A Randomized Study Put a Causal Number on AI Overviews

Until now, almost everything published about AI Overviews and traffic loss has been correlational. Traffic fell. AI Overviews spread. People drew a line between the two.

On August 17, 2026, two academics published something different: a randomized experiment that can separate cause from coincidence.

What the researchers did

Saharsh Agarwal, Assistant Professor of Information Systems at the Indian School of Business, and Ananya Sen, Associate Professor at Carnegie Mellon University's Heinz College, built a custom Chrome extension.

They recruited over 1,000 users who installed it in early 2026, then randomly assigned each user to one of two conditions: standard Google Search, where AI Overviews appeared whenever a query triggered them, or a modified version where the Overview was removed in real time. They observed actual behavior for two weeks.

Random assignment is what makes this different. Because the two groups differ only in whether they saw the feature, the gap between them is the effect of the feature, not a reflection of what people were searching for or how the web changed around them.

What they found

• The presence of an AI Overview reduced organic clicks to third-party sites by 39.8%.
• It increased searches where the user clicked no links at all by 34.5%.
• There was no measurable improvement in users' perceived search quality or in how easily they found information.

Why the third finding is the important one

The first two numbers confirm, more rigorously, what site owners have been reporting for a year. The third one is the finding that changes the argument.

The standard defense of any platform change is that users prefer it. Traffic moving away from publishers is framed as a side effect of serving people better. Here that defense was tested directly, on real users, and it did not hold. Roughly 40% of the clicks that used to reach independent websites stopped, and the people the feature exists for could not tell the difference.

That reframes AI Overviews from an improvement with costs into a redistribution with a measurable loser and no measured winner among users.

Is this a publisher problem?

Only if you define publisher narrowly, and almost nobody should.

Any business whose website answers questions is publishing. Your service pages explain what something involves. Your pricing page explains what it costs. Your FAQ answers the objection before the call. That is the exact material an Overview summarizes, and the summary is where the reader stops.

The practical consequence is that the top of your funnel is being answered on someone else's surface. Not blocked, not penalized. Answered.

What to do about it, and what not to

The reflex is to optimize for citation inside the Overview. That deserves scrutiny before it becomes a budget line.

A citation returns a fraction of the visits the same ranking used to return, and it is hard to attribute. Recent measurement supports treating it soberly rather than as a replacement: separate data from agency Brainlabs across 54 client accounts found organic sessions down 10.5%, while key events from AI assistant referrals rose sharply and converted at 1.5 times the rate of organic search traffic. Fewer visitors, further along in their decision.

Put those two studies together and a more honest strategy emerges:

• Stop managing to the sessions number alone. It no longer describes your business, and the randomized study says roughly 40% of the decline is structural rather than something you did wrong.
• Segment AI referrals before judging the damage. If they convert better, volume loss and revenue loss are different problems.
• Be specific enough to be summarized accurately. If a machine is going to answer for you, vague capability language becomes a competitor's opportunity.
• Build a reason to be reached directly. An owned audience is the only channel that a summary layer cannot sit in front of.

The honest limitation

The Agarwal and Sen experiment is one study, run over two weeks, with just over 1,000 self-selected users who agreed to install a research extension. That population may search differently from the general public. The 39.8% figure should be read as strong evidence of a large effect, not as a constant that will hold for every site and every query type.

What it does settle is the direction and the rough magnitude, and it removes the most common counterargument: that users are getting a better experience in exchange.

Sources

• Saharsh Agarwal and Ananya Sen, "How To Preserve the Online Information Ecosystem in the Presence of Google AI Overviews," ProMarket, August 17, 2026.
• Digiday, "In Graphic Detail: How AI search has impacted the web traffic of over 50 advertisers," August 25, 2026, reporting Brainlabs data across 54 client accounts.