The AI Overviews click studies do not agree

Ten measurements, one question, an 84 point spread

Seer Interactive reports a 65.2% organic CTR decline when an AI Overview is present and your page is not cited. Amsive reports an 18.68% CTR increase on branded keywords that trigger AI Overviews.

Both were measured on real Google Search Console data. Both are correct. That is an 84 point spread on what looks like a single question, and it is why a client sent you two links this week and asked which one to believe.

The correct answer is that the question was underspecified. I have read every write up I can find on this topic, and they all do the same thing: they cover one study, quote its headline, and treat the others as noise. The one aggregation I found that lists twelve of them, IDEAVA's roundup, says the magnitude varies "depending on methodology, query type, and measurement window" and then never connects a single methodological choice to a single number.

That connection is the whole job. Here it is.

The reconciliation table

Every row below comes from a page I fetched. The last column is the part that matters, because it is the one nobody publishes.

Read the table by column four rather than column five and the disagreement stops existing. Inside each design the numbers are consistent. Across designs they were never talking to each other.

Use this graphic on your site

Free to republish with a link back to this page. Copy the embed code:

<a href="https://josephtimpson.com/insights/ai-overviews-click-through-rate"><img src="https://josephtimpson.com/assets/infographics/ai-overviews-click-through-rate.svg" alt="Bar chart of reported AI Overview click-through rate changes from ten published measurements, grouped by measurement design" width="1200" style="max-width:100%;height:auto"></a><p>Graphic by <a href="https://josephtimpson.com/insights/ai-overviews-click-through-rate">Joseph Timpson</a></p>

The four variables that produce the entire spread

Every number above is a function of four choices. Change any one and the headline moves by tens of points.

The measured quantity. Position-one CTR, portfolio CTR, per-visit click rate and zero-click share are four different metrics. Pew's 8% is not a CTR at all. It is the proportion of browsing sessions in which the user clicked anything. A page can hold its CTR perfectly while that number collapses, because the two have different denominators.

The query population. Ahrefs restricted to informational keywords for a stated reason: 99.2% of keywords that trigger AI Overviews are informational in intent. Authoritas restricted to UK news terms, which is the single most AI-summarised query class on the internet. Amsive restricted to keywords the client already ranked for. Three restrictions, three different worlds. This is the same sampling problem I write about in how a local AI visibility study should be built, and it is not solved by adding more keywords.

The comparison design. Ahrefs compared two cohorts of different keywords. Amsive and Semrush compared the same keywords to themselves before and after. Cohort designs will always report larger effects, because the with-AI-Overview cohort was selected by Google on the basis of query characteristics that already predicted low click rates. Same-keyword designs are cleaner and consistently produce smaller numbers. That single distinction explains most of the gap between the 58% and the 15.49%.

The window and the baseline. Ahrefs ran the identical method twice and got 34.5% and then 58%. Nothing about the method improved. The window moved. Any number in this field is a snapshot of a system that Semrush measured swinging from a 6.49% AI Overview trigger rate in January 2025 to 24.61% in July and back to 15.69% by November.

The control group nobody quotes

Seer's study contains the most important number in this entire literature and almost nobody cites it.

On the same 3,119 query set, across the same 42 organizations, over the same period, queries where no AI Overview appeared at all saw organic CTR fall 46.2% year over year.

Sit with that. The control group lost nearly half its clickthrough rate with no AI Overview involved. Against a 46.2% baseline decline, the 65.2% figure for uncited pages represents an incremental AI Overview effect of roughly 19 points, not 65.

This is the number that should end most client conversations, and it is the number that gets cut from every roundup because it makes the headline smaller. Search result pages have been getting denser, more commercial and more feature-heavy for a decade. AI Overviews are an accelerant on a fire that was already burning. If you have ever told a client that AI Overviews took their traffic, you owe them this control group.

The mechanism nobody names: the denominator moved

CTR is a ratio. Practitioners argue about it as though only the numerator can change. In 2026 the denominator is what changed.

Here's what we know for certain: clicks were essentially flat in October, moving from 398K to 400K. Impressions more than doubled, from 15.8M to 33.1M. CTR compressed because the denominator grew far faster than the numerator.

That is one month of one dataset, and it reframes the whole argument. Clicks were flat. Impressions doubled. CTR halved. A practitioner reporting "CTR down 50%" that month would have been factually right and strategically wrong, because zero clicks were lost.

Seer Interactive's 2026 update, whose lead author is McDonald, states its own limit plainly: "We cannot claim causation. Higher-authority brands are also more likely to be cited." That sentence is more useful than most of the headlines built on top of the study.

The practical consequence is that CTR is now a broken primary KPI. If impressions and clicks can move in opposite directions at that scale, you have to report both, separately, or your client reporting will tell a story the revenue does not support. Seer's own paid media team found the same shape on the ad side: nonbrand impressions up 28% while CTR fell from 16.39% to 14.13%.

The two findings that went the other way

Two results in the table point in the opposite direction from the consensus. Neither is a fluke and both are load-bearing.

Semrush found zero-click going down. Taking keywords that had no AI Overview in May 2025 and did have one by October, zero-click fell from 33.75% to 31.53%. This is the cleanest natural experiment anyone has published, and the authors offer the explanation themselves: maybe these searches were already unlikely to end in a click. That is the selection effect stated out loud. Google adds AI Overviews to queries that were already answer-shaped. The queries were low-click before the summary arrived.

Amsive found branded keywords gaining.

Interestingly, branded keywords that triggered AI Overviews saw a CTR increase of +18.68% on average.

Only 4.79% of branded keywords triggered an AI Overview in that dataset, so the sample is thin. But the direction is consistent with everything else we know. A user searching your brand name has already decided. A summary of your brand confirms the decision and the click follows. That is why brand mentions matter more than backlinks for AI visibility, and why a portfolio that is 60% branded will see a completely different AI Overview story than one that is 90% non-branded.

Before you quote any figure in this field to a client, check the brand and non-brand split of the study and the brand and non-brand split of the client. If they differ, the number does not transfer.

Google's counter-claim belongs in the table, not above it

Google disputes all of this. That position deserves a row, and it deserves to be judged by the same standard as every other row.

Overall, total organic click volume from Google Search to websites has been relatively stable year-over-year.

In the same post Reid writes that "This data is in contrast to third-party reports that inaccurately suggest dramatic declines in aggregate traffic". She may well be right. Google is the only party with the clean dataset, and every study in the table above is an outside reconstruction of something Google can simply read.

But Google published no sample size, no date range, no methodology and no data. On the columns that let you evaluate a study, Google's row is the emptiest one in the table. When Authoritas published its UK news figures, a Google spokesperson told The Guardian the study was based on flawed assumptions, and again showed nothing.

The asymmetry is the argument. Nine independent measurements with published methods point one way. One unsourced corporate assertion points the other. That is not a tie, and it is not a reason to dismiss Google either. It is a reason to hold every number loosely, including the ones you like.

Google's own instrument still cannot answer the question

Here is the detail that should end the argument about whose data is better.

As of February 2026, Ahrefs noted that "there is still no way to disambiguate AI Overview clicks and impressions from the rest of your Search Console data."

It seems that Google doesn't want us to see the clickthrough rate for AI Overviews.

Google then shipped a generative AI performance report in Search Console in June 2026. Read the official documentation and the report covers impressions only. Google's own help page defines the metric as how many times links to your site were shown in a generative AI feature, and it is rolling out to a subset of website owners. No clicks. No CTR. No queries.

So the single organization capable of settling this has built a reporting surface that reports the numerator's denominator and nothing else. Until that changes, everything in this field is inference, which is exactly why AI search attribution has to be built from your own logs and your own lead data rather than bought from a dashboard. It is also why the tracking tools all disagree with each other.

What you can defensibly tell a client

Strip out everything that does not survive the table and you are left with a short list. This is what I actually say.

What you should not say: any single percentage, without naming the study, the query population and the design. "AI Overviews cut clicks 58%" is a sentence about 150,000 informational keywords measured against a non-matched cohort on desktop between two Decembers. Said without those qualifiers it is not a finding, it is a vibe.

Rand Fishkin's framing is the one I keep coming back to, because it puts AI Overviews in proportion rather than at the center:

Ten years ago, ~45% of Google searches were zero-click. Today it's 68%.

Twenty-three points of that shift accumulated over a decade, most of it before AI Overviews shipped. Fishkin also notes his own data limitation, that "it's likely the percentage of zero-click searches is even higher than what we've found here," which is more methodological candour than Google has offered on this subject in two years.

What nobody has measured yet

Four gaps, stated plainly, because the honest version of this post has to include what is missing.

Nobody has run a matched same-keyword cohort at scale. Amsive and Semrush did same-keyword designs on modest samples. Ahrefs did a large sample with an unmatched design. The study that does both does not exist, and it is the only one that would settle the causal question.

Nobody has published revenue. Every measurement in this field counts clicks, impressions or citations. Not one counts closed business. If you want an outcome claim, it has to come from your own client data, which is the reason an AI visibility audit starts with the lead ledger rather than a rank tracker.

Nobody has isolated AI Mode. Every study above measures AI Overviews inside classic Search. AI Mode is a different surface with different retrieval behavior and its own referral profile, and no independent CTR study of it exists.

Nobody has confirmed the recovery. Seer's rebound from a 1.3% floor in December 2025 to 2.4% in February 2026, reported by Search Engine Land as an early sign of recovery, is the most encouraging data point in the field and it is two months long, from one portfolio, published by the same firm that produced the most alarming figure. It needs replication before anyone builds a strategy on it. Treat it the way you should treat any small-sample AI visibility signal.

The field will keep producing these studies, and each one will get a headline that strips the method. Keep the table. When the next number lands, put it in a row, fill in columns three and four, and see whether it is telling you something new or telling you the same thing from a different angle.

Most of the time it will be the second one. That is the actual finding, and it is why I spend far more of a client's budget on crawler access than on chasing a metric Google will not let anyone see. More published reconciliations live in the insights archive, and if you want the version applied to a real site, that is what a working session covers.