The bounce-click defence is a measurement story, and Google is losing it

Liz Reid has now made the same argument three times in public. AI Overviews aren't taking your real traffic, she keeps saying — they're removing the "bounce clicks," the ones where someone landed on your page, grabbed a fact, and left within seconds. The clicks that disappeared were the ones that didn't matter anyway.

It's a tidy story. It's also, at this point, the only story Google has, and it's starting to wear thin.

This week the Indian School of Business and Carnegie Mellon published the first randomised controlled trial on AI Overviews. 1,065 US participants, three groups, a Chrome extension assigning conditions. When AI Overviews appeared, organic clicks dropped 38%. Zero-click sessions rose 33%. And — this is the part Reid has to answer for — removing AI Overviews didn't reduce user satisfaction, perceived quality, or ease of finding information.

If the lost clicks were genuinely low-value, removing the feature that eliminates them should have made users *less* satisfied. It didn't. The satisfaction needle didn't move.

That's a problem for the bounce-click defence, and it's a bigger problem for everyone trying to plan around AI search without real measurement infrastructure. Because what's actually happening here isn't a debate about user behaviour. It's a debate about whose data we're allowed to see.

The shape of the argument Google keeps making

Reid's framing is internally consistent. She says publishers were getting visits that weren't really visits — quick lookups, fact-grabs, the SEO equivalent of someone shouting a question through your letterbox and walking away before you opened the door. AI Overviews, in this telling, just save everyone time. The deep, engaged visits — the ones that produced sign-ups, sales, ad impressions worth billing for — those are still arriving.

It's a coherent story. The problem is it's also entirely unfalsifiable from outside Google.

Reid hasn't shared the data. Not the engagement-time distributions, not the bounce-rate breakdowns, not the conversion-event splits between AI-Overview-present and AI-Overview-absent SERPs. Three appearances making the same argument and zero supporting datasets. At some point the absence of evidence becomes its own evidence.

Meanwhile every external source pointing the other way is showing the same thing. Ahrefs documented in February that AI Overviews now correlate with a 58% reduction in CTR for top-ranking pages — nearly double the 34.5% drop measured a year earlier. Index Exchange found 69% of publishers experienced year-on-year ad opportunity declines through 2025, average drop 14%. The randomised trial puts a causal stake in the ground: 38% click reduction, no satisfaction trade-off.

Three independent lines of evidence pointing one way. One unfalsifiable narrative pointing the other. That's not a debate. That's a measurement standoff.

The earnings numbers don't settle it either

Alphabet's Q1 2026 was a barnstormer. $109.9 billion total revenue. Google Search alone hit $60.4 billion, up 19% year-on-year — actually accelerating from 17% growth in Q4 2025. Pichai told analysts queries are at an all-time high and that AI experiences are tied to increased Search usage.

You'll see this number wielded both ways in the next few weeks. The Google-side argument: search is healthier than ever, queries are growing, the doomers were wrong. The publisher-side argument: revenue is up because ad yield per query is up, which is consistent with fewer organic clicks reaching publisher pages.

Both can be true simultaneously. They almost certainly are.

The earnings data describes the ad business. It says nothing about whether your specific pages are getting more or fewer clicks from AI-influenced results. Higher Google ad revenue and lower publisher organic traffic are not contradictory — they're the predictable shape of a system that's compressing the journey from query to answer and capturing more of the monetisable surface inside the AI layer rather than handing it off to the open web.

So when Reid says queries are up and search is growing, she's right. When publishers say their organic clicks are down, they're also right. These are not the same conversation, and treating them as if they are is how the industry keeps talking past itself.

What the randomised trial actually changed

I want to be clear about why the ISB/Carnegie Mellon paper matters, because it's going to get cited a lot in the next few weeks and most people will quote the 38% figure without understanding what it represents.

The satisfaction finding is the one that quietly destroys the bounce-click defence.

Every prior study on AI Overview impact — Seer, Chartbeat, Pew, Ahrefs — has been observational or correlational. Researchers measured what happened on real SERPs and inferred causation. Useful work, but always vulnerable to the obvious counter-argument: maybe the queries that triggered AI Overviews were *also* the queries where users would have clicked less anyway. Maybe AI Overviews appear on informational queries, and informational queries always had lower CTRs. The correlation is real but the causation is contestable.

A randomised controlled trial removes that ambiguity. Same users, same queries, randomly assigned to conditions. If clicks drop in the AI Overview group and not the control group, you can say AI Overviews *caused* the click reduction. That's a different epistemological category than anything we had before.

The satisfaction finding is the one that quietly destroys the bounce-click defence. If the clicks being removed were genuinely low-value — quick fact-grabs that wasted the user's time — then removing AI Overviews should have made users feel like they were working harder, finding it more difficult, getting lower-quality results. None of those needles moved. Users were equally satisfied without AI Overviews. Which means the clicks that disappeared weren't a tax being lifted. They were just clicks.

That doesn't mean every lost click was a high-value visit. It means Reid's specific claim — that AI Overviews remove a measurably low-value class of traffic — is now contradicted by the first study designed to test it directly.

The deeper problem: nobody can audit Google but Google

Here's what nobody is saying loudly enough. The bounce-click argument is impossible to refute *or* confirm without Google's internal data. Reid could end this debate tomorrow by publishing engagement-time histograms by SERP feature. She hasn't. Three appearances, zero data.

That's not a small thing. It's the structural problem of this entire era. Google is making product decisions that materially affect the revenue of millions of publishers, and the only entity with the data to evaluate the impact is Google. The randomised trial cost researchers a Chrome extension and 1,065 participants. To do it at meaningful scale across query categories, languages, and verticals, you'd need access to query-level engagement data that only Google has.

We are, in other words, debating the impact of a system using fragments of information leaked through external proxies, while the entity with full visibility makes confident public claims and provides no supporting data.

This is the measurement problem the SEO industry has been trying to solve from the outside for two years. It's also why every "GEO platform" promising to solve it is selling something they can't deliver. You cannot reverse-engineer the bounce-click question from outside Google's logs. Nobody can.

What actually changes for the work

If you're running SEO for a UK business right now and you're trying to make decisions based on the bounce-click debate, stop. The debate isn't going to resolve in a useful timeframe and the data needed to resolve it isn't going to be made available.

What you can act on is what the trial — and the broader pattern across Ahrefs, Index Exchange, and the publisher revenue data — actually establishes:

Organic clicks for pages competing on informational queries are genuinely down. Not by a fatal margin in every vertical, but by enough that traffic-only KPIs are now lying to you. If your reporting still treats sessions as the primary success metric, you're measuring something that's drifting away from the underlying business reality. Conversion events, qualified enquiries, branded search volume, direct traffic — these are now the real signals.

Pages that exist purely to capture quick-answer informational queries are the most exposed. This isn't new advice but the trial confirms it empirically: the queries where AI Overviews dominate are precisely the queries where users were happy not to click. If your strategy is built on ranking for "what is X" and "how does Y work," that strategy is bleeding out and the bleeding is causal, not correlational.

Pages tied to task completion, proprietary data, transactional intent, or brand-led demand are far less exposed. This lines up with what Cyrus Shepard found in his 400-site analysis and what Rand Fishkin has been arguing for months. The sites surviving the AI search transition aren't the ones with the best content — they're the ones AI can't disintermediate. Different question, different answer.

Brand search is the metric that matters most and gets measured least. If branded queries to your domain are growing, you're winning regardless of what happens to informational organic. If they're flat or declining while non-branded organic falls, the strategic problem is bigger than AI Overviews and SEO tactics won't fix it.

The honest limits

The trial is one study with 1,065 participants in the US. It's the best evidence we have but it's not the final word. The findings could look different in other markets, on mobile, across query types we haven't isolated yet, or once AI Mode in Chrome rolls out properly and changes the baseline again. The 38% figure is a population-level effect — your specific site, in your specific vertical, may be hit harder or softer.

It's also possible Reid is right about *something*. Some of the lost clicks probably were low-engagement. The argument isn't that bounce clicks don't exist; it's that the data Google has provided doesn't show they're the dominant share of what disappeared. A more honest version of Reid's claim would be: "Some of the lost traffic was low-engagement, some wasn't, and the net publisher impact is mixed." That's defensible. The current version isn't.

And the earnings numbers genuinely do complicate the doomer case. Search revenue accelerating from 17% to 19% growth is not the shape of a dying product. Google has built something users are using more, not less. The user-facing product is succeeding. The publisher-facing externalities are the cost of that success — and Google is, predictably, downplaying the externalities while celebrating the user numbers.

Where this leaves us

The bounce-click defence is the most important piece of corporate communication in search right now because it's the framing Google needs to make stick. If "AI Overviews remove low-value clicks that didn't matter" becomes the consensus narrative, the regulatory and commercial pressure on Google to share data, change the product, or compensate publishers eases significantly. If "AI Overviews caused a 38% click reduction with no satisfaction gain" becomes consensus, the pressure intensifies.

Google has the data to settle it. Google won't share the data. The first randomised trial points the other way. Three external datasets point the same way as the trial. Publisher ad revenue points the same way as the trial.

That's the loop. And it's not closing in Google's favour.

What I'd take away if I were planning the next two quarters of SEO work for a UK business: assume the publisher-side numbers are real, assume the bounce-click defence is partial at best, stop measuring success in sessions, and put your effort into the categories of work — brand, task completion, proprietary value, transactional intent — that the trial *and* the earnings data *and* the publisher data all agree on. That's a rare alignment. Use it.

The measurement debate will continue. The strategic implications already don't depend on resolving it.