A recent University of Washington Working Paper quantifies something Wikipedia editors and marketers have long suspected: that Google’s AI Overviews are pulling search traffic away from the site. Although the estimate is around 5%, which is small compared to some of the more extravagant claims made by some claims in the SEO world, it is substantiated by a method that is good to understand.
This paper is by Mehrzad Khosravi and Hema Yoganarasimhan, last updated on September 2. It is not peer reviewed yet, so be wary of the numbers on the page. Google, for its part, disputes whether the underlying referral data can actually isolate its own effect from everything else happening in search.
Wikimedia has said that many people now reach its content through AI tools and search summaries without ever visiting wikipedia.org, but this paper aims to look at one aspect of that story: by how much did outside search referrals shift once AI overviews became the default in the U.S.?
How Researchers Used Wikipedia As A Testing Ground
Wikipedia proves to be a good subject for this kind of study because Wikimedia publishes monthly clickstream data aggregated by article and language edition. External search engines are bucketed together and pairs with low volume are removed, so that the researchers tested several assumptions about the missing values and the estimates were consistent despite that.
Google turned on AI Overviews by default in the U.S. in May 2024, while the paper states that Germany and France had not done so during the period of analysis. It was used as the comparison group, against which the effect size was estimated. Roughly 40% of traffic to English Wikipedia comes from the U.S., while the visitors to the German and French editions are predominantly from countries where AI Overviews had not been turned on by default. If a given week saw a lower percentage of English referrals compared to German and French for the same articles, it was considered to represent the effect size.
The period covered in the analysis was December 2023 to December 2024, with May 2024 representing the first month after the default switch. The researchers matched approximately 500,000 English-German article pairs and 530,000 English-French pairs and applied a statistical model designed to evaluate the percentage change in the count. They found a -5.45% change in referrals relative to German Wikipedia and -4.82% relative to French.
The effect size is equivalent to approximately 100 million fewer monthly search referrals for English Wikipedia or 1.2 billion per year. The latter number represents an annualized version of the December 2023–December 2024 monthly change, assuming that the pattern was consistent throughout the year. In both cases, the number refers to direct search referrals, not the overall number of visitors to Wikipedia.
It is notable that the previous iterations of the estimate used different methods and found a much larger effect. Earlier versions of the analysis used daily pageviews and the decrease attributed to AI Overviews was as high as 15%. The authors changed to monthly statistics and added a comparison to German and French editions in the revised version from August. It also narrowed the timeframe to 12 months in the latest revision from September. A similar calculation using the Japanese edition appears in the current version as a separate analysis, showing a steeper -16.5% decrease, but the authors have cautioned that it should be treated as a directional estimate due to the use of a different comparator and a smaller window of analysis.
Where The Estimate Falls Short
The paper acknowledges its own limitations, some of which could be significant. First, since Wikimedia’s public data does not separate Google from other search engines, Google has pointed out that the study cannot cleanly disentangle Google’s own product. Yoganarasimhan has countered that the use of other engines comprises only a small enough fraction of traffic that the timing of introduction of the AI Overviews would still drive the analysis.
Second, the study measures something narrower than it perhaps appears to at first glance: how many referral counts change after the introduction of AI Overviews by default, rather than the more immediate response of a user clicking on a particular search result. An earlier study in April, which the paper cites, saw an even larger drop-off in organic clicks for a similar second scenario, so the two are not directly comparable.
On the matter of referrals, a single user who follows up from a Google search link and reads several other Wikipedia articles during the same session would only count as a single referral. Most of English Wikipedia’s readership is not based in the United States, which could diminish the effect seen for other languages as well. The study cannot rule out the possibility that other events which took place in 2024 could have affected English searches differently from their impact on German or French searches.
The revenue figures should also be taken with caution. The authors estimate that a comparable site with advertising would have seen revenues reduced by somewhere between $10.8 million and $37 million a year at standard advertising rates. Since Wikipedia does not carry any advertisements, the number has to be taken as a hypothetical value, and not an actual loss for Wikimedia or its finances or a source of support for Google.
Finally, the paper does not examine long-term questions such as editing or donations. Whether or not a reduction in visits leads to a reduction in either is a separate question, not answered by this study.
Wikimedia’s Own Numbers Tell A Broader Story
Wikimedia’s internal data tell a different story than the one printed on the paper. Product director Marshall Miller noted last October that overall human pageviews across all Wikipedia language editions had fallen roughly 8% year over year. The figure was taken from Wikimedia’s overhaul of its bot detection and a subsequent review of the traffic data from early 2025, which revealed that a spike of popularity, most notably in Brazil, was driven by bots trying to get past the filters.
Miller cited generative AI and social media as catalysts of the fall, but also warned that the relabeled pageviews should be taken with a pinch of salt in light of the changes. The overall fall in pageviews is not a causal estimate for Wikipedia’s adoption of AI Overviews, so it shouldn’t be stacked on the paper’s figure.
Wikimedia’s draft 2026-27 fiscal year plan from April reveals that the drop in both pageviews and Google referrals, as well as an unprecedented bot traffic surge, is not going to be fixed soon. The figure shows that nearly 90% of Wikipedia’s visitors historically arrived via Google search, which is a sign of the channel’s importance.
Bots Are Adding Pressure From A Different Direction
Search referrals are not the sole issue that damages Wikipedia’s traffic picture. Wikimedia’s engineering team reported last year that the amount of bandwidth used for downloading images and media had risen by 50% since early 2024, driven primarily by bots scraping Wikimedia Commons to train AI models rather than readers. Bots comprised at least 65% of the most resource-intensive traffic and about 35% of pageviews.
A year later, the team stated that it was blocking or throttling approximately 30% of automated requests that ignored Wikimedia’s crawling policies, with one report estimating blocked or throttled requests at nearly 1.5 billion per day. Some bots disguise themselves as regular browsers or route through residential proxies to avoid detection. Wikimedia states that the objective is not to prevent reuse, but to guide large-scale users towards channels where they can identify themselves and receive appropriate access limits.
None of the bot activity is reflected in the referral numbers from the paper. A crawler scraping an article for training data, a search engine summarizing it in an AI answer, and a human clicking through to read it are treated as independent events, and only the last counts as a referral.
What This Means For Anyone Watching Search Traffic
A Pew Research study has already demonstrated that users are unlikely to follow sources cited within AI summaries. An analysis of nearly 69,000 Google searches conducted by a sample of U.S. adults found that traditional search result clicking dropped to 8% when an AI summary was present, compared to 15% when it was not. Just 1% of visits including an AI summary involved a click on a source cited, with Wikipedia, YouTube and Reddit accounting for 15% of those mentions.
Google contends that AI Overviews direct users to a wider range of sites, and Search VP Liz Reid stated last year that overall organic click volume had been broadly stable compared to previous periods, although she gave no specific figures.
Wikimedia has been lobbying AI firms to attribute content correctly and use Wikimedia Enterprise, its paid high-volume access product, rather than directly scraping the site. The Foundation has stated this model represents payment for infrastructure, not content licensing, with the material remaining freely available under its open license. Wikimedia Enterprise revenue is capped at 30% of Wikimedia’s overall revenue and brought in $8.3 million in the 2024-25 fiscal year, with Google, Amazon, Meta, Microsoft, Mistral AI, Perplexity and GNOMI now named partners.
The Bigger Picture Is Still Forming
Wikipedia is an unusually compelling case study due to the amount of clickstream data published for the site in different languages, but that same strength is a weakness for considering the broader implications. Other websites typically do not have such content or the capacity to compare across their various language editions, and it is still unclear if other publishers would be able to adopt a similar model to Wikimedia Enterprise.
This paper has been significantly revised several times since its initial publication in February, and so these figures should be considered more as a current working draft than a final conclusion. In addition to the referral study, Wikimedia’s internal traffic data, and the bot traffic issue, the overall impact of search algorithms on Wikipedia and other potential entrants should be regarded as a multifaceted issue.
FAQs
Q1: How much did AI Overviews reduce Wikipedia search traffic, according to the paper?
The University of Washington study estimates a 5% reduction in monthly search referrals to English Wikipedia, compared to Wikipedia’s German and French counterparts for the same period.
Q2: Who wrote this paper?
“Research by Mehrzad Khosravi and Hema Yoganarasimhan.” This is a working paper, not yet peer reviewed, that has been updated several times since its initial publication in February.
Q3: Does Google refute these claims?
Google argues that Wikimedia’s aggregated public data combines performance across search engines, and therefore cannot isolate Google’s contribution to a traffic decline.
Q4: How did they calculate this number?
By analyzing referral trends for comparable articles in English, German, and French Wikipedias, since the U.S. received the AI Overviews rollout by default in May 2024, while Germany and France had not yet received the feature.
Q5: Why is this number so different from numbers in previous versions of this paper?
Earlier versions of this paper used daily pageview statistics and found a nearly 15% decrease, whereas this paper uses monthly search referral estimates and incorporates findings from English, German, and French Wikimedias to find a smaller average reduction of 5%.
Q6: Is the amount of lost revenue accurately represented in this figure?
The $10.8 million to $37 million number should not be taken as an amount actually lost by Wikimedia. This number represents a hypothetical revenue reduction for an advertising-supported site using a standard revenue rate across all traffic, whereas Wikipedia does not carry advertisements and therefore this number has no bearing on Wikipedia’s actual finances.
