Every Search Console Account Now Has an AI Report. Here’s What It Won’t Tell You.
On 31 August 2026, Google finished rolling the Generative AI performance report out to every Search Console property in the world. If you log in this week, it’s there. You didn’t ask for it, nobody emailed you about it, and it is sitting in your Performance menu next to the reports you already know how to read.
It had been a limited release until now. Google first opened it on 3 June to a subset of UK site owners, under pressure from the Competition and Markets Authority’s work on the Digital Markets Act. That’s why British businesses saw it before anyone else. As of last Monday, everyone has it.
So a great many people are about to open a report they have never seen before and draw a conclusion from it. Most of those conclusions will be wrong, for one reason that Google states plainly in its own documentation and almost nobody reads.
The report counts impressions. It does not count clicks.
That single fact changes what the report is for. Here is how to open it, what each part actually measures, and how to read your own numbers without fooling yourself.
Where it is
Search Console → Performance → Generative AI.
If you’d rather go straight there: search.google.com/search-console/performance/search-analytics/ai
There is a second, separate report for Google Discover. This one covers Search only.
What it counts
The report includes impressions from exactly two surfaces:
- AI Overviews — the AI-generated summary block at the top of a results page
- AI Mode — Google’s conversational search experience
That’s it. Google says the list will grow as Search develops, but today those are the two. Experiments running in Search Labs are excluded entirely, because they’re still in active development.
An impression means a link to your site was shown to a user inside one of those two features. It does not mean the user read it, noticed it, or acted on it.
The four dimensions, and the one that catches people out
You can group the data four ways, using the tabs above the table.
Pages. Groups by the final URL linked after any redirects. This is the one that catches people out: like the ordinary Performance report, page data is assigned to the canonical URL, not to the duplicate the user actually saw. If you have a page reachable at several addresses, all of its impressions land on the canonical one. That’s usually what you want, but it means the URL in the report may not match the URL in your analytics.
Countries. Where the search originated.
Devices. Desktop, tablet or mobile.
Dates. By day, week or month. All dates are Pacific Time, which is the same convention as the ordinary Performance report, and the same reason your daily totals never quite line up with a UK-based analytics tool.
Two more things worth knowing before you start reading numbers into it. The chart aggregates by property — if two of your pages appear in the same AI answer, that’s one impression on the chart. The table aggregates by page when you group by page. That’s why the chart total and the table total often disagree, and it is not a bug. The usual limits apply as well: 1,000 rows, and the newest day or two may be preliminary, shown as a dotted line, and may still change.
If you can’t see the report, or it’s empty
There are three explanations, and they are not equally likely.
One: you genuinely have no impressions. Plausible for a small site in a niche where Google isn’t generating many AI answers. Not something to panic about.
Two: the rollout hasn’t reached your property. Google’s own note says not all properties have access yet, even after the worldwide announcement. Give it a fortnight.
Three — and this is the one to check — your site is excluded from generative AI features. Google shipped a control alongside the report, on the same day, at Settings → Search generative AI. Direct link: search.google.com/search-console/settings/search-gen-ai
It has three settings: include (the default for all properties), exclude, or inherit control from parent.
That last one is the trap. If you have more than one Search Console property — a domain property and a URL-prefix property, say, or a separate property for a subdomain — a child property inherits its setting from its closest parent by default. Change the setting on example.co.uk and you may have silently changed it for blog.example.co.uk and https://example.co.uk/shop/ too. Worth ten seconds of checking if you manage several properties, or if a previous agency set them up.
While you’re in there, two clarifications that Google makes explicitly and that are widely misreported:
- Excluding your site is not a ranking signal. It affects whether you appear in AI Overviews and AI Mode. It does not affect the rest of Search.
- Excluding your site does not stop your content being used for AI training. That’s a different mechanism entirely — the
Google-Extendeduser agent in your robots.txt. Two separate controls, two separate purposes, routinely confused.
Changes take about one to two days to take effect, sometimes longer.
How to actually read it
The report on its own tells you very little. Read against your ordinary organic data, it tells you quite a lot. Open the Generative AI report and the standard Performance report side by side, set both to the same date range, and look for one of three patterns.
AI impressions rising, organic clicks flat.
The most common pattern, and the least alarming. You’re being surfaced in AI answers and your ordinary traffic is unaffected. Google is quoting you, people are still coming to you by the usual route. Nothing to fix. Keep doing what you’re doing and check again next quarter.
AI impressions rising, organic clicks falling.
This is the pattern everyone is worried about, and it’s the only one that justifies real work. It suggests your content is answering the question well enough that people no longer need to visit. Look at which pages: it is nearly always the informational ones — definitions, how-tos, “what is” pages. It is very rarely your pricing page or your contact page, because those aren’t questions an AI answer can close out.
The response is not to make your content worse or harder to summarise. It’s to be honest about which pages were ever going to convert and which were always top-of-funnel. If an AI answer is intercepting a query that was never going to buy from you, you have lost a session, not a customer.
No AI impressions at all.
Check the control setting first, as above. If you’re included and still seeing nothing, the likeliest explanation is that Google isn’t generating AI answers for the queries you rank for, which is common in narrow B2B and specialist trade sectors. This is worth monitoring rather than acting on. It can change quickly.
One more caution on the data itself: the report is new, so your history is short — there’s no backfill of the years before it launched. Any month-on-month comparison you run before roughly the end of this year is being drawn from a very small sample. Treat early trends as weak signals.
What this report cannot do
It is worth being precise about the gaps, because the gaps are where the bad decisions come from.
- No clicks. You cannot calculate a click-through rate. You cannot tell whether an impression was worth having.
- No queries. You cannot see which questions triggered the AI answer. This is the single biggest limitation, and the one that most obviously separates this report from the Performance report you’re used to.
- No position. There’s no equivalent of average position — no sense of whether you were the first source cited or the fifth.
- No competitor data. You can see that you appeared. You cannot see who else did, or who appeared when you didn’t.
And one gap that is bigger than all of them: this report only covers Google. ChatGPT, Gemini and Perplexity do not appear in Search Console and never will. For a lot of businesses those are now where the questions get asked first. Search Console tells you what happened inside Google’s two AI surfaces. It tells you nothing at all about the rest.
The thing worth noticing
There’s a familiar shape to this. A new metric appears, it is easy to measure, it goes up, and within a few months people are optimising for it — writing reports around it, setting targets against it, and quietly reorganising their content to make the number rise.
An impression in an AI Overview is not a visit. It is not a lead. It is a record that a machine mentioned you to somebody. That is genuinely worth knowing, and it is worth tracking as a trend over quarters. It is not worth treating as a total, and it is certainly not worth optimising for on its own, because you cannot see what happened next.
The useful version of this report is the boring one: open it once a month, put it next to your organic clicks, and look for divergence. If the two lines move together, carry on. If they separate, go and look at which pages.
And the work that makes you appear in AI answers in the first place has not changed at all. Clean semantic structure, proper schema markup, prices and answers in readable text rather than images, fast stable pages, and content that answers a real question in full.
Worth a footnote here. When we wrote about what the llms.txt data actually says in May, we had John Mueller calling the file “not done for search” while the Chrome team was simultaneously shipping a Lighthouse audit that checked for it. Google has since settled its own argument: on 15 June it added a note to its AI optimisation guide stating plainly that llms.txt files aren’t needed for Google Search and will affect your visibility neither positively nor negatively. Add one if you like — we still do — but nobody should be charging you a strategy fee for it.
The new report gives you a number. The old discipline is still what moves it.
Seeing the other half
Search Console now shows you Google. It’s free, it’s yours, and you should absolutely read it every month.
What it can’t show you is what ChatGPT, Gemini or Perplexity say when someone asks them for a supplier like you — whether you’re named at all, whether what’s said is accurate and current, and who’s being recommended instead. That’s AI Visibility Monitoring: we put your real customer questions to all three every month and send you one page in plain English telling you what changed and what we’d do about it. From £95 + VAT pcm, covering up to 25 tracked questions and five named competitors.
If the more useful starting point is understanding why the answers look the way they do — whether an AI system can actually read and use your site — that’s the Agent Readiness Audit, assessed as part of our Technical Business Benchmark. £750 + VAT, credited in full against any work you commission afterwards.
Schedule a call and we’ll tell you honestly which one you need — or whether you need either.