Forty Percent Fewer Clicks, on Forty-One Percent of Searches A preregistered randomized experiment by Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon found that Google AI Overviews cut outbound organic clicks by 39.8 percent on searches that trigger an Overview, according to a working paper dated 17 June 2026. The study, which ran from 7 January to 10 February 2026 with 1,065 US participants using a custom Chrome extension that hid AI Overviews in real time, measured 0.37 outbound organic clicks per search with Overviews versus 0.62 without, and found the share of searches ending with no outbound click rose from 54 to 73 percent. The 39.8 percent figure applies only to the roughly 41 percent of queries in the sample that showed an Overview, and the authors found no measurable difference in click quality or user satisfaction between groups. The headline going around this week is that Google’s AI answers cut search traffic to websites by 40 percent. It comes from a blog post by RevScope, an analytics company, dated 9 November, which pulls together an experiment, a large SEO dataset, Pew Research data and Google’s own statements. It’s a careful post, and it says plainly which of its numbers are causal and which aren’t. But the 40 percent in the headline is a conditional number, and leaving out the condition changes what it means for anyone who runs a website. The number comes from a working paper by Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon. I read the version dated 17 June 2026. It’s the best evidence available on the question, and it’s worth reading for what it measured. The experiment Agarwal and Sen recruited US participants on Prolific and had them install a custom Chrome extension on their desktop computers. Recruitment ran in stages from 7 January, and the study ended on 10 February 2026. 1,065 participants passed the browser checks. The extension randomly assigned each one to one of three arms. The control group saw Google as usual. The “Hide AIO” group got an extension that removed AI Overviews in real time and moved the remaining results up to fill the gap, so the page didn’t look altered. A third group was redirected to Google’s AI Mode. That arm lost so many participants that the authors treat it as exploratory, and I leave it aside here. The study was preregistered in the AEA RCT Registry AEARCTR-0017393 . The design gets around the problem every observational study has. Searches that trigger an Overview are different from searches that don’t: Pew found they tend to be longer and more often phrased as questions or full sentences. Comparing the two kinds of search mixes the effect of the Overview with the effect of the query. Here, both groups searched the same way, and the only difference was whether the Overview was shown. The paper checks that the share of queries that would have triggered one was similar across the two groups, so hiding the summaries didn’t change what people searched for. For queries where an Overview was triggered, or would have been, users who couldn’t see it made 0.62 outbound organic clicks per search. Users who could made 0.37. The paper reports this as a 39.8 percent reduction. That’s a regression estimate with user fixed effects, which is why it doesn’t exactly match the 40.3 percent you get by dividing the two means. The share of searches ending with no outbound click went from 54 to 73 percent, reported as a 34.5 percent increase. On queries that never trigger an Overview, the two groups behaved the same, and the authors use that as a placebo check. Sponsored clicks didn’t change, though at about 0.02 per search they’re too rare to rule out a small effect. People in both groups made the same number of searches, so the fewer clicks per search add up to fewer visits. Two other results matter as much. Clicks from the group without Overviews were no lower in quality: the chance of coming back to Google with the back button, the bounce rate and the time spent on the site were all statistically the same across groups. That bears on Google’s argument, which I come back to below. And the post-experiment survey found no measurable difference in satisfaction with Google, perceived quality of information or ease of finding it. Hiding the summaries didn’t make search feel worse. It’s also useful to read the effect the other way around. A drop from 0.62 to 0.37 is a 40 percent loss. Going from 0.37 to 0.62 is a 68 percent gain. Both describe the same gap. A publisher deciding whether to opt out of Overviews, if that ever becomes possible, should probably think in the second form. The condition in the 40 percent The 39.8 percent applies only to searches that show an Overview. In Agarwal and Sen’s sample, that was about 41 percent of queries. Their footnote does the conversion: across all searches, the effect “translate s into an average reduction in outbound organic clicks of roughly 18.5%.” That’s the number closer to what a site would see in its analytics, if its traffic came from a mix of queries like the one these participants produced. So the share of searches showing an Overview decides most of what the headline means, and the estimates of that share don’t agree. Pew logged the browsing of 900 US adults and found that 18 percent of their Google searches in March 2025 produced an AI summary. Agarwal and Sen measured about 41 percent in January and February 2026, also from real users’ searches. DataForSEO’s monthly keyword panel puts it at 61.43 percent, up from 5.93 percent two years earlier. Agarwal and Sen cite Aral and colleagues at 67 percent and attribute the gap with their own number to query composition: a curated set versus “organically generated user queries in a natural browsing environment.” The four figures measure different things. A keyword panel counts each tracked keyword once, whether a million people search it or ten. A browsing log counts what people actually type, weighted by how often they type it. Some of the rise from 18 to 41 percent is probably real growth over ten months, and some is the difference between two samples. The 61 percent is a share of keywords, not of searches. Still, the direction is clear, and the authors make the point themselves: as Overviews spread, “the overall reduction in outbound traffic could become substantially larger, even if the per-query effect remains constant.” If I scale their 18.5 percent linearly from 41 percent coverage to DataForSEO’s 61 percent, I get about 28 percent. That’s crude. It assumes the newly covered queries draw clicks the way the old ones did, and that a keyword share stands in for a search share, and neither is safe. But it shows why the per-search number and the sitewide number shouldn’t be confused, and why the gap between them is closing. What observational data said first Before the experiment, the best-known evidence was Pew’s. In its July 2025 analysis, users who saw an AI summary clicked a traditional result on 8 percent of visits, against 15 percent for users who didn’t. They clicked a link inside the summary on just 1 percent of visits. They ended their browsing session on 26 percent of pages with a summary, against 16 percent of pages without one. Pew’s click gap is larger in relative terms: 8 against 15 is nearly a halving, while the experiment found about two in five. The metrics differ, so the comparison is loose. Visits with a click aren’t the same as clicks per search. But the direction is what you’d expect if selection inflates the observational gap. Pew compared different searches, and the searches that draw summaries are the ones where users were more likely to be satisfied on the page anyway. The experiment shrinks that gap without closing it: with the query held fixed, the summaries still cost clicks. Google’s August 2025 response to these reports, from Search head Liz Reid, was that “total organic click volume from Google Search to websites has been relatively stable year-over-year,” and that Google was sending “slightly more quality clicks,” meaning clicks where users don’t quickly come back. The post gave no figures. The experiment can’t test the total, since it covers a few weeks and doesn’t follow the market. It can test the per-query claim, and on the same queries the summaries reduced clicks. On quality, it found no difference in returns to the search page, bounce or time on site between the two groups. Once someone clicked, the visit looked the same in both groups. Zero clicks were never all bad Information retrieval researchers have studied queries that end without a click for a long time, and that work changes how the 73 percent should be read. In 2009, Jane Li, Scott Huffman and Akihito Tokuda, at Google, defined “good abandonment”: a query whose need “was successfully addressed by the search results page, with no need to click on a result or refine the query.” They hand-labeled samples of abandoned queries on PC and mobile in the US, China and Japan, deliberately estimating an upper bound. On PC, queries that could in principle be answered on the results page made up 31.8 percent of abandoned queries in the US, 23.3 percent in China and 19.0 percent in Japan. On mobile, the shares were 54.8, 49.8 and 32.3 percent. That’s the baseline I’d compare the 54 percent zero-click rate against. Even with the Overviews hidden, more than half of these searches ended without an outbound click. Weather, addresses, definitions, a quick fact: the results page had been answering those for years. The Overview moves the line further, and the question for publishers is which searches it moves. The paper’s split by query type helps. Informational queries were 71 percent of searches and triggered an Overview 53 percent of the time. Navigational queries triggered one 15 percent of the time and transactional ones 6 percent. The effect is significant for informational queries. The other two samples are too small to show one. The traffic being lost is mostly people looking something up, which is the traffic informational publishers live on. The 2009 paper also bears on the experiment’s one big limit. It’s desktop-only, and Li and colleagues found the potential good-abandonment rate significantly higher on mobile in all three countries. They attributed this partly to how awkward it was to open pages on a phone. Mobile searchers already stop on the results page more often. The per-search effect of an Overview on mobile could be smaller, because there were fewer clicks to take, or larger, because a summary on a small screen pushes results further down. This experiment can’t say, and I wouldn’t apply 39.8 percent to mobile traffic without a mobile test. Why the satisfaction result is the surprising one Most people expected the click result. The survey result runs against earlier evidence. In a 2020 CIKM lab study, Zhijing Wu, Mark Sanderson, B. Barla Cambazoglu, W. Bruce Croft and Falk Scholer showed 30 participants results pages with no answer, one direct answer or several, and tracked their eyes. Users “read fewer search result snippets and paid less attention to snippets when a direct answer is shown,” and the authors concluded that the answer module “reduces user effort, and promotes user satisfaction during the search process.” That’s the mechanism behind lost clicks: attention moves to the top of the page and stays there. Agarwal and Sen find the same thing at the level of position. The extra clicks without Overviews went mostly to results ranked one to three. The difference is in satisfaction. In a lab, with prepared questions and an eye tracker, answers raised it. In two weeks of people’s own searches, hiding the answers didn’t measurably lower it. I don’t think that settles whether Overviews help users. A post-study survey is a coarse instrument, and a null result from about a thousand people doesn’t rule out a small benefit. But it does weaken the strongest defense of the trade, which is that users get enough out of the summary to justify the traffic it costs publishers. There’s also a reason to be careful about reading clicks as welfare in either direction. Filip Radlinski, Madhu Kurup and Thorsten Joachims tested this directly in 2008, on the arXiv search engine, using ranking functions whose relative quality they knew by construction. None of the eight absolute usage metrics they tried, including number of clicks and abandonment, “reliably reflect retrieval quality for the sample sizes we consider.” Paired designs that interleaved two rankings did. Lost clicks are a direct loss to a publisher. For the user, a click that doesn’t happen might be a need that was met or a need that was given up on, and the count alone doesn’t tell you which. The experiment’s survey is an attempt to look past the count, and it found nothing either way. What I’d take from it The 40 percent is real and well identified, at least for US desktop users in early 2026. It’s the effect on a search where the Overview appears. For a site’s total Google traffic, the paper’s own figure is about 18.5 percent, at the 41 percent coverage it observed. That figure will keep rising as coverage does, without the per-search effect changing at all. If you run a site and want to estimate your own exposure, the per-search number isn’t the first thing to find. Find out what share of the queries that bring you traffic now show an Overview. A health site and a restaurant site will get very different answers: by RevScope’s reading of the DataForSEO report, Overview presence is 69.68 percent for health keywords and 29.37 percent for dining. Multiply your share by something like 40 percent and you have a rough first estimate of what you’re losing on desktop. The other number worth watching is the one Agarwal and Sen couldn’t measure well: what happens in AI Mode. Users left that arm in large numbers, and among those who stayed, publishers’ click-through rates and user satisfaction were both substantially lower. A separate preregistered experiment by Stephanie Wang and colleagues, with 1,100 participants, also found that an AI Mode-only experience reduced click-through rates and eroded trust in information found on Google. If Overviews cost a site two clicks in five where they appear, the bigger question is what happens to that site if Google makes AI Mode the default. References 1. Agarwal, S. and Sen, A. 2026 . The Impact of Google AI Overviews on Publisher Traffic and User Experience: Evidence from a Field Experiment https://doi.org/10.2139/ssrn.6513059 . SSRN working paper 6513059, first posted 3 April 2026; version dated 17 June 2026. Preregistered as AEARCTR-0017393. 2. RevScope 2026 . Google’s AI Overviews cut clicks to websites by 40% https://revscope.co/blog/search-seo-geo-2026/ . 9 November 2026. 3. Pew Research Center 2025 . Google users are less likely to click on links when an AI summary appears in the results https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/ . 22 July 2025. 4. Reid, L. 2025 . AI in Search is driving more queries and higher quality clicks https://blog.google/products/search/ai-search-driving-more-queries-higher-quality-clicks/ . Google, 6 August 2025. 5. DataForSEO 2026 . Search Insider 2026 https://dataforseo.com/whitepapers/search-insider-2026 . Whitepaper, accessed 18 November 2026. 6. Wang, S. T., Gleason, J., Bart, Y., Wilson, C. and Metaxa, D. 2026 . AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence https://arxiv.org/abs/2608.18352 . arXiv:2608.18352, 18 August 2026. 7. Li, J., Huffman, S. and Tokuda, A. 2009 . Good abandonment in mobile and PC internet search https://doi.org/10.1145/1571941.1571951 . Proceedings of the 32nd International ACM SIGIR Conference on Research and Development in Information Retrieval, 43-50, July 2009. 8. Wu, Z., Sanderson, M., Cambazoglu, B. B., Croft, W. B. and Scholer, F. 2020 . Providing Direct Answers in Search Results: A Study of User Behavior https://doi.org/10.1145/3340531.3412017 . Proceedings of the 29th ACM International Conference on Information and Knowledge Management, 1635-1644, October 2020. 9. Radlinski, F., Kurup, M. and Joachims, T. 2008 . How does clickthrough data reflect retrieval quality? https://doi.org/10.1145/1458082.1458092 . Proceedings of the 17th ACM Conference on Information and Knowledge Management, 43-52, October 2008.