A randomized field experiment involving 1,065 desktop Chrome users has produced what its authors describe as the first causal evidence that Google's AI Overviews divert clicks away from publisher websites without improving how users experience search. The paper, published on the Social Science Research Network on April 3, 2026, and last revised June 17, 2026, found that when an AI Overview was shown, outbound organic clicks fell by 39.8% and zero-click searches rose by 34.5%, while clicks on sponsored results stayed flat.

The study was written by Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon University's Heinz College. Unlike prior research on the topic, which relied on observational comparisons of search traffic before and after AI Overviews rolled out, this paper used a custom-built browser extension to randomly assign real users to see or not see the AI-generated summaries that Google places atop many search results pages. That randomization, the authors argue, lets them isolate the causal effect of AI Overviews rather than merely observing a correlation.

Data collection ran for two weeks per participant, beginning January 7, 2026, and concluding February 10, 2026. Recruitment through the online platform Prolific started January 7 and continued in stages through January 27. Participants who passed a series of eligibility checks, including using Google as their primary search engine and Chrome as their only browser, were split into three groups: a control group that saw the standard search results page, a treatment group in which the extension hid AI Overviews whenever they would otherwise appear, and a third group redirected to Google's AI Mode conversational interface for every query.

What the experiment measured

The two primary groups, control and "Hide AI Overview," were each assigned with a 36% probability, while the AI Mode group received the remaining 28%. Across the 396 users in the control group and 374 in the hide-AIO group, researchers observed 68,089 unique searches over the two-week period following each participant's enrollment. According to the paper, AI Overviews were triggered in approximately 41% of observed queries, a share the authors call "not marginal."

For queries where an AI Overview was intended to appear, removing it increased outbound organic clicks from 0.37 to 0.62 per search, a jump the researchers attribute entirely to the intervention because random assignment ruled out other explanations. The same removal cut the probability of a zero-click search, meaning a search that ends with no external website visit, from 0.73 to 0.54. Neither sponsored clicks nor the number of internal clicks within Google's own ecosystem moved in any statistically meaningful way, which the authors say indicates that AI Overviews substitute specifically for organic website visits rather than reshuffling attention across the page more broadly.

A placebo check strengthens that reading. For queries where no AI Overview was triggered at all, the researchers found no meaningful difference in clicks between the control and treatment groups, exactly what would be expected if the intervention worked as intended and had no spillover effect on unrelated searches.

Where on the page the effect concentrates

The paper reports that more than 87% of AI Overview appearances place the summary in the very top position of the results page, above all organic listings. When the researchers split queries by AI Overview position, they found the click effect was overwhelmingly concentrated in that top slot: an 88% relative increase in outbound clicks when the overview was removed from position zero, against no measurable effect when the overview appeared lower down the page.

Breaking clicks down further by the rank of the organic result clicked, the largest absolute increase after removing the AI Overview occurred at position one, with progressively smaller, though still statistically significant, effects at positions two and three and beyond. The pattern held across query categories too. Informational queries, which made up 71% of all searches in the sample, showed a large and statistically significant effect (a 0.26 increase in clicks per search), while navigational and transactional queries, smaller shares of the sample at 18% and 12% respectively, showed effects that were not statistically distinguishable from zero, though the authors note these estimates are noisier given smaller sample sizes.

No improvement in click quality or satisfaction

A recurring claim from Google is that the clicks users do make after seeing an AI Overview tend to be higher quality, reflecting more targeted intent. The study tested this directly by examining three measures on the downstream websites that users clicked through to: the probability of navigating back to the search results page via the browser's back button, a bounce measure defined as a session lasting less than ten seconds with no further interaction, and total time spent on the page.

None of the three measures differed meaningfully between the treatment and control groups. The paper states plainly that this finding "is at odds with the view that AIOs primarily eliminate low-engagement website visits." In other words, the additional clicks generated when AI Overviews were hidden looked, by these metrics, just as engaged as the clicks that existed already.

User-reported experience told a similar story. An endline survey, completed by 90% of participants who remained active for the full study, asked about overall satisfaction with Google Search, perceived quality of information, and ease of finding information, each measured on a five-point scale. Across all three dimensions, the difference between the control and hide-AI-Overview groups was statistically indistinguishable from zero, with the paper describing the results as "precisely estimated nulls." Removing AI Overviews, in other words, cost users nothing in terms of how they rated their search experience, even as it sent more traffic to outside websites.

The AI Mode arm showed steeper effects, and higher dropout

A third study arm redirected participants to Google's Gemini-powered AI Mode, a more conversational interface than the standard results page with AI Overviews layered on top. The researchers treat this arm as exploratory rather than causal because it experienced substantially higher attrition. Users could bypass AI Mode by appending a technical parameter to their search queries that forces a plain web-only interface, and the study tracked such workarounds as a form of attrition alongside outright extension uninstalls.

Among participants who did remain compliant, the paper found overall external clicks per search were 0.36 in AI Mode against 0.53 in the control group, a drop of about 0.17 clicks per search. The endline survey results for AI Mode users were markedly lower across every satisfaction dimension: down 1.14 points on overall satisfaction, 1.15 points on perceived information quality, and 0.87 points on ease of finding information, each relative to the control group average of roughly 4.0 on the five-point scale. The authors caution that these AI Mode figures are likely biased upward in terms of how negative the true effect might be, since users who disliked the experience most were also the ones most likely to have dropped out and gone uncounted.

Study design and data collection

The behavioral data underlying the paper came from a Chrome extension built on Webmunk, an open-source tool for studying online behavior developed with support from the National Bureau of Economic Research. The extension recorded URL visits, click behavior, scroll depth, and detailed information about each search results page, including whether an AI Overview appeared, its position, and the layout of organic and sponsored results, all while operating without any observable change to page load times. Mean page-load times were statistically indistinguishable across groups, at 2.02 seconds for the hide-AI-Overview condition against 1.97 seconds for control.

Recruitment funneled 3,710 people who started the initial screening survey down to a final analytic sample of 1,065. Along the way, 1,574 were screened out for not meeting eligibility criteria, 447 did not complete the survey or install the extension, and 581 failed browser-history validation checks requiring minimum recent activity on Google. A further 43 participants were excluded for data-integrity reasons, including 19 identified as workers on Google's own RaterHub search-quality evaluation program, whose search behavior the authors judged unrepresentative of ordinary consumers.

Balance tests across the three experimental groups showed no statistically significant differences in age, gender, race, prior trust in AI-generated information, or prior browsing activity, which the authors present as evidence that randomization worked as intended. Over 95% of users in both the control and hide-AI-Overview groups reported in the endline survey either not noticing any change to their search experience or being unsure whether anything had changed, a detail that supports the internal validity of the comparison since participants were largely unaware of which condition they had been assigned to.

The paper situates its findings within an active policy and legal fight over how AI-generated search summaries use publisher content. It notes that competition authorities in the United Kingdom have proposed that publishers be given the option to opt out of AI-generated summaries, and that complaints regarding AI Overviews have been filed with European and Brazilian antitrust regulators. PPC Land reported on June 30, 2025, that the Independent Publishers Alliance, Movement for an Open Web, and Foxglove filed a formal antitrust complaint with the European Commission and the UK Competition and Markets Authority, alleging that Google's use of web content for AI Overviews caused significant harm to publisher traffic, readership, and revenue.

That regulatory pressure escalated through the following year. The European Commission opened a formal antitrust investigation into Google's AI content practices on December 9, 2025, examining whether the company violated EU competition rules by using publisher and YouTube content for AI purposes without appropriate compensation or a viable opt-out mechanism. In the United States, Penske Media Corporation filed a 101-page federal antitrust lawsuit against Google on September 12, 2025, alleging the company coerces publishers into supplying content for AI systems without compensation while simultaneously reducing the website traffic those publishers depend on. Penske Media filed a further legal opposition on February 12, 2026, arguing Google had broken what the filing called the fundamental fair exchange that sustained the open internet, quoting a USA Today executive who stated that AI Overview traffic diversion is "just 100% false" to deny.

The paper also references separate concerns that shareholders of information platforms, including Reddit, have filed lawsuits alleging that Google's AI Overviews are diverting traffic that would otherwise reach those platforms, and that scholars have proposed the conduct could support a monopolization case against Google under the Sherman Act tied to its dominance in search.

How the findings compare with earlier industry research

The paper's headline effect sizes sit within, though toward the upper end of, a range that industry researchers using observational methods have reported over the prior two years. Ahrefs research published April 17, 2025, and covered by PPC Land, found that AI Overviews reduced organic clicks to top-ranking websites by 34.5% based on an analysis of 300,000 keywords, a figure that became one of the most widely cited data points in the publisher-traffic debate. A follow-up Ahrefs study published February 4, 2026, found the reduction had grown to 58% for top-ranking pages, nearly double the earlier figure, suggesting the effect intensified as AI Overviews expanded in coverage.

Other measurements have varied by methodology. Pew Research Center data from July 22, 2025, found users clicked on traditional search results roughly half as often when AI-generated summaries appeared, a finding Google publicly disputed as relying on "a flawed methodology and skewed queryset." Seer Interactive research published November 4, 2025, and analyzing 42 client organizations, found organic click-through rates for AI Overview queries fell 61% between June 2024 and September 2025, with paid click-through rates on the same queries falling 68% over the same period.

Because those prior studies compared periods before and after AI Overviews existed, or compared queries that triggered the feature against queries that did not, none could rule out that other simultaneous changes in search behavior, competing products, or Google's own algorithms were driving part of the observed decline. The Agarwal and Sen paper argues its randomized design addresses that gap directly, since the only systematic difference between its control and treatment groups was whether the extension displayed or suppressed an AI Overview that would otherwise have appeared.

The paper's finding of no change in downstream engagement quality also stands in some tension with public statements Google has made. PPC Land reported on July 5, 2025, that Google's Vice President of Product for Search, Liz Reid, characterized AI Overview clicks as "higher-quality clicks," arguing they demonstrate greater purchase intent and longer downstream engagement compared to traditional search clicks. Google's Vice President of Search Nick Fox separately questioned the methodology of traffic-decline studies during a May 21, 2025, podcast appearance, stating "from our point of view the web is thriving". The field experiment's bounce-rate, back-button, and time-on-site measures, none of which differed between groups, provide a data point directly relevant to that quality claim, since they suggest the incremental clicks generated by removing AI Overviews were not lower in engagement than the clicks that already occurred.

What the paper does not claim

The authors are careful to frame several results as suggestive rather than definitive. The AI Mode findings, given the differential dropout in that arm, are explicitly described as descriptive. The heterogeneity analysis by AI Overview position captures a correlation with observed placement rather than a causal effect of position itself, since Google's own systems, not random assignment, determine where an overview appears on any given page. The authors also note that total time spent on a webpage is a measure known to be affected by measurement error across the industry, a limitation they flag rather than paper over.

The study does not measure revenue impact on publishers directly, nor does it estimate how the 41% AI Overview trigger rate observed in the sample might change as Google expands the feature's coverage over time. The paper's conclusion states that as that trigger rate rises, "their aggregate impact on publisher traffic is likely to be substantial," but this is presented as an inference from the measured per-query effect rather than as a separately estimated quantity.

Why this matters for marketers

For search marketers, SEO teams, and publishers, the study adds a methodologically distinct data point to a debate that has largely relied on before-and-after comparisons or self-reported traffic logs. Randomization is a different standard of evidence than the correlational studies that have dominated coverage of this topic, and it arrives at a moment when regulators in multiple jurisdictions are actively weighing whether to require opt-out mechanisms or compensation frameworks for AI-generated search summaries.

The finding that click quality did not differ between groups is particularly relevant to any advertiser or publisher currently deciding whether to treat AI Overview-adjacent traffic as a strategic asset or as a shrinking pool of visits worth defending. If, as this paper suggests, the additional clicks a publisher would receive without AI Overviews are no less engaged than the clicks it currently gets, that undercuts arguments that a smaller volume of AI-filtered traffic compensates for its reduced size through higher value per visit. Marketing teams building content strategies increasingly oriented around zero-click visibility, brand mentions, and citation within AI summaries rather than direct clicks may find in this research a more rigorous basis for those planning assumptions than the observational data that has so far shaped the conversation.

Timeline

  • January 7, 2026: Recruitment for the study begins on Prolific; data collection starts.
  • January 7-27, 2026: Participant recruitment proceeds in staggered stages through pre-screening and baseline surveys.
  • February 10, 2026: Data collection concludes after two weeks of browsing observation per participant.
  • April 3, 2026: The paper "The Impact of Google AI Overviews on Publisher Traffic and User Experience: Evidence from a Field Experiment" is first posted to SSRN.
  • June 17, 2026: The paper is last revised, reflecting the version reviewed for this article.

Summary

Who: Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon University's Heinz College conducted the research, working with 1,065 US-based Prolific participants recruited for the study.

What: A randomized field experiment using a custom Chrome extension found that Google's AI Overviews reduce outbound organic clicks by 39.8% and increase zero-click searches by 34.5% when the feature appears, without improving user-reported satisfaction, perceived information quality, or engagement on downstream websites.

When: Data collection ran from January 7 to February 10, 2026. The paper was first posted to SSRN on April 3, 2026, and last revised on June 17, 2026.

Where: The study observed desktop Chrome users in the United States, capturing behavior on Google's standard search results page as well as its AI Mode conversational interface.

Why: The findings matter because they provide randomized, causal evidence in a policy debate that has so far relied largely on observational data, arriving as the European Commission, the UK Competition and Markets Authority, and multiple US courts weigh whether Google must offer publishers compensation or opt-out mechanisms for AI-generated search summaries built from their content.