Publishers have been saying for two years that Google’s AI-generated answers are eating their traffic. Now there’s experimental evidence — and it comes with an awkward second finding: the AI answers don’t even make users happier.
The paper, “AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence” (arXiv:2608.18352), was submitted on August 18, 2026, by Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson and Danaé Metaxa. It’s a preregistered field experiment — meaning the researchers publicly committed to their methods and hypotheses before collecting data, which makes the results harder to dismiss as cherry-picked.
How the experiment worked
The team recruited 1,100 participants and ran the study on Google Search itself, the dominant search platform. Participants were assigned to different search experiences: the standard Google results page, a version with AI Overviews and AI Mode removed, and an AI Mode-only experience. The researchers then measured what users actually did — click-through rates to publishers — alongside how users perceived the experience: satisfaction, trust in the information, overall experience quality.
This design matters. Most of the debate about AI Overviews has relied on observational data — publishers watching their traffic fall as AI answers rolled out, with Google arguing correlation isn’t causation. A randomized field experiment isolates the causal effect: change only the presence of AI-generated answers, and see what moves.
What they found
Two results, both bad news for the current direction of search:
- Removing AI Overviews and AI Mode increases click-through rates to publishers. When the AI-generated answers were taken away, users clicked through to publisher sites more. The AI layer is intercepting visits that would otherwise have gone to the sites that produced the underlying information.
- An AI Mode-only experience reduces click-through rates and erodes user experience and trust. Users who got only the AI-generated experience didn’t just click less — they rated the experience worse and trusted the information less.
Put together, the findings describe a lose-lose: publishers lose the traffic that sustains their business, and users get an experience they like less and trust less. The integration of generative AI into web search, the authors argue, reshapes online attention with direct economic consequences for the publishers that supply the content both search platforms and the broader information ecosystem depend on.
Why this study lands differently
Plenty of industry analyses have claimed AI search hurts publishers. What this paper adds is causal evidence from a preregistered design on the actual platform, measuring both behavior and perception. The trust finding is the genuinely new part: the industry conversation has assumed, mostly without evidence, that users prefer getting a synthesized answer to clicking through links. This experiment suggests that preference is weaker than assumed — at least for AI Mode as currently implemented.
There are limits, as with any single study. Eleven hundred users is a solid sample but a slice of Google’s billions. The experiment captures short-term behavior, not how users might adapt over months. And “user experience” in a study setting may not capture the convenience of quick answers for simple factual queries. But the direction of the effects — less publisher traffic, worse perceived experience — is now evidence, not anecdote.
The bigger question
The paper lands in the middle of an unresolved economic negotiation between AI platforms and the publishers whose content trains and grounds them. If AI answers reduce the traffic that funds content creation, while the content is what makes the answers good, the system has a circularity problem: the better the AI gets at answering from publisher content, the less incentive there is to produce that content.
Licensing deals — the ones OpenAI, Google and others have struck with major publishers — are one attempted answer, but they cover a handful of large outlets, not the long tail of the web. A recent appeals court ruling on AI training and fair use — with a catch — is already reshaping the legal half of that negotiation. The study doesn’t solve the business model. It does something more basic: it establishes, with unusual rigor, that the tradeoff is real, and that users aren’t obviously winning it either. As AI assistants take over more of how people find information — the same shift that killed OpenAI’s Atlas browser — the question of who gets paid for the underlying knowledge is only getting sharper.
Why preregistration matters
The study’s preregistered design deserves a note, because it’s doing quiet work in the background of these findings. In a preregistered experiment, researchers publicly commit to their hypotheses, methods and analyses before seeing the data — which rules out the most common ways findings get inflated: trying many analyses and reporting the one that worked, or reshaping the hypothesis to fit the results. For a question this commercially charged, where every stakeholder has a preferred answer, that discipline is what separates evidence from ammunition.
The field-experiment setting matters too. Lab studies of search behavior often use mock interfaces, which participants treat differently from the real thing. Running the experiment on Google Search itself — with real queries, real results pages and real stakes — means the measured click-through differences reflect actual user behavior, not how people think they’d behave. Eleven hundred participants is a substantial sample for this kind of intensive design.
The circularity problem
The paper’s economic argument is the part with the longest shadow. Search engines and AI assistants are trained and grounded on publisher content; AI-generated answers reduce the traffic that funds that content; less funding means less content worth grounding on. It’s a loop, and right now it’s running in one direction.
The industry’s attempted answer — licensing deals between AI companies and major publishers — covers a handful of large outlets, not the long tail of independent sites that make up most of the web’s depth. Nothing in this study suggests those deals change the underlying dynamic for everyone else. And the trust finding complicates the usual defense: if users liked the AI answers more, you could argue the tradeoff serves them. The experiment found the opposite — an AI Mode-only experience that users rated worse and trusted less.
None of this means AI in search is going away; the direction of travel is set. But the study reframes the negotiation. The question is no longer whether AI answers affect publishers — that’s now causal evidence — but what the information economy looks like when the answers get better and the sources get poorer. That’s a design problem for the whole industry, and it’s barely been started.
Featured image: Nenad Stojković via Wikimedia Commons (CC BY 2.0)
