On 31 August 2026, Google’s Generative AI performance report finished rolling out to websites worldwide. For the first time, there is a first-party number for how often your pages appear inside Google’s AI surfaces.
This is a genuine step forward, and we should say so plainly. Until now, almost every claim about AI visibility rested on third-party sampling, prompt-testing tools or inference. A number reported by the platform itself is a different class of evidence.
It is also considerably narrower than the coverage implies. The gap between what this report measures and what people will assume it measures is wide enough to produce some badly wrong conclusions, so it is worth being precise.
What it actually reports
Google’s documentation is admirably clear. The report covers impressions from two generative AI capabilities:
- AI Overviews
- AI Mode
Google adds: “We expect to update this list over time as we develop Google Search.” So the surface list is a moving target by design.
The metric is defined as “how many times links to your site were shown to a user in a generative AI feature on Google Search.”
That is the whole thing. One metric, two surfaces.
What it does not report
This is the part that matters, and it is where most summaries stop being useful.
No clicks. No click-through rate. No average position. The report is impressions only. You can learn that you were shown. You cannot learn from this report whether anyone did anything about it.
No Search Labs data. Google states plainly: “Search Console doesn’t include data from experiments in Search Labs, as these experiments are still in active development.” Whatever Google is testing does not appear here.
A minimum impression threshold. Your site must clear it to appear at all. Smaller B2B sites — which is most of our clients and probably you — may see nothing for a while, and nothing is not the same as zero.
The standard 1,000-row limit from Performance reports applies, so on a large site you are seeing a truncated view.
Preliminary data. Chart data may render as a dotted line, meaning it is provisional and can change within hours. Do not screenshot a dotted line into a board pack.
Chart and table totals can disagree, because they aggregate differently — property-level versus page-level. If two numbers in the same report do not match, that is expected behaviour, not a bug.
And one more, which Google does not need to state because it is a matter of scope: this is Google only. It tells you nothing about ChatGPT, Claude, Perplexity or Copilot. If your buyers research in ChatGPT — and in B2B many now do — this report is silent on the surface you may care about most.
Impression, citation, and influence are three different things
Here is where we would encourage genuine caution, because the vocabulary is about to get abused.
An impression in this report means a link to your site was shown inside an AI feature. It is mechanical and countable.
A citation — the thing most GEO conversations are actually about — means the answer was constructed using your content, and attributed. Related, but not identical.
Influence means someone read it, believed it, and it changed what they did next. Neither of the first two establishes the third.
Over the next few months you will see agencies present rising impression counts as proof that AI visibility work produced business results. That inference does not hold on this data alone. Impressions are the floor — evidence that you are present and eligible. They are a necessary condition for everything downstream and a proof of none of it.
We would rather be the firm that says this out loud than the one that quietly benefits from the confusion.
What it is genuinely good for
Having been sceptical, let us be fair about the value, because it is real.
It closes part of a gap we have been open about. We published an article in August on which AI crawlers can actually reach your site, and were explicit in it that crawl access is a floor rather than a guarantee — a crawler that cannot reach your content certainly cannot use it, but reaching it proves nothing on its own. This report gives one Google-reported signal on the other side of that gap. It does not close it. It narrows it.
It is a directional trend line you did not have. Absolute numbers matter less than movement. If AI impressions climb after a technical fix, that is meaningful signal. If they flatline after three months of content work, that is meaningful too, and considerably cheaper to learn from a report than from a renewal conversation.
It is diagnostic when it is empty. If you have solid organic impressions and nothing in the AI report, that is a specific and actionable finding. It points at eligibility problems — crawler access, snippet controls, indexation — rather than at content quality. That is precisely the class of problem we wrote about last week regarding Cloudflare’s 15 September crawler defaults, and precisely the class of problem that is invisible without a measurement surface.
It is first-party. Whatever its limits, it comes from the platform rather than from a tool sampling prompts and extrapolating. That is a meaningful upgrade in evidence quality.
How to actually use it
Four suggestions, in order.
Check whether you have access yet. Rollout is staged. Absence of the report is not absence of visibility.
Set a baseline now, before you change anything. The most common measurement mistake is starting to measure after the intervention. If you intend to do AI visibility work this quarter, record where you stand this week.
Read it alongside your logs, not instead of them. Impressions tell you the outcome. Logs tell you whether the crawlers that produce that outcome are reaching you. When impressions are flat, the logs usually explain why, and the configuration usually explains the logs.
Report it honestly. State the metric, state the two surfaces, state that it excludes clicks and excludes every non-Google assistant. A number with its limits attached is more persuasive to a serious buyer than a number without them, and considerably safer.
What we cannot tell you
We cannot tell you how AI impressions correlate with revenue. Nobody can yet — the data is days old and the correlation would need to be established per business, over time.
We cannot tell you whether impression counts here predict citation behaviour in ChatGPT or Perplexity. Different systems, different indexes, no published mapping between them.
We cannot tell you that this report will keep working the way it does today. Google has said the surface list will change as Search develops.
What we can tell you is that a first-party measurement surface now exists where none did, that it measures one specific thing well, and that the fastest way to lose credibility over the next year will be treating it as though it measures more.
The short version
Google now reports how often your links appear in AI Overviews and AI Mode. Impressions only — no clicks, no CTR, no position, no Search Labs, Google surfaces only, staged rollout, 1,000-row cap.
Use it as a baseline and a trend line. Use it as a diagnostic when it is empty. Do not use it as proof that AI visibility work produced commercial results, because it does not contain that information.
Set your baseline this week, before you change anything else.