbers side by side, and you'll have a far more complete answer than either instrument gives alone.