What Search Console's New AI Report Shows and Hides

Google gave every website owner a new report on 31 August 2026 and, as far as I can tell, hardly anyone has noticed. It sits in Search Console under Performance, and what it shows is how often Google’s own AI has been putting your pages in front of people. Yours has one. Odds are you have not looked.

I did, last week, on a small UK site. Under 3,000 Google impressions a quarter, the kind of property nobody writes case studies about. One in six of those impressions turned out to be inside an AI Overview or an AI Mode answer. And the page doing 84% of that work? A how-to piece from 2025 that nobody had touched since.

None of it was planned. That is rather the point. The report tells you where Google’s AI already leans on your pages, long before you have done a thing about it, which makes it worth ten minutes even if the rest of Search Console bores you rigid.

What follows is roughly the order I wish someone had explained it to me in: where the thing is, what it counts and what it quietly leaves out (clicks, for a start), and then what one small site’s numbers actually say.

Where the report lives, and how it got here

You find it by opening Search Console, clicking Performance, and looking one line down. Generative AI, with a Beta badge still stuck to it. If you would rather go straight there, the address is search.google.com/search-console/performance/search-analytics/ai. Same date presets as the report you already know, from 24 hours up to 3 months, and one purple card with one number on it. That number is the whole show.

It did not arrive quietly, even if it felt that way.

The part most write-ups skip: this report is a regulatory product as much as a marketing one. Google did not wake up wanting to show you how often its AI borrows your pages. The Competition and Markets Authority imposed publisher controls on Google as a conduct requirement under its strategic market status in the UK, and the reporting came bundled with the switch. The main obligations do not even bite until 3 December 2026, so Google shipped the toggle early. Knowing that tells you something about where it is going: more control, eventually, and probably click data, because the regulator is expecting it.

What does the generative AI report actually show?

One metric. Impressions.

Google’s help page defines them plainly: “Impressions are how many times links to your site were shown to a user in a generative AI feature on Google Search.” You can slice those impressions four ways: by page, by country, by device, and by date. That is the lot.

There is no clicks column. No CTR. No average position. And no queries, so you cannot see which question triggered the AI answer that included you. For anyone used to the normal Performance report, that last gap stings the most. (A twin report covers generative AI features in Discover. Same shape, same gaps, and for most business sites, a row of zeros.)

One more thing the help page says that people miss: the report “includes data from the Web search type in the Performance report”. In plain English, the AI impressions are already inside your normal Web impression total. They are a slice of it, not an extra on top. Add the two together and you have counted the AI ones twice.

Is the report useless without clicks, then? No. But it is a visibility gauge, not a traffic gauge, and it needs reading next to the classic report rather than instead of it. More on how to do that further down.

How does Google count one AI impression?

This is where the numbers get wonky if you do not know the rules, so here they are.

The third rule matters more than it looks. AI Mode answers routinely fold their sources into a side panel. If nobody opens the panel, your page was used and you were never told. The report shows you the minimum, not the whole story.

What a small site’s numbers look like

Theory is fine. Here is a real property, anonymised: a small UK site, three-month preset, 13 June to 12 September 2026, read directly from Search Console.

Remember the slice rule: those 429 are inside the 2,610. So about 16% of every time Google showed this site to anyone in that quarter, it showed it inside an AI answer. For a site nobody has tuned for AI search, that is a bigger number than I expected.

Now the pages.

Three things jump out.

The AI visibility is lopsided. One page holds 362 of the 429 impressions, 84% of the total. It is a practical how-to post for a specific local trade, the kind of page that answers a question in full. A third of that page’s own impressions now come from AI features. The glossary post, which ranks for far more classic queries, gets a 6% AI share. Google’s AI is not spreading its citations evenly across a site. It picks one or two pages it trusts and leans on them.

Desktop wins, by a mile. Of the 429 AI impressions, 299 were on desktop and 127 on mobile, with 3 on tablet. That is 70% desktop. Every other stat about search says mobile first. AI Mode usage on this evidence skews to people sat at a computer, which for a business-to-business site means people at work, mid-research. Worth knowing before you assume your AI audience is thumbing through on a phone.

The bar for “enough data” is low. This is a site with 22 clicks in a quarter. It still has a fully populated report with 62 countries in the country tab (the UK accounts for 262 of the 429, the US 38, then a long tail). If you have been told your site is too small for AI visibility to matter, open the report anyway. You will probably find something.

And one honest caveat: none of this tells me whether any of those 362 AI impressions turned into a visit. The classic report shows 13 clicks for that page across the quarter. Whether those came from a blue link or from inside an AI box, Search Console will not say.

Should you switch your site out of AI features?

Almost certainly not. And I say that knowing plenty of publishers disagree.

Here is the switch. In Search Console go to Settings, find the section called Search generative AI, and you will see it set to Include. Change it to Exclude and Google says your site “will not receive traffic or impressions from our generative AI features”. Google also says the setting is not used as a ranking signal outside those AI features, which is the CMA’s doing.

For a business site, the trade is a bad one. You give up being cited in the box that now sits above the results on close to half of the commercial queries BrightEdge tracks (48% by February 2026), and in return you get… the same ten blue links you had before, with fewer people reading them. Ahrefs measured a 34.5% fall in click-through rate for the top organic result when an AI Overview is present, across 300,000 keywords, back in 2025 when the boxes were rarer than they are now. Opting out does not remove the AI Overview. It removes you from it. The box stays, your competitor’s link goes in it instead.

Where the switch does earn its keep is a publisher whose entire business is the click: paywalled news, ad-funded content sites. If that is you, you are having a different conversation. For the plumber, the accountant, the agency, the training company, being named in the answer is the point.

Two things people confuse with this switch, so you do not have to:

• The nosnippet tag removes a page from AI Overviews, but it also strips the normal snippet from classic results. Blunt tool.
• The Google-Extended robots.txt token controls whether your content trains Gemini models or grounds answers in the Gemini app and Vertex AI. It has no effect on AI Overviews or AI Mode. Blocking it does not take you out of the answers.

The mildly awkward truth is that Google gave everyone a switch most people should never touch, because a regulator required it. Fine. Leave it on Include and use the report instead.

Five things to do with the report this month

The report will not change your rankings. It will change what you decide to work on, if you read it properly.

1. Write down your baseline. Open the 3-month view, note the total AI impressions and the Web total from the classic report, work out the share. Do it today, because the report is in Beta and the counting rules have already been clarified once. Next quarter you want a number to compare against.

2. Find your one page. Sort the Pages tab by impressions. On most small sites one page will dominate, the way the how-to post did above. Read it as Google’s AI reads it: is there a clear answer near the top, does each claim carry a source, is the page dated? The eight checks in the AI citation signals post apply here directly. Fix that page before you write anything new.

3. Check the device split. If yours is desktop-heavy like the site above, your AI readers are researching at work. That changes the call to action: a contact form and a phone number, not an app download.

4. Cross-read with the classic report. Put the same page in both reports for the same range. High AI impressions with falling clicks means Google is summarising you and keeping the reader. High AI impressions with steady clicks means you are being cited and people still come through. Same number, two different problems, and only the classic report tells you which.

5. Diary a monthly look, then leave it alone. Google has said more metrics are coming, and the CMA wants page-level controls by March 2027. Clicks may well follow. Check on the first of the month, note the totals, and get back to work. Staring at a report with one metric in it is a right faff for no gain.

The report is a mirror, not a lever

Nothing in the Search Console generative AI report moves your visibility. It reflects what Google’s AI is already doing with your pages, which on the small UK site above was a good deal more than anyone had noticed: one in six impressions inside AI answers, one 2025 page doing 84% of the work, seven in ten of those impressions on a desktop.

Read it that way and it earns its ten minutes. Baseline it, find the page the AI already trusts, make that page better, and check back monthly. If you want the fuller method for getting cited in the first place, the GEO playbook covers it, and the local business version shows how differently ChatGPT and Perplexity pick their sources.

And if you open your own report and find a page carrying a third of its impressions from AI features that you had written off as old, do not be surprised. That is, it turns out, the normal case.