Google AI Mode cuts publisher click rates by 19 percent

What the numbers say

A field experiment published on 18 August 2026 put a hard number on what publishers have been watching in their analytics for two years. Assigning users to Google AI Mode reduced their external click-through rate by 18.8 percentage points compared with standard search (95% confidence interval:, 22.2pp to, 15.3pp; p<0.001).^[1] The finding comes from researchers at the University of Pennsylvania and Northeastern University, whose paper, "AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence", was published on arXiv.^[2]

Hiding AI Overviews (the synthesised answer blocks that appear above standard search results) entirely from participants pushed click-through rates up 8.8 percentage points (95% CI: 2.3pp to 15.3pp; p=0.008).^[1] AI Overviews and AI Mode together compound the drag on referral traffic rather than one offsetting the other.

How the study was built

The researchers recruited 1,444 participants between 17 and 19 March 2026, drawing mostly from Prolific and a smaller Northeastern work-study cohort, and restricted enrolment to US-based Google Chrome users.^[1] Each participant installed a browser extension that controlled whether AI Overviews and AI Mode appeared during a ten-day observation window. Of those recruited, 1,100 made at least one search and 956 completed the post-experiment survey.^[1]

The design is a randomised assignment measuring what participants actually did, not a survey asking what they think they do. That makes the click-through figures unusually hard to argue away on methodological grounds.

The satisfaction problem

Users assigned to AI Mode did not just click less; they felt worse about the whole experience. AI Mode users reported satisfaction 0.73 standard deviations lower than standard search users, usefulness 0.59 standard deviations lower, and sense of agency 0.66 standard deviations lower (all p<0.001).^[1]

Google has asked publishers to accept fewer clicks in exchange for a better overall experience that keeps users inside the Google product. The experiment returned fewer clicks and a worse experience at the same time.

Google's position, set against the data

On 17 July 2026, Google Senior Vice President of Knowledge and Information Nick Fox wrote on LinkedIn: "In fact, we're now sending billions of clicks to websites every week through AI features in Search alone, and we're just getting started."^[3] The Penn and Northeastern paper was published 32 days later, and its experimental design directly tests whether that directional claim holds at the individual search level.

Google's Vice President of Search Liz Reid told Ars Technica: "Google search is AI search."^[4] That framing makes the click-through findings harder to treat as a niche product problem. AI Mode (a fully conversational search experience powered by Google's Gemini models, launched in May 2025) is the product, and the product reduces referrals by nearly 19 percentage points.

Google has also introduced an undisclosed-rate payment pilot with some publishers, running alongside these product changes.

What it means for brands and agencies in Australia

Australian agencies and their clients have been absorbing anecdotal evidence of declining organic search traffic since Google introduced AI Overviews in May 2024. The Penn and Northeastern paper converts those anecdotes into a randomised controlled estimate with tight confidence intervals.

Any campaign sold on AI search visibility, meaning the appearance of a brand or its content inside an AI Mode response, is being bought in an environment where the same feature measurably suppresses the click-through that conversion tracking depends on. Traffic models built before May 2025 are calibrated to a product that no longer exists in the same form.

The study covers US Chrome users searched in March 2026, so direct translation to Australian conditions requires caution. The mechanism, an AI answer that satisfies the query before the user reaches a publisher link, is not geographically bounded. The paper is available in full at arXiv, published 18 August 2026.