A preprint^[1] by two researchers from the University of Washington's Foster School of Business

"[...] estimate[s] the impact of Google's AI Overviews (AIO) on Wikipedia's search traffic using AIO's staggered geographic rollout and Wikipedia's multilingual structure. Our difference-in-differences design compares monthly external-search referrals to English Wikipedia articles with referrals to the same articles in German and French, and finds that default AIO availability reduced English search traffic by 5.45% and 4.82%, respectively"

(Google's AI Overviews add an AI-generated natural language response on top of the usual results list of Google Search. The paper's appendix provides a detailed comparison in case of the search term "e=mc2", featuring screenshots of the resulting page with and without AIO.)

The researchers point out that (contrary to popular assumptions) finding such a decline was not a given:

A priori, the effect of AIO on Wikipedia traffic is ambiguous. On the one hand, Wikipedia’s concise factual content resembles the information that AIO can synthesize and present directly, making visits to Wikipedia potentially easy to replace. On the other hand, Wikipedia ranks prominently in organic search and is among the sources most frequently cited in AI-generated summaries (Harsel et al. 2025). These citations could increase Wikipedia’s visibility and direct users to its pages. AIO could therefore either substitute for Wikipedia visits by satisfying users’ information needs directly or complement them by increasing exposure to Wikipedia’s content.

(Here "Harsel et al. 2025" appears to a November 2025 post by SEO firm Semrush. Compare also our April 2025 coverage of results by another SEO firm, Brightedge: "Wikipedia probably not among the victims of Google's 'AI Overviews'".)

Given the many factors that affect Wikipedia usage, it can be challenging to find evidence for causal impact of a particular external change (a topic examined in depth some years ago in a foundational paper by the Wikimedia Foundation, see our coverage). The authors address this by exploiting the fact that Google rolled out its AI Overviews at different times in different countries:

We construct article–language–month panels from Wikimedia’s public clickstream data covering December 2023 through December 2024. Our main samples contain 499,927 matched English–German article pairs and 530,873 matched English–French article pairs. We use English Wikipedia as the treated edition because a substantial share of its readership comes from the United States, where AIO became part of the default search experience in May 2024. We use the German and French editions as controls because their readership is concentrated in European markets where AIO did not become part of the default search experience during our sample period [...] We then estimate a PPML difference-in-differences model that compares each English article’s monthly search traffic with that of its corresponding German or French version, controlling for article–language and calendar-month fixed effects to absorb persistent cross-language differences in baseline traffic and common monthly shocks.

(Difference-in-differences is a commonly used but somewhat fraught statistical method, which was also used in an earlier paper – coauthored by Nobel-winning economist Daron Acemoglu – to study the impact of the November 2022 launch of ChatGPT on Wikipedia. See our March 2025 review and the associated talk page.)

Among a few other robustness checks, the researchers also conducted a similar difference-in-differences comparison with Japanese Wikipedia, which resulted in a quite different estimate of "a 16.53% decline in English search traffic relative to Japanese." The authors deemphasize this result, observing that "the Japanese edition differs substantially from the main controls [i.e. French and German Wikipedia] in audience composition and [...] this comparison uses a shorter post-treatment window".

The paper is motivated by the importance of AIOs' economic effect on web publishers in general, and chose to study "Wikipedia as a large, transparent, and economically important empirical setting." In line with that, the authors also tried to convert their traffic reduction estimate into a rough assessment of how much financial damage English Wikipedia might sustain from AIOs if it were running ads in a typical setup:

an illustrative advertising-revenue equivalent of approximately $0.901M–$3.090M per month, or $10.82M–$37.08M per year if the estimated monthly effect persisted

The UW preprint has not yet been published in a peer-reviewed venue. Usually, the publications we cover here in preprint form don't see substantial changes when that happens. However, in this case the preprint's latest two versions (v6 from September 2, 2026, the current version reviewed here, and v5 from August 26) differ greatly from its four earlier revisions issued since it was first published in February 2026, including in the main results. Earlier revisions of the paper had claimed that "AIO exposure reduces daily traffic to English articles by approximately 15%", a much higher estimate than the 5.45% and 4.82% estimates offered in the current version. What's more, amplifying that divergence, that 15% estimate pertained to English Wikipedia pageviews overall, which the external search referrals studied by the current version form only a part of – about half (47.71%) according to a 2023 study^[2] of Wikipedia reader behavior (see also below), with over 90% of those coming from Google Search as of 2021.

The authors appear to have had second thoughts about their earlier methodology, going so far as to formally withdraw an earlier version of their preprint: "We decided to work on a new, more comprehensive sample of the data. As this could affect the conclusions, we decided to withdraw the paper until we have the final results". In v4 from May 12, this withdrawal notice was gone, while the abstract maintained the same 15% reduction in total pageviews claim that was removed later. However, in the appendix the authors added a different statistical approach. "As a robustness check to our baseline two-way fixed-effects difference-in-differences model in levels [...], we also estimate a multiplicative specification using Poisson pseudo-maximum likelihood ("PPML"). This resulted in a much smaller 3.5% decline estimate.

In the preprint's most recent versions (v5 and v6), the authors kept the PPML DiD method, but also switched to a different data source for their estimate: The Wikimedia Foundation's aforementioned public clickstream data, which offers the advantage of containing information about search engine referrals. Despite its name, this dataset does not contain full clickstream trajectories, but is redacted for privacy reasons, which prevented the authors from extending their findings to conclusions about English Wikipedia's overall traffic: "The public clickstream data identify only the immediate referrer for each article transition; they do not link an internal click to the external source that initiated the user’s session [...]. We therefore cannot determine how much internal traffic originates from search-referred sessions or quantify the resulting downstream effect."

A third difference concerns the choice of comparison languages: The earlier revisions relied on Hindi, Japanese, Portuguese and Indonesian Wikipedia to compare English Wikipedia's traffic against, whereas, as mentioned, the current version uses German and French instead.

The earlier, now withdrawn 15% estimate for pageviews overall had already circulated among Wikimedians and the wider public. For example, it was featured in a July 2026 post by the Institute for Public Relations. At this year's Wikimania in July, the annual "State of Wikimedia Research" overview presentation lauded it as "a rare example of a credible causal estimate of the effect of AI". (It cited a separate but apparently identical version of the preprint published on SSRN. That version now appears to have been completely deleted from public view, as the authors updated it and SSRN does not retain earlier revisions publicly, in contrast to ArXiv.) The Wikimania talk also highlighted interesting findings about heterogenous effects depending on the topic area, for example concluding the AIO-induced decline was smallest for STEM articles – but the authors completely omit such analyses from their current version.

While these divergences don't change the overall takeaway that it's fairly safe to assume that AIO have caused some traffic declines on English Wikipedia, they illustrate how such difference-in-difference estimates can depend greatly on researcher degrees of freedom: Aside from change in the exact outcome metric examined (pageviews overall or only the share of pageviews referred by search engines), the non-English Wikipedias chosen as controls and the statistical methods (PPML- vs. levels-based difference-in-difference) also appear to have substantially affected the reported numbers.

Other recent publications that could not be covered in time for this issue include the items listed below. Contributions, whether reviewing or summarizing newly published research, are always welcome.

From the abstract:^[2]

[...] we present the first systematic large-scale analysis of how readers browse Wikipedia. Using billions of page requests from Wikipedia’s server logs, we measure how readers reach articles, how they transition between articles, and how these patterns combine into more complex navigation paths. We find that navigation behavior is characterized by highly diverse structures. Although most navigation paths are shallow, comprising a single pageload, there is much variety, and the depth and shape of paths vary systematically with topic, device type, and time of day. We show that Wikipedia navigation paths commonly mesh with external pages as part of a larger online ecosystem, and we describe how naturally occurring navigation paths are distinct from targeted navigation in lab-based settings. Our results further suggest that navigation is abandoned when readers reach low-quality pages.

From the abstract:^[3]

This paper presents a comprehensive comparative analysis of AI Search and traditional web search (Google). Through a series of large-scale, controlled experiments across multiple verticals, languages, and query paraphrases, we quantify critical differences in how these systems source information. Our key findings reveal that AI Search exhibit a systematic and overwhelming bias towards Earned media (third-party, authoritative sources [including Wikipedia]) over Brand-owned and Social content, a stark contrast to Google’s more balanced mix.

A few example results from section 5.2 "Comparative Analysis":

In consumer electronics [...] Claude’s top domains included TechRadar, Tom’s Guide, and RTINGS, while GPT [GPT-4o] leaned on TechRadar, Tom’s Guide, and Wikipedia.[...]
[In the] automotive vertical [...] Claude [Claude 3.5 Sonnet] favored Consumer Reports, Car and Driver, and US News, while GPT mixed Wikipedia, Automoblog, and news outlets like AP. [...]
For electric cars, [...] Google’s top electric car sources included Reddit, Quora, and Facebook alongside Earned outlets like Car and Driver and government portals such as energy.gov. ChatGPT concentrated on encyclopedic and news publishers such as Wikipedia, AP News, and Reuters, with limited inclusion of automotive review sites. [...]
ChatGPT shows a relatively concentrated pattern, relying heavily on Wikipedia (∼19.7% of citations) and a small set of consumer-review and food sites (e.g., accio.com, Tasting Table, Sporked) (Fig. 38). Perplexity distributes citations across a broader long tail (the “Others” bucket ≈ 60.6%), with YouTube as the single most frequent domain (∼8.9%) and additional coverage from Tasting Table, The Daily Meal, Caffeine Informer, Statista, Visual Capitalist, TikTok, and others [...].

From the abstract:^[4]

[...] we examined the role of zoos and a television program featuring animated animals in shaping public interest in and support for animals including threatened species from 2011 to 2018 in Japan. [...] [T]he broadcast of a Japanese animated TV program featuring animals (Kemono Friends) increased the Google search volume and Wikipedia pageviews for animal species featured in the program. The total increases of search volume and Wikipedia pageviews were estimated to be approximately 4.66 million for 37 species and 1.06 million for 63 species, respectively."

From the abstract:^[5]

Objectives

To examine sourcing, framing, editorial dynamics and readership of autism-related Wikipedia pages in six languages (English, Spanish, French, Italian, Norwegian and Georgian).

Methods

We constructed a corpus of the main autism article in each language and used Python-based pipelines to extract references, revision histories (2004–2024) and monthly pageviews (2020–2024). [...]

Sourcing patterns differed markedly, with English and Italian relying more heavily on journals, French and Spanish on websites, Norwegian on institutional documents and Georgian under-sourced. Pageviews were highly unequal in absolute terms but, once normalised per million speakers, smaller language communities (Georgian, Norwegian and Italian) showed the highest relative demand. These high-demand editions nonetheless had the smallest editorial bases and fewest cumulative edits."

From the abstract of this opinion paper by former Wikimedia Foundation trustee Dariusz Jemielniak (subtitled "What might break it: AI parasitism, red tape,navel-gazing"):^[6]

This paper examines the sustainability of Wikipedia, arguing that despite its unique success as a non-profit, community-governed check on corporate information power, its future is far from assured. The author identifies three interrelated threats to the encyclopedia's relevance and viability:
The AI Extraction Crisis: Large Language Models (LLMs) are heavily trained on Wikipedia's high-quality content without attribution or contribution, severing the vital feedback loop that recruits new editors and posing an existential threat to the source of original human-generated knowledge.
The Enemy Within: The core editing community has developed a conservative culture marked by "red tape" and defensive bureaucratization. This resistance to technical innovations (like simplified editors or AI summaries) creates barriers for newcomers and prioritizes established processes over Wikipedia's underlying mission.
The Tragedy of the Commons, Revisited: Policy and rules are controlled by a self-selected group of long-term editors, leading to a fossilization of rules that may not align with the needs of the vast readership and potential casual contributors."