The evidence publishers needed arrived in November, and it arrived fourteen months after the door it would have opened was already shut. A randomized field experiment has now measured, with real people in their own browsers, what Google’s AI Overviews do to the websites those summaries are assembled from. On searches where an overview appeared, clicks out of Google to an outside site fell about 40 percent. Readers did not get a better search experience in exchange. They got a shorter page.

That is worth saying without the usual garnish. Google is not summarizing the web to help you find things. It is summarizing the web to keep you, and the people who did the reporting the summary is built on receive the bill. the field experiment by Saharsh Agarwal and Ananya Sen is the receipt, and the abstract states the result in the flat, unbothered language of people who know their method will be attacked: conditional on appearing, overviews reduce outbound organic clicks by 39.8 percent and raise zero-click searches by 34.5 percent, with no change in sponsored clicks or search frequency.

Two researchers did the work. Saharsh Agarwal is at the Indian School of Business; Ananya Sen is at Carnegie Mellon’s Heinz School. They recruited 1,065 US adults who search Google on desktop Chrome, paid them to install a custom extension, and ran the experiment for two weeks in early 2026. Participants were not told what was being tested. One group had the overviews stripped out and the remaining results shifted up so the page looked ordinary. One group was pushed into Gemini-powered AI Mode for every search. One group searched untouched. Superpower Daily walked the design arm by arm, and the detail that matters is that this was not a comparison of people who happened to see a summary against people who did not. It was an assignment, and then a reversal: the groups were swapped and the effect flipped with them.

Amit Mehta — He Told Publishers to Come Back With Evidence, and the Docket Had Already Closed

In September 2025 the federal judge who found Google had illegally monopolized search declined to give publishers a way to keep their work out of Google’s AI products. Judge Amit Mehta noted that the record carried no testimony from a single publisher, and reasoned in substance that a remedy cannot be granted on evidence nobody produced. It was a defensible ruling on the record in front of him, and it was also an invitation.

The invitation has been accepted, and the case that needed it is over. What is not over is the queue: the European Commission, Britain’s Competition and Markets Authority, Brazil’s CADE and India’s competition authority are all examining Google’s AI search and its effect on the publishers who supply it. Courtney Radsch laid out the numbers on November 20, and her framing is the one that should travel — a judge who asked for evidence has now been handed a pre-registered experiment he did not have. The remedy he declined to order in 2025 is the remedy four other jurisdictions are drafting in 2026.

The numbers, plainly. Overviews appeared organically on roughly 41 percent of the searches in the trial. Across those searches, outbound clicks fell from 0.62 per search to 0.37. Zero-click searches climbed from 54 percent to 73 percent. Courtney Radsch laid out the numbers on November 20. Sponsored clicks did not move. Neither did how often people searched. Neither did how long a visitor stayed on the site they did reach, or how fast they left it. The summary consumed the click and returned no measured benefit to the person who made it.

Saharsh Agarwal — AI Overviews Take the Click and Hand Back Seven Percent

Here is the part that should end the argument about citations being the consolation prize. Links cited inside the overview generated only about 7 percent of the outbound clicks on those searches. Being named by the summary is not the same as being read by the human, and the gap between those two things is the entire publishing business. A citation is a receipt. It is not a reader.

The losses also land where they hurt most precisely: the top of the page and the first three unpaid results. That is the shelf space a publisher competes for, spends years earning, and staffs a newsroom to hold. the working paper, revised in July 2026 notes the AI Mode arm went further — fewer external clicks per search, more zero-click searches — and describes it as exploratory, with higher dropout and some participants installing workarounds to get ordinary results back. People will fight to get their search results returned to them. That is a fact about demand, and Google should sit with it.

Courtney Radsch — An Opt-Out Per Publisher Is a Cure That Leaves the Disease Standing

Radsch, who directs the Center for Media & Digital Governance at the Open Markets Institute, served as discussant when the paper was presented through the Centre for Economic Policy Research’s media plurality series, and she has drawn the policy conclusion the data invites. she argues that per-publisher opt-outs leave the overview standing because removing one publisher’s material does not remove the summary — the overview gets reassembled from whoever else is still feeding it. A publisher that opts out competes for a click underneath an answer built from its competitors. She favors opt-in rules and compensation instead.

Say the boundary out loud, because it is where honest analysis beats advocacy: its own brief concedes the experiment never tested the remedies. An experiment that hides a summary from a user is not an experiment in letting a publisher withhold its content. Radsch’s remedy is a serious proposal and it is not a finding. Both statements are true, and the second one does not weaken the first.

Ananya Sen — The Gemini Arm Clicked Even Less, and Users Fought Their Way Back

One experiment is an argument. Two is a pattern. a second randomized experiment posted to arXiv in August 2026 — from Stephanie Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson and Danaé Metaxa, across Pennsylvania and Northeastern — ran its own three-arm trial on a different quarter’s data with about 1,100 Chrome users. Hiding the overviews raised the click-through rate to outside sites by 8.8 percentage points. Routing every search into AI Mode cut it by 18.8 points, cut the share of users clicking through to news sites by 12.5 points, to Reddit by 21.2 and to Wikipedia by 9.9. the Pennsylvania and Northeastern team put the trust drop at 0.34 standard deviations.

Two independent teams, two recruitments, two codebases, same direction. The Google defense has been that the summaries spare people effort. The trials measured effort — search frequency, session counts, time on site — and found the effort unchanged, the satisfaction flat or worse, and one arm so unpleasant that participants quit it. You can argue with a survey. It is much harder to argue with an assignment that keeps landing the same way when you flip it.

Google — The Defense Was That Summaries Help People, and the Users Did Not Say So

In 2025, when Pew found people clicked a traditional result on 8 percent of visits with an AI summary against 15 percent without, Google disputed the methodology. That is the standard move, and it usually works, because most studies of search behavior are observational and can be picked apart on selection. A pre-registered randomized trial that changes the page and then reverses the groups is the version of this study that survives that move. Radsch reported that the two-week window was clean of algorithm changes, and searches without an overview showed no effect — the control inside the experiment.

A newsroom that used to receive 62 readers out of every 100 searches now receives 37, and the seven who might have arrived through a citation inside the answer do not cover the difference. That arithmetic belongs to a specific desk: the city hall reporter, the trade publication with four staffers, the local site that covers the water board because nobody else will. The overview that answers the question exists because someone sat through the meeting. It has no byline, because the byline is the part it removed.

The AI-mekna did not learn to write the news. It learned to stand in front of the news and take the ticket. Google will say the experience is better. The people in the trial, who did not know what was being tested, said it was not.

Sources: SSRN, Tech Policy Press, Superpower Daily, arXiv.