A College of Washington working paper now estimates that Google’s AI Overviews lowered month-to-month search referrals to English Wikipedia by roughly 5% after the characteristic grew to become the U.S. default in 2024. The paper, by Mehrzad Khosravi and Hema Yoganarasimhan, was final up to date on September 2 and hasn’t been peer reviewed.
The newer estimate took the place of a bigger one which the identical authors revealed again in February, and this revision is all a part of the story. Google disputes whether or not the mixed referral metric can isolate its characteristic.
Wikimedia says many individuals entry its volunteers’ work via AI and search instruments with out visiting wikipedia.org. The revised paper explores a particular side of this, taking a look at how default AI Overview availability influenced referrals from exterior search engines like google and yahoo.
Wikipedia Affords A Uncommon Visitors Take a look at
Wikimedia releases month-to-month clickstream knowledge on the article stage for numerous Wikipedia language editions, equivalent to English, German, and French. These recordsdata categorize exterior search engines like google and yahoo collectively and omit referrer-article pairs with low quantity. The authors examined a number of values for the suppressed post-treatment counts and reported practically an identical estimates.
Google switched AI Overviews on by default in america in Could 2024. In response to the paper, it hadn’t but accomplished so in Germany or France through the pattern interval. The authors see language editions as a approach to signify totally different areas. In the course of the pattern interval, about 40% of visitors to English Wikipedia got here from america. In the meantime, visitors for German and French editions principally got here from international locations with out default AI Overviews. If, after Could 2024, English referrals fell in comparison with German and French referrals for a similar articles, that distinction is the estimate.
The information covers the interval from December 2023 to December 2024. The authors take into account Could 2024, when the U.S. rollout began, as the primary post-treatment month. The panel contains 499,927 matched English-German article pairs and 530,873 English-French pairs. They used a Poisson pseudo-maximum chance difference-in-differences mannequin, which helps estimate proportion adjustments in counts. The authors estimate that default AI Overview availability lowered month-to-month external-search referrals to English Wikipedia by 5.45% relative to German and 4.82% relative to French.
For the English version as an entire, the estimate works out to round 100.27 million fewer search referrals every month, totaling about 1.20 billion yearly. The month-to-month quantity is model-based, and the yearly estimate assumes this impact stays constant all year long. Each figures account just for direct referrals from search engines like google and yahoo, not all visitors Wikipedia receives.
Earlier variations used each day pageviews and reported a decline of about 15%. The paper then switched to month-to-month search referrals and added German and French controls in v5 on August 26, with extra updates in v6 on September 2.
An English-Japanese examine within the present model reveals a 16.53% decline, utilizing a shorter timeframe and a distinct comparability group. The authors word that this result’s directional assist solely.
What The Estimate Does Not Show
Wikimedia’s public knowledge lumps all exterior search engines like google and yahoo collectively, a limitation the paper acknowledges. Google instructed UOL Tilt that this prevents the evaluation from isolating Google, however Yoganarasimhan stated different engines deal with solely a small portion of visitors, and the evaluation makes use of the timing break created by the AI Overviews rollout.
This estimate addresses a particular query. It measures what occurred to referrals when English-language readers have been in an surroundings the place AI Overviews have been on by default. The sooner experiment I mentioned in April confirmed an even bigger drop when AI Overviews appeared throughout searches. So, whereas associated, they’re totally different findings.
It counts referrals from search engines like google and yahoo made by human readers, so if somebody arrives from Google after which reads three extra articles, it counts as one go to. Most of English Wikipedia’s guests come from exterior the U.S., which waters down the measured impact. Plus, the design can’t fully rule out different 2024 occasions affecting English search however not German or French search.
The income estimate needs to be seen with care. The authors calculate {that a} comparable ad-supported web site would possibly lose someplace between $10.82 million and $37.08 million annually at typical advert charges. Since Wikipedia itself doesn’t show adverts, this determine is only hypothetical and doesn’t signify Wikimedia’s actual funds or any funds moved to Google.
Lastly, the paper doesn’t look into long-term results, so it doesn’t present insights into modifying exercise, donations, or whether or not fewer visits would possibly affect future content material creation.
Wikimedia Sees A Wider Visitors Change
The Wikimedia Basis’s knowledge presents a distinct perspective. In October 2025, product director Marshall Miller shared that total human pageviews throughout all Wikipedia languages had decreased by about 8% in comparison with the identical interval in 2024. This data got here after Wikimedia improved its bot detection and reclassified visitors from March via August 2025. Visits spiked and appeared to return from actual customers, primarily from Brazil, however a lot of that visitors was pushed by bots designed to bypass detection.
Miller attributed the decline to generative AI and social media platforms. Nevertheless, he additionally identified that the relabeled knowledge needs to be seen with some warning as a result of detection strategies have modified over time. That is about total human pageviews, not a causal estimate about AI Overviews, so it will probably’t be added to the paper’s 5% estimate.
I mentioned this replace again in October. By April, the Basis’s draft plan for the 2026-27 fiscal yr highlighted that lowering pageviews and fewer referrals from Google, together with bot visitors the plan calls unprecedented, are anticipated to proceed somewhat than be short-term points. The plan additionally notes that just about 90% of Wikipedia guests have historically come from Google search.
Machine Demand Retains Rising
In April 2025, the Wikimedia Basis’s engineers noticed that the bandwidth for downloading pictures and media recordsdata went up by 50% since January 2024. Most of this enhance got here from bots scraping Wikimedia Commons for AI coaching somewhat than from common readers. In addition they famous that bots accounted for at the very least 65% of the costliest visitors, which reaches the core knowledge facilities, and about 35% of pageviews.
A yr later, the crew talked about they have been blocking or limiting round 30% of those automated requests from crawlers that ignore their insurance policies, and a chart word put blocked or throttled requests at about 1.5 billion a day. The report explains that some bots imitate common browsers or use proxies by way of dwelling web connections. The intention is to information large-scale reusers towards channels the place they’ll determine themselves and obtain greater limits, somewhat than to cease reuse.
None of this exercise seems in referral counts. Whether or not it’s a crawler copying an article for coaching, a search engine quoting it in an AI response, or a reader clicking via from that response, all these actions are thought of separate occasions. Solely the click-throughs are counted as human external-search referrals within the paper’s end result.
Quotation, Visits, And Enterprise Entry
Pew’s evaluation discovered that folks hardly ever clicked on sources inside AI summaries. They checked out 68,879 Google searches from March 2025, which got here from the looking exercise of 900 U.S. adults. They repeated the searches from April 7 to April 17 to seize the outcome pages, which means the AI-summary classification is predicated on these later knowledge. When visits included an AI abstract, folks clicked on conventional outcomes 8% of the time, in comparison with 15% when no abstract was there. Only one% of visits to pages containing an AI abstract led to a click on on a supply inside it. Wikipedia, YouTube, and Reddit collectively made up 15% of cited sources, though Pew didn’t specify Wikipedia’s actual share.
Google’s personal paperwork say that AI Overviews information customers to a broader vary of internet sites. Liz Reid’s August 2025 weblog publish shared that the full natural click on quantity stayed “comparatively secure” in comparison with the earlier yr, regardless that she didn’t present actual numbers. She additionally talked about this on Bloomberg’s Odd Heaps podcast in April.
In November 2025, the Basis urged AI builders to correctly credit score Wikipedia of their outputs and requested them to entry it via Wikimedia Enterprise. They imagine giving credit score encourages a cycle wherein readers change into editors and donors, however in addition they warn that fewer visits would possibly result in fewer volunteers and donors.
Wikimedia Enterprise is the Basis’s paid service that gives high-volume API entry, together with contracts, uptime ensures, assist, and structured knowledge feeds. It expenses for entry, not the content material itself, which stays freely accessible via public channels underneath open licenses, topic to its license phrases. Lane Becker, who oversees this system, shared in July that this mannequin is predicated on entry somewhat than licensing. The revenue from Enterprise is capped at 30% of the Basis’s annual income, he stated, and the unit reported $8.3 million in income for the 2024-25 fiscal yr.
In January, Wikimedia Enterprise introduced that Google is an present buyer, and for the primary time, Amazon, Meta, Microsoft, Mistral AI, and Perplexity have been launched as companions. In July, Wikimedia Enterprise introduced GNOMI as one other accomplice. Wikimedia describes Enterprise as a paid service designed for high-volume entry and assist, somewhat than compensation tied to referral losses.
Why Wikipedia Is An Uncommon Take a look at
The options that make Wikipedia measurable additionally restrict how extensively its findings can be utilized. Wikimedia shares knowledge on articles in lots of languages, and the paper compares matching pages throughout totally different language editions. Most web sites don’t have the flexibility to make these sorts of comparisons.
The paper doesn’t set up whether or not information websites, retailers, or different kinds of web sites would expertise comparable results. Moreover, the sources don’t deal with whether or not different organizations might create a paid entry service much like Wikimedia Enterprise.
Trying Forward
This paper continues to be a working paper, rebuilt as soon as and revised a number of instances since February, so the figures are the most recent estimates, not closing numbers. We’ll replace if new figures from the authors, Wikimedia, or Google change the image.
Extra Sources:
Featured Picture: Kateryna Deineka/Shutterstock
