A University of Washington paper says Google's AI Overviews cut Wikipedia's search referrals from English readers by about 5 percent

A working paper by two University of Washington researchers estimates that Google's AI Overviews reduced monthly external-search referrals to English Wikipedia by 5.45 percent against German and 4.82 percent against French, once AI Overviews became the U.S. default in May 2024. Mehrzad Khosravi and Hema Yoganarasimhan posted the latest version, v6, on arXiv on 2 September; Search Engine Journal reported on it on 11 September. The paper has not been peer reviewed. Google disputes the referral metric, saying the combined external-search figure Wikimedia publishes cannot isolate its product.

The method, in plain terms

Wikimedia releases monthly clickstream files that count, per article, how many visits came from external search engines. The files bundle all search engines together and drop pairs with low volume, and the authors report the estimate barely moved when they changed the count they used for the suppressed pairs. The paper's difference-in-differences design compares against German and French Wikipedia, which draw most of their traffic from countries where AI Overviews were not yet on by default. The panel matches 499,927 English-German article pairs and 530,873 English-French pairs over December 2023 to December 2024, with May 2024 as the first post-treatment month.

Multiplied out to the full English edition, the authors put the drop at around 100.27 million fewer search referrals a month, or about 1.20 billion a year. Both are model-based, and the yearly figure assumes the effect holds evenly across the year. An English-Japanese check on a shorter window shows a 16.53 percent decline, and the authors call this directional support only.

What the paper does not claim

The paper measures search referrals. A reader who arrives from Google and opens three more articles counts as one referral. Most English Wikipedia readers are outside the United States, which pulls the measured effect down. And because Wikimedia's data lumps all search engines together, the design cannot fully rule out other 2024 events that hit English search but had no equivalent in German or French.

The revenue line reads carefully. The authors calculate that a comparable ad-supported site might lose between 10.82 million dollars and 37.08 million dollars a year at typical ad rates. Wikipedia runs no ads, so this describes a hypothetical site and no money changes hands.

How this fits with what Wikimedia is seeing

Wikimedia has its own numbers to compare. Product director Marshall Miller said in October 2025 that human pageviews across all Wikipedia languages were down about 8 percent year on year, after the foundation improved its bot detection and reclassified traffic from March to August 2025. That figure is a raw count across every language, so it cannot be added to the paper's 5 percent estimate. Wikimedia's draft plan for fiscal year 2026-27 says fewer referrals from Google and rising bot traffic are expected to continue, and notes around 90 percent of Wikipedia visitors have historically arrived through Google search.

For a site owner, the useful part is the method. Wikimedia publishes per-article clickstream data that made the natural experiment possible, and language editions gave the authors an untreated comparison group. Most publishers have no equivalent public panel. The next-best move is a within-site version of the same comparison: query clusters that trigger AI Overviews against query clusters that do not, held over the rollout window in a market, with Search Console click data as the outcome. It is a smaller estimate, checkable against your own logs.

Source

Search Engine Journal: What Wikipedia Reveals About AI Overviews And Web Traffic, and the working paper Khosravi and Yoganarasimhan, Impact of AI Search Summaries on Website Traffic (arXiv 2602.18455v6).