“Where Exactly Is That Number From?”
“Our AI exposure looks like it’s going up.” I said something close to that in a monthly review once, and the question came straight back: going up according to what?
Nothing empties your hands faster than being asked for the source of a figure you half-remember from somebody else’s blog post. I skipped that slide two months running rather than answer it.
But since June, Google’s been handing out a report that counts your exposure on its own search surfaces, with Google’s name on the front of it.
So the column I’d been filling in with vibes now has an issuer. Let’s go through what the documentation says is in it, and what isn’t.
How To Check How Often Your Pages Appeared In Google’s AI Answers
Search Console has it, under the name generative AI performance report. Google announced it on 3 June 2026, and it counts the times a link to your site was shown inside AI Overviews and AI Mode, broken out by page, country, device and date (R).
AI Overviews is the AI summary that sits above the ordinary blue links. AI Mode is the conversational surface, where you put a question in the search box and get an answer back you can keep talking to.
What the report carries is a single metric, impressions. Clicks, CTR (click-through rate), average position and the search queries themselves are all absent from it (R).
So the surfaces being counted are the two where Google itself is doing the answering. Discover’s generative AI features sit in a separate report, and the help page says the list is expected to be updated.
What The Four Slices Actually Give You
Pages group by their final URL, the canonical one — the single address Google picks to stand in for a page when the same content is reachable from several.
Country is where the search happened, device splits desktop, tablet and mobile, and dates run from hourly granularity up to monthly. The clock underneath all of it is Pacific Time, which matters the moment you line these figures up next to anything you keep in your own time zone.
The metric side is thinner than the slicing side. One number, defined as the count of times a link to your site was shown to a user inside a generative AI feature.
It’s a till receipt with the total printed and the item lines left blank. You can see how many times somebody went through the checkout; what they put in the basket isn’t on there.
Two exclusions are written into the documentation. Search Labs (Google’s experimental programme for features still in testing) is kept out of the data, and a site that has opted out of appearing in generative AI features doesn’t show up in this report at all.
The source of the figures is the Web search type of the existing Search performance report, which is why the shape of the thing feels familiar.
What To Write Down When The Report Isn’t In Your Property Yet
The rollout is staged — Google is releasing the report to some site owners at a time, not to everybody at once.
So an empty panel tells you something about the rollout and nothing about your brand. Having no data and having had no exposure are two different states, and only one of them is a fact about how the models treat you.
Why does that distinction earn a whole section? Because if you skip it, a blank gets filed as “zero mentions”, and somebody downstream makes a decision on the back of that number.
Hold the distinction and the reverse happens. You become the person in the room whose figures get believed, which over a year is worth considerably more than any single figure in the deck.
Google Says Your Existing SEO Still Does The Work Here
The reasoning behind that sits in an optimisation guide for generative AI features, announced on 15 May 2026 (R, R).
The half of the guide worth reading twice is the list of things it says you don’t have to do:
- “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search … as Google Search itself doesn’t use them.”
- “There’s no requirement to break your content into tiny pieces for AI to better understand it.”
- “You don’t need to write in a specific way just for generative AI search.”
- “Structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add.”
Two mechanisms are offered as the reason. The first is RAG, which the guide also calls grounding: a technique for improving the quality, accuracy and freshness of AI responses “by relying on core Search ranking systems to retrieve relevant, up-to-date web pages”.
The second is query fan-out, described as “a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results”. The model asks itself several versions of your question at once and goes shopping for pages for each one.
Since both mechanisms come in through the same door — core ranking — the guide lands where you’d expect: “The best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems.”
Whether AI-specific files get read at all has been measured from outside Google too, and it pointed the same way. I went through that one in llms.txt against plain HTML.
The Two Columns This Report Leaves Blank
One official figure is one column, and the useful question is what you stand next to it.
The first blank column is every assistant that isn’t Google. Mentions and citations inside ChatGPT, Perplexity and Claude appear nowhere in this report, and they can’t — the issuer is Google and the scope is Google’s own surfaces.
The second is how you got talked about. An impression records an appearance and passes no judgement on it, so being named as the first recommendation and being wheeled in as the comparison also-ran both land in the data as the same 1.
That second gap is the part I find genuinely useful about the first column arriving. Once the base layer — how often you turned up on Google’s surfaces — comes with an issuer’s name attached, you can finally see the exact shape of what’s still missing.
Google-surface impressions, appearances inside the other assistants, and the wording around your name. Get those three and AI exposure turns into something you can plan against.
How To Tell Where A Measurement Tool Gets Its Numbers
The optimisation guide has a line about measurement tools tucked into it:
Be wary of third-party tools that promise ranking success or claim to use “internal” Google metrics. No third-party tool has access to our internal ranking or AI systems.
Read closely, that’s aimed at a specific claim: a tool telling you it can see inside Google’s ranking or AI systems. That’s a claim about access, and Google is saying flatly that nobody outside has it.
A tool that sends questions to an assistant and counts what comes back in the answer never touched anything internal. What it’s reading is the published response, the same text you’d get typing the question in yourself.
So one question does the sorting for you. Does the tool’s own description claim figures from inside Google, or figures counted from outside, off responses anybody could reproduce?
Open Search Console And See Whether The Panel Is There
One screen today. Open Search Console, go to your property, and look for the generative AI performance report.
If it’s there, that’s your first official number for AI exposure, and sorting it by page tells you which of your articles have been turning up in Google’s AI answers. Take that into the next monthly review and the “according to what?” question answers itself.
If it isn’t there, you’re waiting on the rollout, and the whole job is writing “not available yet” in the cell instead of a zero. Boring, and it’s the judgement call that matters most this month.
Either way, keep the edges of the thing in view. What arrives is impressions on Google’s surfaces; the other assistants and the language wrapped around your name stay exactly as blank as they were yesterday.
Maybe I’m making too much of one report — one metric, from one company, about that company’s own surfaces. What I’d still bet on is what changes in a meeting when you can name where a figure came from and nobody else can.
Opening the screen takes three minutes. Go and do that one today.
