Google Now Counts Your AI Search Impressions, Not Your Clicks

Open Search Console this week and there may be a menu entry that was not there in the spring. It sits under Performance, it is called generative AI, and for a lot of independent sellers it is the first time Google has admitted, in writing, that its AI answers show their pages to people. The number is real. It is also thinner than most of the commentary around it suggests, and one setting two clicks away from it can quietly remove your shop from those answers for good.

Here is what the report actually contains, what it deliberately leaves out, what the regex trick doing the rounds can and cannot substitute for, and the two decisions worth making before the end of the month.

What exactly did Google add to Search Console?

Google added a dedicated view for impressions inside its generative search surfaces, and only impressions. The Search Console help page for the generative AI performance report defines an impression as a count of how many times links to your site were shown to somebody inside a generative AI feature on Google Search, and it names the two surfaces covered: AI Overviews and AI Mode. Discover has its own separate report.

The dimensions available are pages, countries, devices and dates. You can regroup the table by any of them and export both the chart and the table. That is the whole feature.

Two details in the help page matter more than they look. The first is that data from Search Labs experiments is excluded, because those experiments are still moving. The second is the deduplication rule: if two results from your site appear in the same generative AI feature, they count as one impression, not two. A shop whose category page and product page both get cited in the same AI Overview sees a single tally mark.

The rollout is still partial. Google's help page says it is going to a subset of website owners, so an empty menu is not evidence that your shop is invisible in AI answers. It is evidence that you are not in the test group yet.

One more caveat is written into the same page and gets ignored in every screenshot going round social media. The most recent days are preliminary: the figures are still being gathered and can change within hours. Anybody publishing a week over week comparison built on the last two days of data is comparing a settled number against one that has not finished arriving. Wait three days before you believe the tail of the chart.

The ordinary Search Console row limit applies here too. A thousand rows sounds generous until you group by page on a catalogue with several thousand products, at which point the table is showing you the top of a distribution and silently dropping the rest. Filter to a directory before you export if your catalogue is large, and export each directory separately rather than trusting one truncated pull.

Why are there no clicks in the report?

Google has not published a reason, and any explanation offered elsewhere is inference rather than fact. What can be stated plainly is the practical consequence: an impression inside an AI answer is not a visit, and the report gives you no way to convert one into the other.

This is the number that will get misread most often. A shop that sees forty thousand generative AI impressions in a month has not learned that forty thousand people considered buying from it. It has learned that its links were rendered inside an AI answer forty thousand times, in a format where the answer above the link frequently satisfies the question on its own. Without a click count, the ratio between the two is unknowable from Google's own data.

There is one thing you can do about that gap, and it costs nothing. Your analytics platform still records referrals. Traffic arriving from an AI Mode session or from an AI Overview link carries a Google referrer, so segmenting landing pages by entrance and comparing the trend against the impression curve gives you a directional read. It is not attribution. It is two lines on a chart that either move together or do not, which is more than the report alone offers.

Can you see the prompts people typed?

Not really, and the honest version of the answer matters because the workaround is being shared as though it were a prompt log. What the regex methods find is long queries in the ordinary performance report, on the reasonable theory that conversational phrasing looks different from the two word searches people typed for twenty years.

Search Engine Land published a method for surfacing conversational queries with a custom regex filter, credited to Chris Long of the agency Nectiv and first posted on the Nectiv blog in February 2026. The pattern matches any query of ten words or more. Long is candid in the piece that there is no direct evidence such queries come from an AI surface at all, and that no perfect rule exists for deciding which ones to track.

Practical Ecommerce ran a simpler variant of the same long query filter in August 2026, with two caveats worth repeating to anyone who reaches for this. Search Console never states which queries are prompts, so every classification is your judgement. And a long query with an implausible impression count may be a rank tracking tool rather than a shopper, which inflates exactly the rows that look most exciting.

Used carefully, the filter still earns its keep. A run of ten word queries mentioning a material, a size, a compatibility question or a use case tells you which product attributes shoppers are reasoning about before they buy, and those are the attributes to put in writing on the page. That is a content brief, not a traffic report, and it is a good one. The same instinct sits behind our note on what makes a product page readable to an AI shopping assistant: the fix is almost always stating a fact the page was leaving implicit.

Treat the count as an upper bound. A ten word filter also catches genuine long tail searches that have nothing to do with any assistant, so the number it returns is the largest the conversational share could be, not the size it is.

The setting that makes your shop invisible

Two clicks from the new report there is a switch that deserves far more caution than it is getting. Under Settings, the Search generative AI control lets a site owner remove their content from AI Overviews, AI Mode and the generative features in Discover. Google's help page is unusually direct about the effect: your content is prevented from being visible to users in those features, including being linked to inside them.

Read that twice. Opting out does not mean your pages appear without a summary. It means they do not appear.

The same page is equally clear about what the control does not do. It does not change your ranking in ordinary Search results. It does not affect other Google products. It does not govern whether your content trains a model, which is a different mechanism handled by the Google Extended crawler token. Changes take a day or two to propagate.

That last distinction is where sellers get tangled, because the two levers are constantly discussed as one. Blocking crawlers is a decision about who may read your pages and on what terms. The generative AI control is a decision about whether Google may show your pages to a shopper inside an answer. You can plausibly want the first without wanting the second, and we walked through the trade in detail in our piece on whether to allow, block or charge the AI crawlers hitting your shop.

For almost every independent seller the answer to the generative AI control is to leave it alone. The publishers with a reason to consider it are the ones whose entire product is the text on the page, where an AI summary substitutes for the visit. A shop selling objects is not in that position. Being named inside the answer a shopper is reading is the outcome you want.

What does Google say does not work?

Google's guide to optimizing for generative AI features is blunter than most vendor documentation, and it contradicts several tactics currently sold as essential. It states that llms.txt files and similar markup are ignored by Google Search. It says there is no need to chunk content into small pieces. It says rewriting your pages specifically for AI systems is unnecessary, and that manufactured brand mentions scattered across the web do not help. Structured data remains useful for the rest of your SEO but is not a requirement for the AI features.

The prerequisite it does state is simple. A page has to be indexed and eligible to appear in Google Search with a snippet. Everything else in the guide is ordinary quality advice: material with a point of view rather than recycled description, semantic markup, working JavaScript rendering, a page that behaves on a phone, and for sellers, a properly maintained Merchant Center and Business Profile.

None of that is new, which is the point. If you were waiting for a technical trick to unlock AI visibility, Google has said in its own documentation that the trick does not exist. Our longer treatment of what actually moves citations in AI answers reaches the same conclusion from the other direction, by looking at what cited pages have in common.

What should a small shop do this month?

Four things, none of which take an afternoon.

Check whether the report exists in your property, and if it does, export it before you do anything else. Search Console truncates at a thousand rows and keeps a limited window of history, and because the generative AI data is absent from both the API and the bulk export there is no automated safety net. A monthly export into a spreadsheet is the entire archive you will ever have of this period.

Group the report by page and look at which URLs are being cited. This is the most useful dimension in the whole feature and the one people skip while staring at the headline number. If AI answers are citing your shipping policy and your size guide rather than your product pages, that tells you what the machine currently finds quotable about your shop, and it is rarely what you would have chosen.

Run the long query filter once, read the rows as language rather than as metrics, and write down the recurring attributes. Then fix the pages. A shopper asking a ten word question about whether something fits a particular device is asking for a sentence that is missing from your page.

Leave the generative AI control switched on unless you can articulate why being absent from AI answers helps you sell more. If a supplier, an agency or a forum thread tells you to switch it off, ask them what they expect to gain, and check their answer against Google's own wording about what the setting does.

Does an impression in the report mean somebody saw my brand?

It means a link to your site was rendered inside a generative AI feature. Whether the person read it, scrolled past it or never reached that part of the answer is not something the impression count distinguishes, which is true of impressions in the ordinary Search report as well.

Why is my generative AI number so much smaller than my Search impressions?

Partly because the AI surfaces are a fraction of total Search volume, and partly because of the deduplication rule: several of your pages appearing in one AI answer count once. The two numbers are not built to be compared as a ratio.

Will blocking AI crawlers remove me from AI Overviews?

Those are different controls with different scopes, and Google's help page for the generative AI control says so explicitly by pointing at the Google Extended token for training instead. Decide them separately, because the commercial consequences are not the same.

Should I build my product pages differently for AI answers?

Google's optimization guide says no, and specifically calls out AI targeted rewrites and llms.txt as things Search ignores. Writing a clearer page helps a shopper and an AI answer for the same reason, which is that both are looking for a stated fact rather than a mood.

The part nobody has solved yet

A seller running a shop alone now has an impression count with no clicks, a query filter with no proof of origin, and a switch whose only setting worth choosing is the default. That is a strange package, and it is worth naming rather than dressing up. The report is a first draft of a measurement that the industry needs and does not yet have.

What it does give you, immediately and for free, is the list of pages Google's AI is willing to cite. Read that list honestly and it will tell you where your shop is clear and where it is vague, which is the same information a good rebuild of a product catalogue starts from. If you are at the point of restructuring the pages themselves rather than patching them, our AI store builder generates the catalogue structure and the copy on code you own outright, which keeps the fix in your hands rather than a platform's.

Export the report. Read the page list. Leave the switch alone. Everything else about AI search visibility this month is still guesswork, and the people selling certainty about it are working from the same thin data you now have in front of you.