Google Just Shipped AI Performance Reports in Search Console
It shows where you're cited, not whether you mattered. The clicks, conveniently, are none of your business.
Yesterday, Google shipped a dedicated Generative AI performance report in Search Console, with separate views for Search and Discover, covering how often your pages surface inside AI Overviews and AI Mode. The Search team (Hillel Maoz and Moshe Samet) announced it as the company’s first real cut at first-party AI search visibility. The teams that took AI search seriously were not waiting on it. They had been measuring this on a dedicated visibility platform while Google offered nothing, and now Google has finally shown up to the category.
This is a real step, and I want to say that plainly before I get into the parts everyone is going to skim past. We have wanted first-party AI visibility data since AI Overviews started eating the top of the page. Now a slice of it exists, inside the tool we already trust for organic, and attributed to Google itself.
Then you read the help documentation, and the report quietly tells you what Google is willing to measure and what it is not. The second list is longer, and a lot more interesting.
Start with one sentence
Skip the launch bullets. The line that matters is in the help doc, in the definition of an impression. Google counts an impression as how often “links to your site were shown” inside a generative AI feature on Search.
Read that twice. The unit is a link shown, not your content used. That distinction is the entire AEO game, and Google just drew a hard line straight through the middle of it.
AI Overviews and AI Mode do two different things with your pages. Sometimes they surface a link, a citation chip a user can click. Often they pull a sentence, a stat, a definition off your page to ground the answer and never surface a link at all. The first is citation. The second is influence. AEO has always been about both, and you can argue the second matters more, because being the source the model trusts is more durable than being one of six links in a carousel.
This report sees the first and is blind to the second. When your content shapes an answer without earning a visible link, your impression count does not move. So the number in front of you is not “how much am I showing up in AI,” it is “how often did Google decide to link me.” Those are very different questions, and treating them as one is the first mistake people will make with this data.
One tell that Google knows the difference: its companion blocking control (more on that below) lets a site opt out of being used “as links or for grounding purposes.” Google separates the two when it suits them. The report only reports one of them.
The pros spotted both problems within a day
Watch who said what under Google’s announcement, because the split is the real signal. The “finally, thank you” replies landed fast. So did sharper reads from people who do this for a living, and they converged on the same two complaints.
The first is granularity. The report blends AI Overviews and AI Mode into one impressions bucket. Giorgio Taverniti, who wrote the book on liquid search, did not soften it: aggregating AI Overviews and AI Mode, and shipping nothing on grounding queries, is the kind of choice you make when you have drifted away from the people working in the field. Rebekah Logan made the operator’s version of the same point. Citation behavior in AI Overviews versus AI Mode can differ wildly, so one blended number hides the thing you most want to know, which surface is citing you and which is not. Simone De Palma called the aggregated sub-menu “data engineering 101” and noted that Bing at least ships query-level data from its engine.
The second is the grounding gap, the same one sitting inside the impression definition. Jayme Welch put it cleanly: this is a measure of visibility, not understanding, and it tells you nothing about how often AI features used your content without a citation. Radu Stoian and Yongjoon Yang both asked for query fan-out, what Bing calls grounding queries. That request is going to define the next year of this conversation, and Google shipped without it on purpose.
Notice the divide. The loudest praise was excitement that the report exists at all. The most specific reactions came from the people who pressure-test AI visibility for a living, and they found the same two holes I did, inside a day. Excitement that a thing exists is not evidence the thing is good.
The missing column is not an oversight
There are no clicks. No click-through rate, no average position, no query data. Just impressions, sliced by page, country, device (Search only), and date, down to hourly granularity, with data that appears to begin around May 18.
Barry Schwartz asked Google directly about clicks and got the line you would expect: they are working with site owners on what is helpful and will add metrics over time. Maybe. But I would not hold your breath for clicks, and not because instrumenting them is hard.
A clicks column would quantify the one thing Google has the least incentive to quantify: how much traffic the AI answer keeps for itself. The whole economic logic of AI Overviews is to resolve the query on the page so the user never needs to leave. A per-page, exportable click number sitting in everyone’s Search Console would put a hard figure on that cannibalization. “Additional metrics over time” is a real roadmap. Clicks are the last room in the house they unlock.
So treat the impressions number as a denominator, not an answer. It can show you the shape of your AI visibility and how it moves. It cannot tell you whether AI search sent you a single visitor. For that you still need your server logs and your analytics, and you still need to do the matching yourself.
Why the UK got it first (not for the reason you think)
The rollout is a small subset of UK site owners, reportedly limited to .co.uk properties at first, ahead of a global expansion with no date attached. The instinct is to read that as a routine phased launch. It is not.
The UK went first because the Competition and Markets Authority is forcing the issue. The CMA’s remedies require Google to give publishers genuine control over how their content feeds AI, including the ability to opt out of having it used to fine-tune models, and Google has roughly nine months to roll that out across the UK. The legally mandated piece is the controls. The report is not itself required by the CMA, but it shipped in the same market on the same day, which is not a coincidence. When a regulator makes you hand publishers an opt-out, you want to be able to point to a reporting product launched alongside it and call the whole thing a publisher win.
That reframes the timeline for the rest of us. The pace of access is being set by antitrust pressure in the UK and the EU, not by Google’s eagerness to hand operators better data. If you are in the US waiting for this, you are waiting on regulators in another jurisdiction, not on a product roadmap you can lobby.
Read the rollout cynically and you land somewhere uncomfortable. This looks timed to coincide with the CMA controls and scoped to do just enough to keep regulators satisfied. And notice the detail that should bother you most: the data starts in May 2026, with no history before it. So you cannot line up a decline in your own clicks against the rise of AI Overviews, even if you wanted to. Convenient.
The sleeper launch is the toggle, not the report
Shipped the same day, and arguably more consequential: a control that lets a site opt out of appearing in AI Overviews, AI Mode, and generative features in Discover, both as links and as grounding. Opt out and you get no impressions and no traffic from those surfaces. Google says it will not use the choice as a ranking signal in core Search, so in theory you can leave AI without leaving Search. For the test cohort it takes effect June 17.
In an earlier poll, about a third of SEOs said they would block Google’s AI features. I think almost none of them should, and I suspect almost none of them will once the toggle is real and the thing they would be zeroing out is their own impression count. The honest use of this control is narrow: publishers whose entire business is the click and who have decided AI exposure is pure leakage. For nearly anyone in B2B or considered purchases, AI surfaces are now part of how a buyer builds a shortlist before they ever search your brand. Opting out to protect a click you were probably going to lose anyway is a bad trade. It is still good that the lever exists, because consent should not require a robots.txt hack.
What to actually do with it on Monday
A few concrete moves, whether you have access now or are planning for when you do.
Build the page-level citation inventory. The Pages view is the most useful tab here. It tells you exactly which URLs Google is willing to link inside an AI answer, which is your current answer-worthy set in Google’s eyes. Put it next to the pages you assume are your strongest answer assets. The gap between those two lists is your roadmap, and closing it is a production problem more than an analysis one. This is where a content engine like AirOps earns its place, because turning a list of under-cited topics into answer-shaped pages at real volume is the work, and doing it by hand does not scale to the surface area AI search now covers.
Use impressions as answer-share input, not a traffic proxy. Track the trend and the page mix, watch the country and device splits (AI behavior diverges on both), and kill the reflex to convert impressions into sessions. They are not sessions, and a chart that implies they are will get you in trouble in a board meeting. Plenty of operators in that LinkedIn thread were already drafting the client line, “your content showed up 9,000 times in AI answers this week,” and that is a fair story to tell as long as you are honest that it is a visibility story and not a value one. The moment it gets dressed up as traffic, you have lied to a client with Google’s logo on the slide.
Triangulate, do not switch. This does not replace a cross-engine AI visibility platform like Profound, and anyone telling you GSC makes one redundant has not understood either tool. They are not measuring the same thing. Profound tracks your share of voice and citations across the engines that actually carry the conversations, ChatGPT, Perplexity, Gemini in its own app, Copilot, while GSC reports Google’s own count of links displayed on Google’s own surfaces. Those numbers will not line up, and they are not supposed to. One is a cross-engine market-share lens, the other is a single-vendor, single-metric ground truth. Read them together. GSC becomes the first-party Google reference point you anchor to, and the platform stays the only thing giving you the rest of the map, which is most of it, because Google is one engine among several and AI search is not happening only inside Google.
Keep an eye on what Bing already shows. Bing’s AI performance report has been global for a while and exposes more signal than Google’s, including citations and grounding queries. No clicks there either, but the contrast is the point: Bing surfaces the grounding-level data Google is choosing to withhold. If you operate at any scale, treat Bing’s view as a useful second instrument, not because its volume matters, but because it sees the half of the picture Google won’t show you.
The measurement gap is half-closed
Here is the state of play. We went from no first-party AI data to a real, Google-attributed view of where our links land in AI answers. That is progress and I am glad to have it.
But the half that is still missing, clicks and queries and grounding without a link, is exactly the half that would settle the argument everyone is actually having: is AI search a traffic apocalypse or a new front door. Impressions tell you that you are in the room. They do not tell you whether anyone walked through the door, what they asked to get there, or how often your words ended up in the answer with your name stripped off.
We got a better altimeter this week. We still have no airspeed indicator. Plan accordingly, and read the fine print before you put any of it in a deck.