Home
Tuesday, 22 September 2026 WARSAW 12:56 NEW YORK 06:56
AI · Search · Performance · E-commerce
Just in

Arcade runs Europe’s first creator upfront. 150 brands came to London

SEO & AI search

AI Mode cut clicks to websites by 18.8 points in a 1,100-person test

A randomised experiment with 1,100 searchers found Google's AI Mode cut click-through rates to websites by 18.8 percentage points, while trust in what people read fell too.

Grzegorz Kubicki 22 Sep 2026, 09:29 reported from 2 sourcesarXiv - source document, Search Engine Journal
A laptop screen showing a page of search results, seen from close up.

Photo: cottonbro studio / Pexels

A randomised field experiment run in March 2026 found that Google’s AI Mode cut click-through rates to websites by 18.8 percentage points. The study comes from the University of Pennsylvania and Northeastern University. It went online on arXiv on 18 August 2026 and reached the search trade press on 21 September.

Penn and Northeastern tracked 1,100 searchers for ten days

The team recruited 1,444 people in mid-March and asked them to install a browser extension. It logged what they searched for and what they clicked. Results come from the 1,100 people who actually ran searches, and 956 of them answered a survey at the end.

Everyone was measured as usual for three days. Then each person spent seven days in one of three groups. One group was pushed into AI Mode. A second had AI Overviews and AI Mode stripped out. The third was left alone. Because the groups were drawn at random, the gaps between them can be read as effects rather than coincidence.

AdSense in-article 336x280

News sites lost 12.5 points, Reddit 21.2 points

The drop under AI Mode was not spread evenly. Clicks to news sites fell 12.5 percentage points and clicks to Wikipedia 9.9 points. Reddit lost the most, 21.2 points. Numbers like these sit behind decisions such as the 220 editorial jobs Reach cut while pointing at lost Google traffic.

Ad clicks fell 42.7 points, for a plain reason: AI Mode carried no ads at all during the test. A results page without ads is a page nobody can buy.

Stripping the AI out raised clicks by 8.8 points

The opposite condition moved the other way. With AI Overviews and AI Mode removed, click-through rates rose 8.8 percentage points against normal search. That group also spent 0.59 minutes less per session. People found what they wanted and left.

AI Mode users behaved differently. They ran 0.92 fewer search sessions a day and stayed 0.43 minutes longer in each one. Google has been trying to settle the money side of this separately, and now pays some publishers when its AI uses their pages.

Trust in Google fell 0.34 points on a seven-point scale

The survey answers pointed the same way. Trust in the information people found dropped 0.34 points on a seven-point scale. Satisfaction fell 0.73 standard deviations and usefulness 0.59. More people in the AI Mode group also ran searches on Bing, DuckDuckGo or Yahoo.

Two limits are worth naming. The paper is a preprint, so it has not been through peer review. And the panel was American, testing a product that keeps changing: Google added Gemini 3.8 Flash to AI Mode after the fieldwork ended.

Open Search Console and pull 90 days for the queries where your pages show up in AI answers. Put impressions next to clicks. A flat impression line with falling clicks is the pattern this experiment produced. Then check whether those same queries still return paid results, because the ad slot the researchers measured was empty.

Sources

  1. arXiv - source document — Working paper 'AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence', submitted 18 August 2026
  2. Search Engine Journal — Trade coverage of the study, 21 September 2026

Published by IMARKETING Sp. z o.o., Al. Jana Pawla II 150/127, 31-864 Krakow, Poland. Company register (KRS) no. 0001203282, VAT ID PL6751820741.  ·  Editorial office: Al. Jana Pawla II 150/127, 31-864 Krakow, phone +48 516 110 813, editor@marketingnewsroom.com.  ·  Editor-in-chief: Grzegorz Kubicki.  ·  Corrections: editor@marketingnewsroom.com.