AI Overviews: what changed in the click, measured

Every conversation about AI Overviews eventually arrives at a number. Clicks fell by some percentage, somewhere, for someone. The numbers come from studies with different samples, different countries and different definitions of a click, and they are not wrong so much as they are not yours.

Your site has the only sample that decides your budget. This is how to read it.

What an AI Overview is, from the outside

When a search result page carries an AI Overview, the block arrives in the results payload like any other block, next to the organic results, the featured snippet and the local pack. It has a position on the page, and it usually carries links to the sources it drew from.

That means two things are observable without guessing. Whether the block appeared for a given query, and which domains it cited. Both come back from a normal read of the results page, and both can be stored with a timestamp so that "it appeared" becomes "it appeared on this date, and here is what it cited".

The same is true of the assistants, in a different shape, and measuring a brand inside them is its own method. What does not come back is anything about the person. No impressions for the block, no reading time, no indication of whether the block was expanded. Those live inside Google, and nobody outside has them.

The two halves you can actually measure

Half one, the surface. Read the results page for the queries you care about and record, per query and per date: was there an AI Overview, which domains did it cite, were you among them, and where did your organic result sit. Repeat weekly. After a month you have a presence series, which is enough to say "this query started carrying an AI Overview in the second week of the month".

Half two, your own property. Search Console gives you impressions, clicks and average position per page and per query, for free, for your site. It is the only place with the second half of the story, and it belongs to you.

Now the honest caveat, and it is the reason this post exists. Search Console does not label AI Overview impressions separately. There is no filter that says "show me only the times my page was cited inside the block". So you cannot attribute a click change to the block from Search Console alone.

What you can do is compare, and be clear that it is a comparison.

A before and after you can defend

Pick a window and split your queries into two groups: the ones that gained an AI Overview during the window, and the ones that did not. Same site, same period, same seasonality, same brand.

Then compare the two groups on the same three measures from Search Console: impressions, clicks, and clicks divided by impressions. If the group with the new block moved differently from the group without it, you have a difference worth investigating. If both moved together, you have found seasonality, and you have saved yourself from telling a client a story about AI.

This is a control group, not an attribution. Say that out loud in the report. It is a much stronger position than a borrowed percentage, because the client can see their own query list in it.

Two practical notes. Split by intent as well, because informational queries and brand queries behave differently and mixing them turns the comparison to mush. And use enough queries that a single unusual one cannot swing the result.

The change that broke a lot of tooling

There is a second story running underneath this one, and it matters to anyone comparing tools right now.

On the twenty-sixth of August 2026, Google confirmed that the links on organic results had changed shape. Where the href used to be the destination URL, it became an address on google.com carrying an encrypted token. The visible parts of the result did not change: the title, the site name, the breadcrumb and the snippet are all still on the page. What changed is the href and what you see when you hover.

For anyone reading results by scraping the page, that turned a direct read into an extra step: resolve the token to find out where the link actually goes. The resolution itself is cheap, one request that comes back with the destination in a redirect header, and it does not touch the destination site at all. But scraping in general got harder at the same time, and a lot of tooling had a bad week.

Two things follow, and we would rather be plain about both.

We do not scrape search results. We buy them resolved from a data provider, and that provider had organic links resolved within a day of the confirmation and AI Overview links within two. So this change cost us nothing, and we are not going to claim we solved something we bought.

The defensible position was never "we can read the hover". It is that data is pulled when you ask for it, about the thing you are looking at, instead of crawling everything on a schedule and paying for freshness you never read.

What the tool records

For each front a client watches, the daily check reads the results page for that formulation and stores a row: the date, the position, who occupies the top of the page, and which blocks were present, the AI Overview among them. Rows are never overwritten, which is what makes the thirty day line a record instead of a snapshot. When the position drops three places or more, or leaves or enters the top ten, it raises a flag.

Reading results costs money per query, so the daily check is deliberately narrow: only the fronts explicitly marked to be watched, only up to the plan's count, and every call goes through the same gate as a manual one, with the charge taken before the call and refunded if it fails. Everything else in the product waits for a click.

On the other side, the reports module pulls clicks and impressions per page from Search Console, which is free, and keeps the history so the comparison above can be built without asking anyone to export a spreadsheet.

What we are not saying

We are not going to tell you that AI Overviews killed the click, and we are not going to tell you they did not. We do not have a sample that would support either sentence, and neither does most of what gets quoted.

What we can say is narrower and more useful. The presence of the block is observable and datable. The domains it cites are observable. Your own clicks and impressions are yours. Put those three together across a month of your own queries and you get a defensible read on your own site, which is the only one that ever decides anything.

The wrong response to a number you cannot verify is a different number you also cannot verify. The right one is a control group on your own property.

Crownler stores every check as a row, keeps the thirty day line, and reads Search Console alongside it so the two halves sit in the same report. What the tool does · Questions, answered short · What generative engine optimization is, and what it is not