Does AI search summaries divert traffic away from Wikipedia?

The paper estimates that Google's AI Overviews (AIO) feature, which "presents a synthesized answer prominently on the search-results page, before the traditional organic links," caused a measurable drop in search-referred visits to English Wikipedia. Using a difference-in-differences design, the authors compare "monthly external-search referrals to English Wikipedia articles with referrals to the same articles in German and French," and find that "default AIO availability reduced English search traffic by 5.45% and 4.82%, respectively" against the two control languages. The excerpt frames this as evidence that "answer-producing digital intermediaries can materially reallocate attention away from informational publishers."

The mechanism rests on AIO's staggered geographic rollout and Wikipedia's multilingual structure as a natural experiment. English Wikipedia is the treated group because "approximately 40% of its traffic originates in the United States," where AIO became a default part of search during the study period; German and French editions of the same underlying articles serve as controls that were not (or less) exposed to default AIO. The outcome measured is narrow and explicit: "visits that arrive directly from external search engines," interpreted as "the first-order effect" of the feature, not a full accounting of user welfare or publisher revenue.

This sits apart from the other near neighbors in the library, which concern AI as a content producer competing for attention rather than AI as an answer layer displacing referral clicks. Can AI creators match human creators through posting volume alone? and Does AI help individual scientists while narrowing scientific focus? both describe AI-mediated systems reallocating attention or output away from a prior producer base — one across creators on a platform, one across subfields of science — while overall activity continues or even grows. This paper adds a third, distinct channel for the same general pattern: a search intermediary absorbing the user's information need before a click to the source ever happens, with the authors noting the "fundamental difference between link-based" intermediation and answer-based intermediation.

The excerpt does not measure publisher revenue, ad impressions, or subscription loss directly — Wikipedia itself "does not monetize its traffic through advertising or subscriptions," so the study's outcome is a traffic proxy, not an economic-harm measure for the monetized publishers the authors say are more exposed. It also does not test smaller, less-authoritative sites, non-English-dominant markets, or whether AIO's citations recover any of the lost clicks over time. Given that even a uniquely authoritative, top-ranking, non-monetized site lost roughly 5% of search referrals to a single interface change, the plausible implication — unproven here — is that smaller or ad-dependent publishers face effects at least as large, which is the "policy" stake the paper gestures at but does not quantify.