AI search cuts publisher traffic. New research finds users are not better off.

The debate over whether AI search is taking traffic away from publishers has largely suffered from one methodological problem. Most of the evidence has been observational. Publishers see traffic falling. AI Overviews become more prevalent. Researchers observe lower click-through rates when AI answers appear. The association can be compelling, but it is difficult to establish that the AI feature itself caused the change.

New research takes a considerably stronger approach. In a paper submitted 18 August 2026, researchers from the University of Pennsylvania and Northeastern University report results from a preregistered randomised field experiment involving 1,100 Google users. Instead of simply observing how people behaved when AI appeared, the researchers manipulated which version of Google Search participants experienced. Their central conclusion is unusually clear: when Google's AI search features were removed, people clicked through to external websites more. When searches were forced into AI Mode, they clicked through substantially less and reported a worse search experience.

That makes this more than another zero-click-search study. It provides causal evidence about what happens when generative AI is inserted between information producers and the people searching for their information. And the result presents Google, and potentially the entire AI-search ecosystem, with an uncomfortable question: if publishers receive less traffic, what exactly are users getting in return?

<h2>This was a real-world experiment, not a survey</h2>

The study, "AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence," was conducted by Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson and Danae Metaxa. Participants were recruited in the United States and had to be adults who primarily used both Google Chrome and Google Search. They installed a browser extension and initially spent three days using ordinary Google Search so researchers could establish baseline behaviour.

Participants were then randomly assigned to one of three conditions for seven days. The first condition, Current Search, was ordinary Google Search with AI Overviews appearing naturally and AI Mode available. The second condition, No AI Search, attempted to hide AI Overviews and redirected AI Mode searches back to conventional search. The third condition, AI Mode Search, redirected searches into Google's Gemini-powered AI Mode. The researchers then measured actual search and click behaviour rather than simply asking participants what they thought they would do. After the experiment, they also measured attitudes including trust, usefulness, satisfaction, agency and perceived relevance. That experimental design is what makes the study particularly important. Random assignment gives the researchers much stronger grounds for making causal claims about the effects of AI search.

<h2>Removing AI increased click-through</h2>

The first major result concerns Google's AI Overviews. Exposure to the No AI Search condition increased external click-through rate by 8.8 percentage points, with a 95% confidence interval of 2.3 to 15.3 percentage points. The result was statistically significant at p=0.008. Put plainly, when users were exposed to fewer AI Overviews, they were more likely to leave Google and visit external websites. That provides experimental support for something publishers have argued for some time: an AI-generated answer can substitute for the visit that might otherwise have gone to the original information source.

There is, however, an important methodological qualification. The researchers' browser extension did not successfully remove every AI Overview throughout the experiment. Google changed the HTML underlying AI Overviews during the study, breaking the researchers' intervention. Overall, 51.1% of AI Overviews were successfully hidden in the No AI condition. The researchers therefore used a local average treatment effect analysis to account for this incomplete compliance. That detail matters, and it should not be omitted when interpreting the 8.8-point result. But the direction of the finding remains significant: less exposure to AI Overviews produced more external clicking.

<h2>AI Mode had an even larger effect</h2>

The AI Mode experiment produced a much larger change. Assignment to AI Mode reduced external click-through by 18.8 percentage points, with a 95% confidence interval of -22.2 to -15.3 percentage points and p<0.001. The effect was not confined to one type of website. Compared with ordinary Google Search, assignment to AI Mode reduced the fraction of users clicking through to news sites by 12.5 percentage points, Reddit by 21.2 percentage points, and Wikipedia by 9.9 percentage points.

This is important because AI Mode represents something more fundamental than adding an AI-generated summary above traditional search results. It changes the interface itself. Traditional search essentially says: here are sources that might answer your question. AI Mode increasingly says: here is an answer constructed from sources. That distinction appears to change user behaviour substantially.

<h2>Users stayed inside the AI experience longer</h2>

There is another revealing result. Assignment to AI Mode increased time per search session by 0.43 minutes while simultaneously reducing the number of search sessions per day by 0.92. Users spent more time within individual AI-mediated sessions while visiting the wider web less. The emerging architecture of AI search potentially shifts attention from the pattern of search leading to publisher, toward search leading to AI answer leading to continued AI interaction. The AI platform increasingly becomes not simply the route to the information, but the environment in which the information is consumed. For publishers, that is a fundamentally different distribution model.

<h2>The finding that makes this study especially interesting</h2>

One possible defence of declining publisher traffic is straightforward: perhaps users are simply better off. If an AI system answers the question more efficiently, perhaps fewer clicks represent a superior search experience rather than a problem. The researchers tested that proposition. And AI Mode did not perform well. Compared with ordinary Google Search, assignment to AI Mode significantly reduced trust in information found on Google by 0.34 points on a seven-point scale, perceived usefulness by 0.59 standard deviations, satisfaction by 0.73 standard deviations, agency by 0.66 standard deviations, and perceived personalisation and relevance by 0.42 standard deviations. All of these effects were statistically significant at p<0.001.

This is the finding that changes the story. AI Mode did not simply reduce external traffic while delivering a demonstrably superior experience. In this experiment, it reduced external traffic while users also reported lower trust, usefulness, satisfaction, agency and relevance. That is a much more difficult trade-off to defend.

<h2>Removing AI Overviews did not make users less satisfied either</h2>

The comparison with AI Overviews is equally revealing. When exposure to AI Overviews was reduced, researchers found no detectable deterioration in trust, satisfaction, usefulness or agency. In other words, reducing exposure to AI Overviews increased external click-through, but the researchers did not detect a corresponding loss in the measured user-experience outcomes. The authors summarise the broader result as a trade-off in which AI-mediated search substantially reduces referral traffic without commensurate improvements in user trust or overall search experience. That is substantially different from saying users prefer convenient AI answers and therefore no longer need to click. At least in this experiment, the evidence is more complicated.

<h2>Some users actively went somewhere else</h2>

There is another finding that deserves attention. Participants assigned to AI Mode were 11.2 percentage points more likely to use a competing search engine such as Bing, DuckDuckGo or Yahoo. They also reported a significantly greater intention to switch to Bing. That result runs directly against the researchers' expectation that AI Mode would increase search engagement. The effect was particularly pronounced among heavy Google users: AI Mode reduced daily search sessions significantly more among participants who had previously conducted more than four Google sessions per day. This does not establish how the entire population will react as AI search evolves. But it suggests that forcing conversational AI into the search experience does not automatically create greater engagement or loyalty.

<h2>The qualitative responses help explain why</h2>

After using AI Mode for a week, participants were asked about their experience. Negative responses were somewhat more common than positive ones: 33.6% expressed negative sentiment, versus 29.3% positive and 13.7% mixed. Several complaints are particularly relevant to AI-search design. Some 17.6% mentioned loss of control or agency. Another 15.3% described difficulty navigating directly to specific websites. And 13.4% mentioned limited links or source diversity. Some participants also complained about verbosity and accuracy, while positive comments commonly emphasised efficiency and time savings. That suggests the issue may not be whether users want AI assistance. It may be whether they want AI assistance to replace navigation and source choice. Those are not the same thing.

<h2>This matters directly to AEO</h2>

For brands and publishers, the strategic implication is substantial. Traditional SEO was largely designed around winning a click. AI search increasingly introduces another possible outcome: the source influences the answer but never receives the visit. That means visibility and traffic are becoming progressively separable. A publisher can theoretically become highly influential inside an AI answer while receiving very little direct audience. Likewise, a brand may appear prominently in AI recommendations without generating a conventional search referral.

That reinforces why AEO measurement cannot simply inherit SEO metrics. Practitioners increasingly need to distinguish among retrieval (was the source accessed?), citation (was the source visibly referenced?), representation (did its information shape the answer?), brand visibility (did the brand appear?), recommendation (was the brand actively endorsed or selected?), referral (did the user actually leave the AI environment?) and choice (did the interaction ultimately affect behaviour?). Those are different stages of the AI-mediated decision process. Clicks remain economically important. But they are no longer a sufficient measure of influence.

<h2>Citation quality becomes more important when fewer users verify sources</h2>

There is an uncomfortable implication for publishers. If fewer users click through, then fewer users independently inspect the source material behind an AI-generated answer. That shifts the burden of accuracy from the reader to the system. When a traditional search result linked to a publisher, the user could evaluate the source directly. In an AI-mediated environment where the answer is consumed without visiting the source, the citation itself becomes the primary trust signal. For AEO practitioners, this means the quality and verifiability of the evidence that AI systems retrieve is not merely a technical concern. It is an economic one. Brands whose evidence is strong enough to survive scrutiny without the user ever visiting the page are better positioned than those relying on click-through to demonstrate value.

<h2>AEO Updates Takeaway</h2>

This study provides the strongest causal evidence to date that AI search features reduce publisher traffic without delivering a compensating improvement in user experience. For AEO practitioners, the implication is threefold. First, the separation between visibility and traffic is not speculative; it is now experimentally demonstrated. Second, measurement systems that rely solely on referral traffic will increasingly undercount genuine brand influence inside AI answers. Third, the quality of the evidence environment, not just its optimisation, determines whether a brand's contribution survives in a world where users consume answers without visiting sources. The experiment also suggests that users do not uniformly prefer AI-mediated search, and that forced AI experiences can reduce both satisfaction and platform loyalty. That creates a more complex competitive landscape than a simple zero-click narrative implies, and it reinforces why AEO strategy must account for user experience outcomes alongside visibility metrics.

<h3>References</h3>

[1] 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." arXiv:2608.18352. Submitted 18 August 2026. <a href='https://arxiv.org/abs/2608.18352'>https://arxiv.org/abs/2608.18352</a>