A field experiment that enrolled 1,444 Google users and tracked their searches for a week found that assigning people to AI Mode reduced how often they clicked through to outside websites and made them trust Search less. The result contradicts Google’s public position, repeated by CEO Sundar Pichai and by Liz Reid, that users love AI Mode and that it sends the open web more traffic than ever. The researchers tested that claim with a randomized design instead of relying on aggregated company metrics.
The paper, titled “AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence,” comes from researchers at the University of Pennsylvania and Northeastern University in Massachusetts. It is a draft, posted in August 2026, so its findings have not yet been through formal review. Of 1,444 people enrolled, 1,100 had their searches tracked through a browser extension for seven days, and 956 completed a follow-up survey.
Participants were randomly split into three groups: one saw Google Search behave normally, one had AI Overviews hidden and AI Mode queries redirected to classic results, and one was routed into AI Mode regardless of what they typed. The extension reliably pushed 94.7 percent of the AI Mode group into that experience. The no-AI condition worked less well over time: it suppressed 90 percent of AI Overviews on day one, but a mid-study edit to Google’s AI Overview markup broke the extension’s detection logic, and suppression fell toward zero as the week wore on. Hidden success reached 51.1 percent overall, with the median participant seeing about half of all AI Overviews removed.
Being routed into AI Mode cut clicks to external sites and lowered how many search sessions people ran per day. Removing AI Overviews had the reverse effect: it increased clicks to outside sites without any measurable drop in how people rated the search experience. Google has said AI Mode sends more traffic to the web than ever. This study’s own control arm, where AI features were stripped out entirely, found no evidence that doing so hurt anyone’s experience.
AI Mode also lowered how much participants trusted Search, how useful and satisfying they found it, and how personally relevant it felt, according to the post-experiment survey. Being placed in AI Mode made people meaningfully more likely to search on a rival engine such as Bing, DuckDuckGo, or Yahoo, and more likely to say they intended to switch to Bing, the researchers found. The pattern held broadly across the sample, with one notable exception: heavy Google users cut their search activity more sharply under AI Mode than light users did. The team found no meaningful difference in how people judged AI Mode versus classic Search specifically for news queries.
Search Engine Journal, which first reported on the draft paper, framed the findings as a direct rebuttal to Google’s own account of AI Mode as a traffic driver rather than a traffic sink. For a search team, the gap between a platform’s self-reported aggregate numbers and a controlled academic experiment is the gap between marketing and evidence, and this is a randomized test of AI Mode’s effect on referral behavior rather than an estimate built from logs one side controls.
Because the paper is a draft awaiting peer review, its exact effect sizes could shift before formal publication. Search teams that have leaned on Google’s traffic claims to justify deprioritizing referral tracking should treat this experiment as a reason to pull their own AI Overviews and AI Mode click data over the next quarter rather than wait for the review process to finish.
Search Engine Journal, in a report by Roger Montti published September 21, 2026, first covered the draft paper’s findings.
