Clicks Are Stable and Clicks Collapsed, Both Are True

Google says click volume from Search remains broadly stable and the clicks it sends are higher quality. Independent researchers say click-through rates on affected queries have fallen by somewhere between a third and a half. Both camps have accused the other of bad methodology. Neither has moved.

The instinct is to pick a side, and that instinct is the mistake. The two claims are not in direct contradiction, because they are not measuring the same thing. Once you see what each one is actually counting, the dispute mostly dissolves, and what is left is a genuine and much more useful question about what you should be measuring in your own account.

The two numbers, and the populations underneath them

Google's position is an aggregate across all query types. Every search, informational and commercial and navigational and branded, summed. On that basis, total click volume can be broadly stable, because most searches never trigger an AI Overview at all.

The independent studies measure something narrower and deliberately so. They compare click-through rates on queries where an AI Overview actually appeared against queries where it did not. That is a subset, chosen precisely because it is where the effect would show up.

Now hold both at once. A steep decline concentrated in one query class can coexist with a stable total, provided that class is a minority of all searches. Which it is. That is not a fudge, it is arithmetic, and it means the honest framing is that these are measurements of different populations rather than a settled dispute in either direction.

What remains unresolved, and genuinely so, is magnitude. The independent studies do not agree with each other either.

The UK evidence, and Google's specific objection to it

The most-cited British measurement comes from Authoritas, a UK search measurement firm. Using a set of news keywords and phrases reflecting stories published by UK mainstream publishers between 16 and 22 April 2025, it found that when an AI Overview was present, publisher click-through rate dropped 47.5% on desktop and 37.7% on mobile. AI Overview penetration across that sampled keyword set averaged 12.2% during the week measured.

The study was submitted as part of a legal complaint to the Competition and Markets Authority about the impact of AI Overviews on news publishers. Authoritas also reported that a site previously ranked first could lose around 79% of its traffic for that query if the result appeared below an AI Overview.

Google's response was specific rather than general. A spokesperson said the study was inaccurate and based on flawed assumptions and analysis, that it used outdated estimations, and that its set of searches did not represent all the queries that would generate traffic for news websites.

Read the objection carefully, because part of it is fair. A one-week sample of mainstream news keywords is not a representative sample of all UK search. Authoritas did not claim it was, and the study's own conclusion was directional: it judged it highly likely that AI Overview penetration would increase across international markets, and noted that Google could return AI Overviews for more than 90% of queries if it chose to.

What the objection does not do is offer an alternative measurement. Google has been widely criticised for asserting relatively stable click volume and higher-quality clicks without publishing figures, charts or methodology, and for not allowing sites to analyse their own AI Overview data. Richard Kint of Digital Content Next called for auditable data on AI Overview click-through rates broken down by query type, content category and geography. That data did not exist at the time and largely still does not.

Pew, and why it is the hardest study to dismiss

Most of this evidence is rank-based, derived from tracking tools and inference. Pew Research Center did something different and harder.

It observed actual behaviour: 68,879 Google searches performed by 900 US adults in March 2025, published on 22 July 2025. Users who encountered an AI summary at the top of results clicked through to a traditional search result in 8% of visits. Users who did not encounter one clicked through in 15% of visits. And links cited inside the AI summary itself were clicked in only 1% of visits.

That last figure deserves attention because it complicates the optimistic reading of citation. Being cited is worth having, but the click rate on citations in that dataset was low. The UK behavioural evidence points slightly differently, with around 34% of respondents in the Association of Online Publishers study saying they would click at least one link shown, so this is not settled and the two are measuring stated intent versus observed behaviour, which are not the same thing.

Google disputed Pew as well, citing flawed methodology and a skewed query set, and repeating that it consistently directs billions of clicks to websites daily. The pattern of response is consistent enough to be worth naming: attack the sample, decline to publish an alternative.

The worst cases, which are real and are not averages

Some publisher-reported figures are far more severe than any of the studies, and they matter as much for how they get misused as for what they show.

DMG Media, which owns MailOnline and Metro, told the CMA that click-through rates dropped by as much as 89% when AI Overviews appeared for their content. The Guardian separately reported MailOnline experiencing a 56% drop in desktop clicks and 48% on mobile for affected articles.

Those are genuine measurements from real publishers with real analytics. They are also worst cases for specific query types, not averages, and quoting 89% as though it described a market-wide effect is the same category error as quoting Google's aggregate as though it described AI Overview queries. A range that runs from 15.49% to 89% depending on sample and query class is not a range you can average.

Source	Figure	What was measured, and where
Authoritas	-47.5% desktop, -37.7% mobile	Per-query publisher CTR, AI Overview present versus absent. UK news keywords, one week in April 2025. Google disputes the methodology.
Ahrefs	-34.5%	Top-ranking page CTR on informational keywords. 300,000 keywords, March 2025. Global and largely US.
Pew Research Center	8% versus 15%	Observed click behaviour. 900 US adults, 68,879 searches, March 2025. Citation links clicked in 1% of visits.
AOP and Ipsos	-18%	Click suppression when an AI Overview was present. 1,000 UK adults, nationally representative. The most useful UK behavioural figure available.
Seer Interactive	1.76% to 0.61%, then 2.4%	Organic CTR on AI Overview queries, June 2024 to September 2025, with partial recovery by February 2026. US and global.
SISTRIX	27% to 11%, a 59% relative loss	Position one organic CTR. 100 million German keywords. Germany, not the UK.
Amsive	-15.49%	Average CTR decline. Notably lower than most other figures, and rarely quoted for that reason.
DMG Media, to the CMA	up to -89%	Publisher-reported CTR for specific affected content. A worst case for particular query types, not an average.
Google	Broadly stable	Total click volume aggregated across all query types. No supporting figures, charts or methodology published.

Prevalence is a separate question, and a worse one

Alongside the click debate runs a second argument about how often AI Overviews appear at all. It is worth handling separately, because the numbers here are even less comparable.

For the UK: Authoritas measured 12.2% presence across its news keyword sample in April 2025. Ofcom's Online Nation reporting has been cited at roughly 30% of keyword searches. For global and US tracked query sets: BrightEdge has reported around 48% as of February 2026, Semrush peaked at approximately 24.61% in July 2025 and reported approximately 15.69% in November 2025, Pew's behavioural data found around 18% of searches, and Ahrefs found approximately 9.46% of desktop keywords in a 590 million search sample.

Those range from 9.46% to 48%. They cannot be placed on one trend line, because they use different query universes, devices, dates, geographies and feature-detection methods.

One more caution worth stating plainly, because it caught our own cross-validation. Figures of 47% on desktop and 53% on mobile for the UK circulate widely and are commonly attributed directly to Semrush and BrightEdge. Trace them and they resolve to a secondary aggregator citing those firms rather than to either firm's own published report. That does not make the underlying figures fabricated. It does mean anyone presenting them with a specific report title and publication month has upgraded a citation, and the resulting claim is weaker than it looks.

The practical conclusion: do not quote a UK prevalence figure. Measure prevalence on your own query set, where you control the sample and know what it contains.

What both sides actually agree on

Strip out the contested magnitudes and a surprising amount of consensus survives.

Every methodology tested points in the same direction on AI Overview queries. The magnitudes differ because the methods differ, and several commentators have made the point that this disagreement is collectively a point in the studies' favour rather than against them: independent researchers using different samples and different metrics arriving at different numbers but the same sign is what genuine measurement of a real effect looks like.

Both sides also agree that informational intent is where the effect concentrates and that commercial and transactional queries are far less affected. Google's high-quality clicks framing and the researchers' informational-decline framing are compatible descriptions of the same shift.

And Google's own dominance is not seriously contested by anyone. In its Strategic Market Status decision the Competition and Markets Authority found that AI assistants accounted for 0 to 5% of Google Search query volume when measured on search-grounded queries. Google holds roughly 91% of UK search across devices. The 2025 and 2026 story is coexistence and redistribution, not replacement.

How to hold this in a client conversation

If someone shows you a single figure for how much AI Overviews have cost you, ask three questions. Which query set. Which window. Which geography. Most single figures fail at least one of those.

Then move the conversation to the account. Prevalence on your own queries, impressions and clicks reported separately, branded and non-branded split, and query class segmented. Those four are knowable, and they are what a budget decision should rest on. Whether the market figure is 34.5% or 47.5% or 89% has no bearing on what your own commercial pages are doing.

The one thing worth saying to anyone still waiting for a definitive number: the reason it does not exist is not researcher incompetence. It is that the platform holding the data has chosen not to publish it, and the regulator has now set a date by which it must. That is a more honest account of the situation than either side's press release, and it is the account we work from when we build AI visibility measurement for UK clients, where the entire discipline rests on measuring your own citation footprint rather than inheriting somebody else's average.

---

Frequently Asked Questions

How can Google say clicks are stable while studies show they have fallen?

Because the two claims measure different populations. Google reports an aggregate across all query types, and most searches never trigger an AI Overview at all, so a total can hold steady. The independent studies deliberately compare click-through rates on queries where an AI Overview appeared against queries where it did not, which is a minority subset chosen precisely because that is where the effect appears. A steep decline concentrated in one query class is arithmetically compatible with a stable total. Treating this as a settled dispute in either direction misreads both.

What is the Authoritas study and why is it contested?

Authoritas is a UK search measurement firm that examined a set of news keywords and phrases reflecting stories published by UK mainstream publishers between 16 and 22 April 2025. It found publisher click-through rate dropping 47.5% on desktop and 37.7% on mobile when an AI Overview was present, with AI Overview penetration averaging 12.2% across that sample. It was submitted as part of a legal complaint to the Competition and Markets Authority. Google responded that the study was inaccurate and based on flawed assumptions and analysis, using outdated estimations and a set of searches unrepresentative of all queries generating traffic for news sites.

Which figure should a UK business actually use?

For a UK behavioural measurement, the 18% click suppression from the Ipsos component of the Association of Online Publishers study is the most defensible, because it is UK-specific with a stated nationally representative sample of 1,000 adults. For rank-based UK evidence, Authoritas is the only substantial option and comes with the caveats above. Most robust click-through research including Ahrefs, Pew and Seer is US or global and should be labelled as directional for Britain rather than presented as British data.

What did the Pew Research Center study find?

It observed behaviour rather than inferring from rank tracking, which makes it the hardest study to dismiss. Across 68,879 Google searches performed by 900 US adults in March 2025, published on 22 July 2025, users encountering an AI summary clicked a traditional search result in 8% of visits against 15% for those who did not. Links cited inside the AI summary itself were clicked in only 1% of visits. Google disputed this study too, citing flawed methodology and a skewed query set.

Does being cited inside an AI Overview actually generate clicks?

The evidence pulls in two directions and is not settled. Pew's observed behavioural data found links inside AI summaries clicked in only 1% of visits. The UK Association of Online Publishers study found around 34% of respondents saying they would click at least one link shown. Those are measuring different things, observed behaviour against stated intent, which are reliably different in survey research. The honest position is that citation is worth competing for but should not be modelled as equivalent to an organic click.

Is the 89% figure real?

Yes, and it is a worst case rather than an average. DMG Media, owner of MailOnline and Metro, told the CMA that click-through rates dropped by as much as 89% when AI Overviews appeared for their content, and The Guardian separately reported MailOnline seeing a 56% drop in desktop clicks and 48% on mobile for affected articles. These are genuine publisher analytics for specific query types. Quoting 89% as a market-wide effect is the same category error as quoting Google's aggregate as though it described AI Overview queries.

What is the AI Overview prevalence rate in the UK?

There is no defensible single figure. UK measurements include 12.2% from Authoritas across a one-week news keyword sample in April 2025 and roughly 30% attributed to Ofcom's Online Nation reporting. Global and US tracked query sets give around 48% from BrightEdge in February 2026, approximately 24.61% from Semrush at a July 2025 peak and 15.69% in November 2025, around 18% from Pew's behavioural data, and approximately 9.46% of desktop keywords from an Ahrefs sample of 590 million searches. The range from 9.46% to 48% reflects different query universes, devices, dates and detection methods, and cannot be placed on a single trend line.

Where do the 47% desktop and 53% mobile UK figures come from?

They circulate widely and are commonly attributed directly to Semrush and BrightEdge, but tracing them resolves to a secondary aggregator citing those firms rather than to either firm's own published report. That does not make the underlying figures fabricated, but anyone presenting them with a specific report title and publication month has upgraded a citation from an aggregator to a primary source, which makes the resulting claim weaker than it appears. We do not use those figures for this reason.

Why is the SISTRIX study not evidence about UK search?

Because it is German. SISTRIX analysed 100 million German keywords and found position one organic click-through rate falling from 27% without an AI Overview to 11% with one, a relative loss of 59%. It is a large and careful piece of research, and it describes a market with different search behaviour, a different publisher landscape and a different language. It appears in UK arguments frequently and should not, unless labelled clearly as directional evidence from another market.

Do the disagreements between studies mean none of them can be trusted?

The opposite argument is stronger. Independent researchers using different samples, different windows and different metrics arriving at different magnitudes but the same direction is what genuine measurement of a real effect looks like. If every study produced an identical number from different methods, that would be the suspicious outcome. Amsive's minus 15.49% and DMG Media's minus 89% mark the ends of a range that cannot be averaged, and every methodology tested points the same way on AI Overview queries.

When will there be auditable data to settle this?

A date exists. Under the conduct requirement following Google's designation with Strategic Market Status in general search, Google becomes obliged to publish disaggregated engagement metrics including impressions, click-throughs and click-through rate through Search Console. Main publisher controls arrive in December 2026 with page-level grounding controls in March 2027. The long-standing complaint, articulated by Digital Content Next among others, was that Google asserted stable and higher-quality clicks without publishing figures, charts or methodology while preventing sites from analysing their own AI Overview data. That complaint now has a regulatory answer with a timetable attached.

What should we measure in our own account instead?

Four things, all knowable. AI Overview prevalence across the queries you actually care about, checked by device and location. Impressions, clicks, click-through rate, conversions and revenue reported as separate measures rather than collapsed into traffic, since the diagnosis usually sits in the gap between held impressions and falling clicks. Branded and non-branded demand always split, because branded search volume is increasingly the measurable endpoint of AI visibility. And segmentation by query class, since informational intent carries the decline while commercial and transactional queries are far less affected.

---