Do Google AI Overviews actually cause click loss? The first randomized experiment

Yes. The first randomized experiment shows AI Overviews cause the click loss, not just accompany it. In a 2026 field experiment by Saharsh Agarwal (Indian School of Business) and Ananya Sen (Carnegie Mellon Heinz), an AI Overview was hidden at random for some searches. Showing it cut outbound organic clicks by about 39.8% (0.62 per search without the overview, 0.37 with it) and pushed the share of no-click searches from 0.54 to 0.73. Crucially, the lost clicks were not lower quality, and satisfaction did not improve.

This is a dated reference, not a new dataset. Last reviewed: 7 August 2026. The figures come from Agarwal and Sen's working paper, first posted to the Social Science Research Network on 3 April 2026 and last revised 17 June 2026; it is not yet peer-reviewed. It matters because almost every earlier click-loss number is associational. Our own AI Overview click-loss reference reports Ahrefs figures of roughly 34.5% to 58% and hedges them as associations, not proof of cause. This experiment supplies the causal complement, and its ~40% lands inside that same band, so it sharpens the existing read rather than contradicting it.

What did the experiment actually measure?

The causal question, using random assignment. Instead of counting clicks on pages that happen to sit under an AI Overview, the researchers built a browser extension and randomly decided, per user, whether the overview appeared. That random switch is what turns a correlation into causation: the groups are alike except for the one thing being tested.

Source: Agarwal and Sen, 2026 (SSRN working paper). The design isolates the effect of the overview appearing from everything else about a query, which a raw "clicks are lower where overviews show" count cannot do.

How much did AI Overviews cut clicks?

By about 39.8%, concentrated on informational queries. Removing the overview lifted outbound organic clicks and lowered the no-click rate; both moves reversed when it reappeared.

Source: Agarwal and Sen, 2026 (SSRN working paper). The overview triggered on roughly 41% of all queries and 53% of informational ones, and the losses concentrated there, so the effect on any one site scales with how informational its query mix is. This is the zero-click pattern the featured snippet began and generative answers deepened, and it is the mechanism behind the great decoupling of impressions from clicks.

Were the lost clicks just low-value traffic?

No, and this is the finding that changes the argument. A common industry reassurance is that AI Overviews only skim off low-intent visits nobody wanted. The experiment tested that directly and found no support for it.

• Engagement was unchanged. Bounce rate, time on site, and how often people bounced back to the results page did not differ meaningfully between groups.
• Satisfaction did not improve. An endline survey found no difference in overall satisfaction, perceived information quality, or ease of finding information when overviews were present.
• Ads were untouched. Sponsored clicks did not move, and neither did clicks within Google's own properties; the displacement fell specifically on outbound organic clicks to the open web.

The clicks an AI Overview removes look like ordinary clicks, not junk. Publishers lose real visits, users are no more satisfied, and only sponsored and Google-owned destinations are spared.

Why does a randomized experiment matter more than the usual data?

Because it breaks the tie that association studies cannot. Most AI-visibility claims rest on comparing pages or queries that differ in many ways at once, so any single trait gets unfair credit. Random assignment holds everything else roughly constant, which is the same logic behind reading GEO research critically in does GEO actually work and behind testing changes with a holdout rather than a bare before-and-after, the method in how to test whether a change moved your citations. It is a single study on US desktop Chrome, so it is not the last word, but it is the first causal word on a question that association could only circle.

What are the study's limits?

Real ones, which is why the honest claim is "strong causal signal," not "settled law." Read the numbers with these in view.

Source: Agarwal and Sen, 2026 (SSRN working paper), with limitations as described by the authors and coverage. None of these rescue the "it's only junk traffic" reading; they bound the figure, not the finding.

What should you do with this?

Treat AI Overview click loss as real and causal, then change the scoreboard rather than mourn the clicks. Concretely:

• Stop waiting for proof of cause. You now have it; plan for a structural drop in outbound clicks on informational queries, and size it to your own query mix using our estimation method.
• Measure citations, not just sessions. If traffic is your only metric, AI search reads as pure loss even while your brand is shown and cited; track presence and citation coverage too, per from clicks to citations.
• Compete to be the cited source. You cannot opt out of the overview appearing, but being named inside it is the winnable game, and it is not the same surface as AI Mode, which behaves differently again.

The click loss is now causally established; the number that governs your strategy is whether your brand is still being shown, cited, and recommended inside those answers, measured across engines and over time. That answer-level, cross-engine visibility measurement is exactly what Buffy Intel is built to provide. Questions: [email protected].