A randomized field experiment published in August 2026 found that Google’s AI-powered search experience reduced clicks to external websites and worsened users’ perceptions of Search. For publishers and marketers who depend on organic traffic, the combination of lower referral volume and weaker user satisfaction is a practical problem, not a theoretical one.

What the experiment measured

Researchers from the University of Pennsylvania and Northeastern University enrolled 1,444 volunteers and tracked search activity for 1,100 participants over seven days. A browser extension captured search queries, snapshots of result pages and click behavior so the team could record when Google showed its AI “Overviews” and whether participants clicked through to external sites. Of the tracked users, 956 completed a post-experiment survey.

Participants were randomized into three conditions: forced regular Google Search (AI Overviews hidden), natural Google behavior (AI Overviews shown when Google would normally display them) and forced AI Mode (searches redirected into Google’s AI Mode). The extension initially hid AI Overviews successfully, but an HTML change from Google during the study reduced the hiding rate; overall, about 51.1% of AI Overviews were hidden and the median participant had 50% hidden.

Key findings

The study’s randomized design produced consistent results across behavioral and survey measures:

  • AI Mode reduced clicks to external websites. Participants routed into AI Mode were substantially less likely to click through to publishers.
  • Hiding AI Overviews increased clicks to websites without producing a detectable decline in measured search satisfaction.
  • AI Mode kept users inside Google’s interface longer and acted as a substitute for visiting source sites rather than a discovery aid.
  • Users exposed to AI Mode reported worse perceptions of Search—lower trust, reduced usefulness and lower personal relevance—contradicting claims that users overwhelmingly prefer AI summaries.
  • Assignment to AI Mode increased the share of users who subsequently searched on competitors (Bing, DuckDuckGo, Yahoo) and raised intentions to switch for some participants.
  • Contrary to the expectation that AI Mode would boost search engagement, participants routed into AI Mode conducted fewer search sessions. Heavy users saw a larger drop in activity than light users.

Google’s public claims versus the experiment

Google executives have publicly said users embrace AI-driven summaries and that AI features drive traffic to the open web. This randomized field experiment presents evidence pointing the other way: AI Mode reduced publisher referrals and lowered measured user satisfaction in the study sample. The paper does not prove intent or misrepresentation by Google, but it does show a clear mismatch between the company’s public statements and these experimental results.

Why marketers and publishers should care

Three practical consequences follow directly from the findings.

  • Publisher referral risk: AI Overviews that substitute for site visits can reduce ad impressions, subscriptions and content-led acquisition that rely on visits.
  • Visibility and attribution: fewer clicks mean fewer measurable opportunities to engage visitors and attribute conversions, complicating measurement and ROI calculations tied to organic search.
  • Discovery and channel strategy: lower satisfaction and increased intent to try competitors suggest search behavior and market share could shift—brands should be prepared for fluctuating organic funnels.

Limitations and what to watch

The study has constraints readers should weigh. It ran for seven days, used a browser extension that encountered a Google HTML change (reducing the intervention’s effectiveness), and reflects the behavior of the participants sampled. Those limits restrict how confidently we can generalize across time, user segments and future product iterations.

Still, the randomized assignment, the tracked sample of 1,100 users and the combined behavioral and survey evidence make the results credible enough to warrant attention. Replication by independent teams or complementary analytics from publishers would strengthen the case.

Practical next steps for teams

Marketers and publishers should treat this as an operational risk that can be addressed empirically:

  • Audit query-level referral trends now to identify pages and queries where external clicks have dropped.
  • Prioritize content and pages that historically drove conversions, and measure downstream value—time on site, leads and revenue—not just raw clicks.
  • Diversify acquisition: strengthen direct channels (email lists, social), invest in brand search and test formats that encourage click-through (interactive tools, proprietary data, gated content).
  • Track independent research and Google product changes: adjustments to how AI Overviews are generated or attributed can change referral dynamics quickly.

The central question for any organization that depends on search traffic is empirical: do your analytics show the same patterns? Audit now, treat declines as actionable signals, and prioritize experiments that measure revenue outcomes rather than relying solely on traffic volume.