On August 28, SEO consultant Chris Long posted a screenshot on X that moved through the industry within hours: “Holy moly SEOs: Google is now revealing the AI Mode prompt box BY DEFAULT for some queries.” Google confirmed the change to Search Engine Roundtable and Search Engine Land the same day. For queries Google’s systems select, the AI Overview no longer waits for a click. It expands on load into a full, AI Mode-style answer with the “Ask anything” follow-up box already open, and the list of organic results drops down the page, on many screens below the fold.

Until that morning, the AI Overview loaded as a short snippet with a “Show more” button. That button was a gate, and the gate mattered: seeing the rest of the answer cost a click, and the first organic results stayed visible beside or beneath the snippet. The new design removes the gate. Barry Schwartz’s verdict at Search Engine Roundtable was blunt: “This is not good for publishers, not good at all.”

Google’s defense of the redesign rests on a single empirical claim. The only independent experiment built to test that class of claim has already run, and it published its results this spring. The two events landed eight weeks apart, but they describe the same tradeoff, and read together they produce something the industry has not had until now: a measured cost for the SERP Google is shipping.

What changed on August 28

Google framed the change as a targeted improvement. A spokesperson, identified by The Verge as Jennifer Kutz, said: “For some queries, AI Overviews may dynamically expand for topics where our systems determine it’s most useful for people.” Robby Stein, a product lead on Search, added on X that if a user has already started scrolling, “the expansion stops to keep your place.” Google would not say which queries qualify, what share of searches is affected, or whether the rollout will widen. Schwartz wrote at Search Engine Land: “If you thought AI Overviews take away clicks, just wait for this to roll out to your queries.”

The rollout is inconsistent so far. Jay Peters at The Verge saw auto-expanding overviews on most of his test searches, then watched them retreat to the snippet state, with one query (“what is the best game on switch 2”) triggering the expansion more reliably than others. A test that behaves this way is usually a staged rollout, and Google’s own language, “we’re always testing and tuning these experiences,” points the same direction. The question for anyone who depends on organic search is not whether this design costs clicks but how many, and that question now has an experimental answer.

The experiment that already ran

In June, Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon University’s Heinz College posted the first randomized field experiment on AI Overviews to SSRN (first posted April 3, 2026, revised June 17). Every estimate before it was observational: traffic before versus after the feature launched, or queries with an overview versus queries without. Those designs cannot rule out a third cause, an algorithm update, a shift in user habits, a competing product moving at the same time. Agarwal and Sen removed that problem by assigning the feature itself at random.

The team recruited US participants through Prolific and installed a custom browser extension in their Chrome browsers. Each user was randomly assigned to one of three arms: a control group that saw the standard results page, a treatment group in which the extension hid AI Overviews wherever they would have appeared, and a third arm redirected into Google’s AI Mode conversational interface. Over the study window the researchers logged 68,089 unique searches, and AI Overviews were triggered on roughly 41 percent of queries.

Random assignment means the two main groups differed in exactly one way: whether an overview that would otherwise have appeared was shown. Anything that moved between them can be attributed to the feature. That is the difference the word “causal” marks, and no prior study in this debate had it.

The measured cost

The headline number: showing an AI Overview cut outbound organic clicks by 39.8 percent and raised the probability that a search ends with no visit to any website by 34.5 percent. The April draft of the paper reported a 38 percent loss; the June revision, which added click-quality analysis, a treatment-switch check, and a query-type breakdown, settled at 39.8.

Spread across all searches rather than just affected ones, the drag is smaller but not trivial. With overviews triggered on roughly 41 percent of queries and each affected query losing about a quarter of a click, the arithmetic works out to roughly 0.10 fewer outbound clicks per search, around a fifth of the control group’s outbound click volume. The control group was looking at the collapsed, “Show more” design. The experiment measured the kinder version of the interface Google is now expanding.

Google’s answer to two years of traffic complaints has been that the clicks AI Overviews absorb are the shallow ones. Liz Reid, Vice President of Product for Search, characterized AI Overview clicks in July 2025 as higher-quality clicks with greater purchase intent, and she has said the feature reduces “bounce clicks” without releasing supporting data. Nick Fox, Vice President of Search, questioned the methodology of traffic-decline studies on a May 2025 podcast.

The experiment tested that argument on Google’s own terms, using three downstream measures: how often visitors hit the back button to return to results, how often a visit ended within ten seconds without further interaction, and total time on the destination page. About four in ten same-tab clicks led back to the results page in both conditions; roughly 18 percent of visits ended inside ten seconds; time on site was statistically indistinguishable. The extra clicks websites received when overviews were hidden looked, by every one of those measures, exactly like the clicks that already existed.

An endline survey completed by 90 percent of active participants found no difference between groups in overall satisfaction, perceived information quality, or ease of finding information, results the authors call precisely estimated nulls. Their conclusion in the revised paper: the finding is “at odds with the view that AIOs primarily eliminate low-engagement website visits.”

The June revision also closed the most common escape route for skeptics, the argument that something about the participants, not the feature, drove the result. After two weeks the researchers rotated the assignments. Users who had been seeing overviews and then lost them clicked more; users who got overviews back clicked less. The behavior reversed in both directions when the treatment switched, which is what a true cause looks like. It also means publishers should not expect searchers to adapt around the feature over time.

What the study says about the expansion

Read against the new design, the experiment’s most important data point is the one Google probably wishes had stayed exploratory. The study’s third arm sent users straight into AI Mode, the conversational interface whose behavior the expanded overview now mimics inside the main results page. Those participants averaged 0.36 external clicks per search, against 0.53 for the control group on the standard page, a 32 percent drop, and their satisfaction fell 1.14 points. The arm is exploratory because many assigned users did not comply or dropped out, so treat the magnitude with care. The direction is the point.

That sets up a bounding argument rather than a false precision one. The 39.8 percent loss was measured for a design in which the full answer sat behind a click and the organic results stayed visible. The expansion removes both protections. Nobody has randomized the expanded design yet, so the honest statement is: the measured figure is a floor, the one experimental arm that resembles the new design performs worse than the floor, and the survey evidence says users do not rate the experience better in exchange.

Contrast that with the justification Google actually gave. Kutz’s statement that “our research has shown that with this dynamic experience, users find Search more helpful and engage deeper in follow-up exploration” is not falsifiable in public: the company released no study, no sample size, no metric. The independent experiment measured precisely those outcomes, helpfulness, ease of finding information, downstream engagement, and found nulls for the collapsed design and a satisfaction decline for the AI Mode interface. Two research programs reached opposite conclusions, and only one of them has been published in a form anyone can check.

The number lands inside two years of converging evidence

The randomized estimate did not arrive alone. It sits inside the range that observational studies have been circling since 2025, which is what gives the whole literature its force: the methods differ, the samples differ, the definitions differ, and the direction never does.

The scale anchors are sobering. SISTRIX estimated that AI Overviews cost the German market alone 265 million organic clicks per month across 100 million keywords, by cutting the position-one click-through rate from 27 percent to 11. Seer Interactive’s panel of 42 organizations found organic CTR on overview queries falling 61 percent between June 2024 and September 2025, with paid rates down 68 percent over the same window, and even queries without an overview losing 41 percent of organic CTR. Index Exchange recorded a 14 percent year-over-year decline in ad opportunities across 69 percent of the publishers on its platform through 2025.

One data point inside the Seer study matters for what comes next: brands cited inside an overview earned 35 percent more organic clicks and 91 percent more paid clicks than brands left out. Being inside the answer is currently the only position with a measured upside.

The market has started to reprice accordingly. On July 1, 2025, Cloudflare began charging AI companies per crawl. On July 1, 2026, the same company declared that model insufficient and proposed paying publishers per citation instead, naming the traffic collapse and citing the Pew result as its evidence. From September 15, 2026, it will block training and agent crawlers by default on ad-bearing pages for new domains. An infrastructure company does not tear up its pricing model over a correlation. It does so when the causal evidence arrives.

Who is exposed, and what to do before the rollout widens

The June revision broke the effect down by query type, and the split is the targeting map for the expansion.

Combine the two findings in the revision and the exposure profile comes into focus: informational content, where overviews trigger on 53 percent of queries and the click loss concentrates, and positions one through three, which gained the most when the overview was removed, with position one nearly doubling. Run that logic in reverse and those are the placements that lose the most as the expanded design spreads. A site ranking first for how-to and explainer queries is sitting on the exact combination the experiment measured the largest effect for.

Practical moves while the rollout is still patchy:

• Pull a position-stratified CTR baseline now. Export Search Console CTR by query cluster and position band, split informational from navigational and transactional intent, before the expansion reaches your queries. When the change hits, you will want to separate it from update noise and seasonality, and you cannot reconstruct a baseline after the fact. Position bands one through three on informational clusters are the rows to watch.
• Stop reading flat impressions as safety. In the experiment, impressions of results did not change while clicks fell 39.8 percent on affected queries. The expanded design keeps your listing on the page and pushes it below the fold. Rankings reports and impression counts will look calm while referral traffic drops.
• Treat citation as the primary objective inside affected topics. Seer’s panel found brands cited inside overviews earning 35 percent more organic and 91 percent more paid clicks than brands excluded from them. Structure answer-first content, clear entities, quotable stat blocks, that a summary generator can lift with attribution. A separate measurement study of 55,393 queries published on arXiv in May 2026 found nearly 30 percent of overview-cited domains do not appear on the results page at all, so citation is not just a ranking game.
• Do not budget for user adaptation. The treatment-switch result showed behavior reversing immediately when the feature was shown or hidden. The loss persists for as long as the expanded answer is on the page; it does not decay as searchers “get used to it.”

Caveats, stated plainly: the study sampled US desktop users recruited through Prolific, so the magnitudes may not transfer to mobile-heavy markets, it is a working paper that has not completed peer review, and the AI Mode arm lost enough participants that its figures are directional at best. None of those limitations point toward a smaller effect. If anything, mobile screens, where the fold arrives sooner, give the expanded answer less room before it reaches the fold.

The missing number

Google declined to say what share of queries now auto-expands, and that refusal is the most important sentence in this story. Every input needed to price the damage exists except that one: the per-query loss is measured, the query types are mapped, the direction of travel is documented by the feature’s own expansion history, from 34.5 percent losses in April 2025 to 58 percent in February 2026 as the footprint grew. Agarwal and Sen wrote in their revision that aggregate traffic losses could grow if AI Overviews appear on more queries over time. The expansion announced on August 28 is that mechanism, running.

For two years the industry argued about AI Overviews with numbers it could not fully trust, because every number was a correlation with an available excuse. That era ended in June, when a randomized experiment put a causal floor under the debate: 39.8 percent of organic clicks, gone on affected queries, with nothing given back to the user in measured satisfaction. Five weeks later Google began removing the interface elements that kept the real-world loss near that floor. The next time the company tells the industry that users find the experience more helpful, the published answer is on SSRN.