Routing every query into Google AI Mode cut publisher clicks by 18.8 points

Every query in the AI Mode arm went into AI Mode ^[5]. Everyday use is more selective, as the write-up notes among the study's limits ^[8]. So 18.8 points is the effect at full substitution, measured against an arm running unmodified Google Search, and what Google's current mix does to a given site is a separate number ^[1][5].

To move the number onto a publisher's own referral curve you need a second figure: the share of your incoming search demand that resolves inside an AI answer. Multiply 18.8 by that share. At one query in ten, linear scaling gives about 1.9 points of outbound click rate ^[1]. Linearity is an assumption. Effects were larger for heavier Google users, so the per-user response is not flat ^[6].

The destination breakdown is more useful than the headline. Reddit click-through fell about 21.2 percentage points, news sites about 12.5, Wikipedia about 9.9 ^[4]. In my view that ordering tracks how completely a generated answer stands in for the destination. Reddit's drop is roughly 2.1 times Wikipedia's ^[3].

Two of the measured effects point the same way for a publisher. The AI Mode group clicked out less often and also ran fewer daily Google search sessions ^[1][3]. Referrals per user per day is sessions multiplied by outbound clicks per session, and both terms moved down ^[2].

Session length went the other way. Participants stayed in the AI-mediated experience longer, and the extra time did not produce more engagement with the open web or better sentiment ^[7]. They also reported lower trust in what they found through Google ^[3].

The write-up states the headline result as external clicks falling 18.8 percentage points ^[1]. It states the central finding as a reduced share of participants who clicked through to outside sites ^[10].

The authors of AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence limit the claim themselves, cautioning that the results reflect AI Mode's design and implementation during the experiment and should not be treated as a forecast for every future version of Google Search or every market ^[9]. Participants were US-based and recruited in March 2026 ^[8]. In the third arm, with AI Overviews hidden and AI Mode unavailable, external traffic could rise ^[11].

For planning, the dev.to write-up recommends tracking informational, comparison, branded and conversion-oriented queries separately instead of judging search performance through one aggregate traffic number ^[12]. A page can stay visible inside a search journey and still lose visits when the answer satisfies the immediate need ^[13].