The Click Disappeared.
The Better Answer Never Arrived.
Researchers randomized Google’s AI layer itself. Force AI Mode and clicks fall 18.8 points — trust, usefulness, satisfaction, and the sense of agency fall with them. Where the AI features were successfully hidden, clicks rose, while no measured perception outcome — trust, usefulness, satisfaction, or agency — improved. The click was surrendered for a better experience the preregistered averages could not find.
The zero-click bargain rests on a premise rarely tested alongside publisher traffic: the reader who never clicked got something better in return. In March 2026, researchers at the University of Pennsylvania and Northeastern University gave that premise a controlled test. They enrolled 1,444 U.S.-based Google Search users and had a browser extension quietly reassign the search interface itself for seven days — one group assigned to a No-AI intervention that redirected AI Mode searches and attempted to hide AI Overviews, one with Google unchanged, and one assigned to route every search into AI Mode, the full answer-engine experience; the 1,100 who searched during the window are the paper’s behavioral-analysis sample. The AI Mode routing succeeded for 94.7% of searches. The AI Overview hiding succeeded for 51.1% overall after a Google interface change disrupted the extension mid-study, so the paper estimates that arm’s effect from participants’ actual exposure rather than treating assignment as complete removal. The main experiment’s hypotheses were preregistered before its outcomes were collected; its target sample size was calculated from an earlier pilot. It is the first controlled test this record has identified of what the AI answer layer does to the people on both sides of it.
Assignment to AI Mode cut click-through to publishers by 18.8 percentage points. That half of the result surprises no one; the interface answers so the reader never has to leave. It is the other instruments that break the story open. The same readers — the 956 who completed the closing survey — trusted what they found on Google less: down 0.34 on a seven-point scale. They rated their results less useful, by 0.59 standard deviations. Less satisfying, by 0.73. They reported less agency over their own searching, by 0.66. And their stated intention to switch to Bing rose by 1.25 points on that same seven-point scale. The forced interface did not merely reduce the external visit. It failed to deliver the better experience the click was supposedly traded for.
The other arm of the experiment is the discipline. Greater exposure to successfully hidden AI features raised publisher click-through by 8.8 points and left the reader’s measured experience where it stood. Trust did not rise; the researchers’ own preregistered prediction said it would, and the data declined to cooperate. Satisfaction, usefulness, the sense of agency: no measurable change. Where the hiding succeeded, search did not get better. Forcing the answer made it worse. The two interventions are not mirror images, and that asymmetry is the finding: the click can be taken without the compensating benefit ever arriving.
What the experiment can claim — and what it cannot
What it establishes. A randomized, preregistered field experiment is the strongest instrument this record’s zero-click file has carried: causal, not correlational. Prior evidence inferred harm from traffic patterns. This one randomized the interface itself and watched both sides of the trade.
What it does not establish. The estimates belong to this intervention, this U.S.-based sample — which the authors themselves note skews younger, more educated, and more left-leaning than the U.S. population — seven days of exposure, and the March 2026 interface. Because the No-AI intervention’s compliance deteriorated, its 8.8-point estimate is a local average treatment effect that assumes assignment affected outcomes only through actual exposure to hidden AI features; the authors caution that an interface change affecting participants more broadly could violate that assumption. None of it licenses a verdict on AI search universally.
Instrument class, on the record. An unreplicated preprint, posted August 18, 2026, not yet through peer review, with its preregistration and data filed openly. Much of the public zero-click evidence is vendor-instrumented, or arrives inside litigation argued by a party with a number to defend. A randomized experiment with hypotheses filed in advance is a different class of instrument. That is exactly why its limits deserve stating precisely.
Recognized, retrieved, and rerouted
The finding lands in this record’s oldest seam. Edition No. 021 filed the census: an AI can cite a business by name and send nobody, and 68% of U.S. Google searches were already ending without a click. What the experiment now adds is causal evidence that interface assignment itself can reduce aggregate external click-through. It does not test why any particular source was selected, whether an omitted source merited inclusion, or whether relevance and authority were weighed. In their open-ended responses, participants assigned to AI Mode described difficulty reaching desired sites and seeing fewer, less diverse sources. In Agentics terms, that is evidence consistent with Decision Exclusion at the interface layer — not controlled proof that a named entity was excluded from a specific decision.
The trust ledger carries a narrower second entry. The study did not measure how much trust individual sources would have earned on a direct visit. It found that assignment to AI Mode reduced trust in information found on Google while also reducing external click-through. On these instruments, forcing the answer layer did not preserve trust in the search experience while keeping more attention inside it. The doctrine that follows is Agentics’, not the paper’s: this is the seam Answer Engine Authority™’s fifth phase, zero-click presence, exists to close — not chasing a visit the interface is built to prevent, but making the recognition itself, the name, the claim, the trust that moves inside the answer, do the work the visit used to do.
The zero-click era has been argued from traffic graphs, lawsuits, and platform statements. This time, someone flipped the switch at random and measured both sides. The interface reduced external click-through. Across the preregistered average measures, readers gained no compensating improvement — and several outcomes worsened. And the businesses on the other side of the answer had no vote in the interface that decides their visits.
Stephanie T. Wang (University of Pennsylvania), Jeffrey Gleason (Northeastern University), Yakov Bart (Northeastern University), Christo Wilson (Northeastern University), and Danaé Metaxa (University of Pennsylvania) — “AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence,” a preregistered randomized field experiment (1,444 U.S.-based participants enrolled; behavioral outcomes estimated among the 1,100 who made at least one search, the paper’s abstract N; perception outcomes among the 956 who completed the post-survey), fielded March 2026, posted to arXiv August 18, 2026 — arXiv:2608.18352. All figures cited above are drawn from and verified against the paper’s full text; the preregistration and replication data are filed with OSF and the Harvard Dataverse per the paper. The paper is an unreplicated preprint and is presented as such.
The zero-click census cited from Edition No. 021 is SparkToro’s ongoing research using Similarweb clickstream data: 68.01% of U.S. Google searches ended without a click in the first four months of 2026 — SparkToro.