Search Console shows AI impressions. Not your clicks.

Google's Nick Fox said in July that AI features in Search send "billions of clicks to websites every week." Google also shipped you a report that counts your AI impressions and refuses to tell you how many of those clicks were yours. Both o

Google's Nick Fox said in July that AI features in Search send "billions of clicks to websites every week."

Google also shipped you a report that counts your AI impressions and refuses to tell you how many of those clicks were yours.

Both of these are true at the same time, and the gap between them is the most important thing in search right now for anyone who sells something on the internet.

Google shipped a report that counts everything except the outcome

On 3 June 2026 Google announced new controls and insights for website owners, and the Search Central blog introduced generative AI performance reports in Search Console. They have been rolling out incrementally since, which is why yours may have appeared quietly in the last few weeks without any announcement in your account.

Here is what the report gives you, straight from Google's own help documentation: impressions in AI Overviews and AI Mode, sliced by page, country, device and date. Search Labs experiments are excluded.

Here is what it does not give you: clicks. Also no click-through rate and no position.

That is not an oversight in a v1. Impressions and clicks come from the same pipeline in every other Search Console report you have ever opened. Shipping one without the other is a decision.

The billions-of-clicks claim and the missing click column

On 17 July, Google's Nick Fox told the press that AI features in Search alone send billions of clicks weekly.

It may well be true. The problem is that it is unfalsifiable at the only level you care about, which is your own domain. Google is asking every site owner on earth to accept an aggregate reassurance in place of a number they already have.

When a platform tells you the traffic is fine and simultaneously withholds the column that would prove it, believe the withholding.

— Arthur, Bunny Honey Club

I am not claiming bad faith. There are boring explanations: attribution inside a generative answer is genuinely messy, and a half-accurate click number invites a decade of arguments. But the practical effect is the same either way. You cannot measure what AI Search does to your business, so you have to infer it.

Treat AI impressions as a brand metric, not a traffic metric

This is the reframe that actually changes what you do on Monday.

An AI Overviews impression is not a visit that has not happened yet. It is your business being named inside the answer a customer asked for. Sometimes that produces a click. Often it produces a memory, a search for your name three days later, or a phone call that arrives with no referrer at all.

Stop filing it under traffic. File it under mentions. The moment you do that, the report stops being frustrating and starts being useful, because impressions by page tells you exactly which of your pages the model considers quotable.

That is the same scoreboard we described in the AEO playbook for ChatGPT, Perplexity and Gemini. Being the source is the win. Being the tenth blue link under a complete answer is not a consolation prize, it is a rounding error.

Impressions up and clicks flat is the signature to watch

You cannot derive AI clicks by subtracting one report from another. The help documentation describes the generative AI report as drawn from the Web search type in the main Performance report, so it is a filtered view of data you already have rather than a separate stream. Subtraction will give you a confident wrong answer.

Watch the shape instead. Three patterns, and what each one means:

• AI impressions up, total clicks up. The answer is sending people through. Keep doing whatever that page does.
• AI impressions up, total clicks flat or down. The answer is complete enough that nobody needs you. Your content is good enough to quote and not compelling enough to visit.
• AI impressions near zero on pages that rank well. The models are not selecting you as a source. That is a structure and citability problem, not a rankings problem, and we broke down the fixes in how to build a blog that ranks and gets cited by LLMs.

The middle one is the common case in 2026 and the one people misdiagnose as an algorithm penalty. It is not a penalty. It is a page that answered the question so tidily that the answer engine could finish the job without you.

Local businesses feel this first and measure it last

If you run a clinic, a trade, a restaurant or an agency, the AI answer is increasingly the whole result page for questions like "best physiotherapist near me who takes evening appointments."

Your Google Business Profile, your reviews and your service pages are what gets assembled into that answer. None of that shows up as a click in any report you own. The phone rings, and the call has no source attached to it. This is the measurement hole we flagged in local SEO for small business, and the generative AI report does not close it. It just confirms the impressions are real.

One practical fix that costs nothing: ask new customers how they found you and write the answer down. A tally in a notebook beats an analytics dashboard that structurally cannot see the channel.

What we changed on this blog

We run this blog on the same rules we sell, so here is what we actually did rather than what we recommend in the abstract.

We stopped reporting AI impressions next to sessions, because putting them in the same table made them look like traffic that had gone missing. They now sit in their own line called mentions, and we judge them on whether the right pages are being quoted.

We started reading impressions-by-page as an editing brief. When a page collects AI impressions and no clicks, it usually means we answered completely and gave nobody a reason to come. The fix is not to answer worse. It is to put the thing that cannot be summarised, the real number, the actual opinion, the specific build, where the summary has to stop.

And we compare across engines rather than trusting one. Google is the only one handing you a report at all, which makes it easy to over-index on it. The comparison of Perplexity, ChatGPT Search and Google AI Mode is still the honest picture of where your customers actually ask.

If you want the content built to be quoted and the measurement wired up so you can see it working, that is what our content work is. If you would rather do it yourself, the report is free and the notebook costs three pounds. Both of those are fine answers.