Google AI Overviews cut website clicks about 40% in randomized study
A two-week experiment with 1,065 U.S. adults tested what changed when AI summaries disappeared. The findings sharpen a dispute over whether source links and publisher opt-outs are enough.
A randomized browser-extension experiment involving 1,065 U.S. desktop users found that Google searches with AI Overviews sent substantially fewer visitors to outside websites, while participants reported no improvement in search quality or ease. The result gives publishers experimental evidence of a traffic trade-off, but it covers two weeks of desktop use and does not establish effects in other countries or on other devices. Courtney C. Radsch argues that publisher-by-publisher opt-outs would not solve the issue and proposes opt-in rules and compensation; the study did not test those remedies.
01
AI Overviews appeared on about 41% of searches in the experiment, and links cited in the summaries received only about 7% of outbound clicks.
02
Traffic losses were concentrated on pages with an Overview and among the first three unpaid search results.
03
Ad clicks, clicks to Google-owned pages, search volume, time on visited sites, and visitors’ bounce rates did not change.
Google’s AI Overviews reduced clicks to outside websites by about 40% on searches where they appeared, without improving users’ reported search experience. The randomized study, described by Courtney C. Radsch in a November 20 Tech Policy Press article, puts experimental evidence behind publishers’ concern that AI answers take traffic without delivering better visits.
Changing the search page, not the searcher
Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon recruited 1,065 U.S. adults who used Google through desktop Chrome. Participants were paid to install a browser extension for two weeks in early 2026. They did not know the experiment’s purpose, according to Radsch, who served as a discussant at a presentation of the paper.
The extension let the researchers randomly assign people to different versions of Google. Rather than comparing people who happened to encounter an AI summary with those who did not, the experiment changed what participants saw:
One group had AI Overviews removed. The extension shifted the remaining results upward so the page retained a normal layout.
A second group was redirected to Google’s Gemini-powered AI Mode for every search, replacing ordinary results with that search experience.
The control group used Google without changes, giving the researchers a comparison against the versions altered by the extension.
Radsch describes the study as pre-registered, meaning its research plan was recorded in advance. She also reports two checks: searches without Overviews showed no effect, and switching the groups reversed the effect. The researchers checked for major Google algorithm changes during the experiment and found none.
Source links captured little of the traffic
AI Overviews appeared naturally on about 41% of searches in the experiment. The roughly 40% click reduction applies to searches where an Overview appeared, not all Google searches. Traffic losses were concentrated on pages with the summary at the top and among the first three ordinary, unpaid results.
Links cited within the summaries generated only about 7% of outbound clicks on those searches. Clicks on advertisements and Google-owned pages did not change, nor did the number of searches participants ran. The traffic effect therefore differed between outside websites and those other destinations.
For visitors who reached a website, time spent there and the rate of leaving without further engagement were unchanged. Users also reported no improvement in search quality or ease of finding information. Those findings concern the measured experience of U.S. desktop participants over two weeks; they do not establish that every user gets no benefit.
Removing an Overview is not a publisher opt-out
Radsch’s policy argument turns on a distinction the experiment did not directly test. Removing the entire Overview restores a page without that summary. Letting one publisher exclude its material does not necessarily remove the Overview itself. She argues that individual opt-outs can leave publishers competing for traffic beneath an answer assembled from other sources.
Radsch, who directs the Center for Media & Digital Governance at Open Markets Institute, recommends requiring publishers to opt in before their content is used in AI answers. She also advocates collective bargaining, payment for content use and disclosure of impressions, citations and clicks. These are her proposed remedies, not findings that the experiment tested.
Testing the effects independently carries its own cost. The researchers put the two-week experiment at about $30,000, excluding months spent building the extension, according to Radsch. She calls for replications outside the United States and across devices before treating this desktop sample as the answer for other markets.
Sources
techpolicy.pressNo Cure Without Evidence? AI Overviews Cut Clicks 40% and Give Users Nothing
Reader comments
Newest comments first. Replies stay oldest first.