TL;DR
- University of Washington researchers estimate default AI Overviews cut monthly search referrals to English Wikipedia by 5.45% vs German and 4.82% vs French.
- Earlier versions of the working paper, which is not peer reviewed, used daily pageviews and reported about 15%; version 6 is dated 2 September 2026.
- Google told UOL Tilt the data cannot isolate Google because Wikimedia groups all external search engines together.
- The source says the study does not establish whether news sites, retailers or other kinds of site see a similar effect.
A University of Washington working paper by Mehrzad Khosravi and Hema Yoganarasimhan estimates that Google's AI Overviews cut monthly external-search referrals to English Wikipedia by about 5% after the feature became the default in the United States in May 2024. Search Engine Journal reported on 11 September 2026 that the paper was last revised on 2 September, has not been peer reviewed, and replaces an earlier version from February that put the decline at about 15%. Google disputes whether the data can isolate its feature.
What the paper measured
The study uses a feature of Wikipedia that almost no other website has: the same article exists in several language editions, and the Wikimedia Foundation publishes monthly clickstream data for each one at article level. That lets the authors compare the same topic across audiences that were, and were not, exposed to AI Overviews by default.
Google switched AI Overviews on by default in the United States in May 2024. According to the paper, it had not done so in Germany or France during the sample period. The authors use language editions as a stand-in for regions. During the sample period about 40% of traffic to English Wikipedia came from the United States, while traffic to the German and French editions came mostly from countries without default AI Overviews.
The logic is a difference-in-differences comparison. If, after May 2024, search referrals to an English article fell relative to referrals to the same article in German or French, that gap is the estimate of the AI Overviews effect. The data runs from December 2023 to December 2024, with May 2024 treated as the first month after the change.
The panel is large. It contains 499,927 matched English-German article pairs and 530,873 English-French pairs. The authors used a Poisson pseudo-maximum likelihood difference-in-differences model, a method suited to estimating percentage changes in counts.
The numbers, and how they changed between versions
The current version estimates that default AI Overview availability reduced monthly external-search referrals to English Wikipedia by 5.45% relative to German and by 4.82% relative to French. Applied to the English edition as a whole, that works out to about 100.27 million fewer search referrals a month, or about 1.20 billion a year.
Both of those totals need care. The monthly figure is model-based, and the annual figure assumes the effect stays constant through the year. Both count only direct referrals from search engines, not all the traffic Wikipedia receives.
The revision history is part of the story, which is why Search Engine Journal led with it. Earlier versions of the paper used daily pageviews and reported a decline of about 15%. Version 5, dated 26 August 2026, switched the outcome to monthly search referrals and added the German and French controls. Version 6 followed on 2 September. The current version also includes an English-Japanese check that shows a 16.53% decline, but it uses a shorter time frame and a different comparison group, and the authors describe it as directional support only.
The table below lays out the figures as the source reports them, so each can be read with its own caveat.
| Figure | Value | Caveat in the source |
|---|---|---|
| Referral decline vs German edition | 5.45% | Main estimate, working paper, not peer reviewed |
| Referral decline vs French edition | 4.82% | Main estimate, same design |
| Fewer English referrals per month | About 100.27 million | Model-based |
| English-Japanese check | 16.53% | Shorter window, different control, directional only |
| Earlier versions (daily pageviews) | About 15% | Replaced by the current method |
What the estimate does not prove
Wikimedia's public clickstream data groups all external search engines together, and the paper acknowledges this. Google told the Brazilian outlet UOL Tilt that the grouping prevents the analysis from isolating Google. Yoganarasimhan's response, as reported, is that other search engines account for only a small share of traffic and that the analysis relies on the timing break created by the AI Overviews rollout.
The estimate also answers a narrow question. It measures what happened to referrals when English-language readers were in an environment where AI Overviews were on by default. It is not the same as measuring the drop on the specific searches where an AI Overview actually appeared, which an earlier experiment Search Engine Journal covered in April found to be larger. The two findings are related but not interchangeable.
Several other limits are spelled out in the source:
- The outcome counts referrals from search engines by human readers. A reader who arrives from Google and then reads three more articles counts as one referral.
- Most visitors to English Wikipedia come from outside the United States, which dilutes the measured effect.
- The design cannot fully rule out other events in 2024 that affected English search but not German or French search.
- The paper does not look at long-term effects, so it says nothing about editing activity, donations, or whether fewer visits could affect future content.
The paper also includes a revenue illustration that is easy to misread. The authors calculate that a comparable ad-supported website might lose between $10.82 million and $37.08 million a year at typical ad rates. Wikipedia does not run ads, so this is purely hypothetical. It does not describe Wikimedia's real finances or any money moving to Google.
The wider picture Wikimedia and Google describe
The paper is one data point in a longer argument about AI answers and web traffic, and the source places it next to several others. None of them can be added to the 5% figure, because they measure different things.
Wikimedia's own pageview trend
In October 2025, Wikimedia Foundation product director Marshall Miller said total human pageviews across all Wikipedia languages were down about 8% compared with the same period in 2024. That figure came after Wikimedia improved its bot detection and reclassified traffic from March to August 2025, when a spike of apparently human visits, mostly appearing to come from Brazil, turned out to be largely bots built to evade detection. Miller attributed the decline to generative AI and social media, and also cautioned that the relabelled data should be read carefully because detection methods changed. It is a count of overall human pageviews, not a causal estimate about AI Overviews.
By April 2026, according to the source, the Foundation's draft plan for the 2026-27 fiscal year said falling pageviews, fewer referrals from Google and what it called unprecedented bot traffic are expected to continue rather than prove temporary. The plan also notes that close to 90% of Wikipedia visitors have historically arrived from Google search.
Machine demand that never shows up as a referral
In April 2025, Wikimedia engineers reported that bandwidth used to download images and media had risen 50% since January 2024, mostly from bots scraping Wikimedia Commons for AI training. They said bots accounted for at least 65% of the most expensive traffic, the kind that reaches core data centres, and about 35% of pageviews. A year later the team said it was blocking or limiting around 30% of automated requests from crawlers that ignore its policies, with a chart note putting blocked or throttled requests at about 1.5 billion a day.
That matters for reading the paper. A crawler copying an article for training, a search engine quoting it inside an AI answer, and a reader clicking through from that answer are separate events. Only the last one is a human external-search referral, which is the only thing the paper counts.
Pew's click data and Google's position
The source also cites a Pew analysis of 68,879 Google searches from March 2025, drawn from the browsing of 900 US adults. When a visit included an AI summary, people clicked a traditional result 8% of the time, compared with 15% when there was no summary. Only 1% of visits to pages with an AI summary led to a click on a source inside it. Wikipedia, YouTube and Reddit together made up 15% of cited sources, though Pew did not give Wikipedia's individual share.
Google's stated position is different. In an August 2025 blog post, Liz Reid wrote that total organic click volume had stayed "relatively stable" year on year and that AI Overviews send people to a broader range of websites, without giving exact figures.
Paid access instead of referrals
In November 2025 the Foundation asked AI developers to credit Wikipedia properly and to access it through Wikimedia Enterprise, its paid high-volume API service. Enterprise charges for access, not for the content, which stays freely available under open licences. Lane Becker, who runs the programme, said in July that Enterprise income is capped at 30% of the Foundation's annual revenue, and the unit reported $8.3 million in revenue for the 2024-25 fiscal year. In January, Wikimedia Enterprise said Google is an existing customer and named Amazon, Meta, Microsoft, Mistral AI and Perplexity as partners. Wikimedia describes this as a paid service for high-volume access and support, not compensation tied to lost referrals.
Why a Wikipedia result does not transfer directly to your site
The same properties that make Wikipedia measurable limit how far the result travels. Most websites cannot compare the same page across language editions served to audiences with and without AI Overviews, and the paper does not establish whether news sites, retailers or other kinds of site would see a similar effect. The source is explicit on that point.
There is also a difference in what kind of content is being summarised. Wikipedia is almost entirely informational, the category of query where an AI summary is most likely to satisfy the searcher without a click. A commercial page answering "which supplier should I contact" sits in a different position. Treat the 5% as evidence that the effect exists and can be measured at scale, not as a benchmark for your own traffic.
What this means for Thai marketers
The study says nothing about Thailand or Thai-language search, so any local reading is analysis rather than a finding. With that caveat, three practical points follow.
First, the method is a useful template. If your site runs in Thai and English, or serves several markets, you have a rough version of the same natural experiment. Comparing clicks for equivalent pages across languages or markets before and after an AI feature reaches one of them is more informative than looking at a single traffic line and guessing.
Second, informational content is the most exposed. Blog posts that answer definitional questions, the "what is" pages many Thai businesses publish for SEO, are the closest analogue to a Wikipedia article. Those pages still matter, because being the source an AI answer draws from is its own kind of visibility, but their value increasingly shows up as citations and brand exposure rather than sessions. That is the problem generative engine optimisation is meant to address.
Third, separate human referrals from machine demand in your own analytics. The Wikimedia numbers show how much crawler activity can hide inside or alongside traffic figures. If server load or raw hits rise while Search Console clicks fall, both can be true at once. A sound SEO programme now needs reporting that treats clicks, impressions, AI citations and bot traffic as different measures.
FAQ
Did AI Overviews cut Wikipedia's traffic by 5%?
Not exactly: the paper estimates about a 5% drop in monthly external-search referrals to English Wikipedia, not in all traffic. The figures are 5.45% against the German edition and 4.82% against the French edition, from a working paper that has not been peer reviewed.
Why did the estimate fall from 15% to about 5%?
Because the authors changed the method. Earlier versions measured daily pageviews and found about a 15% decline; version 5 in August 2026 switched to monthly search referrals and added German and French comparison groups, which produced the lower figure.
Does Google accept the finding?
No. Google told UOL Tilt that Wikimedia's data groups all search engines together and so cannot isolate Google, while one of the authors said other engines carry only a small share of traffic and the design relies on the timing of the rollout.
Does this apply to Thai websites?
The source does not say. The study covers English, German, French and a Japanese check on Wikipedia only, and the source states that it does not establish effects for other kinds of sites. Thai site owners should measure their own before-and-after click data instead of borrowing the figure.
Should we stop publishing informational content?
No, but measure it differently. Informational pages are the most likely to be summarised, so their value shows up more in AI citations and brand visibility, and less in clicks, than it did before AI answers became common.
The paper is still a working paper that has been rebuilt once and revised several times since February, so its figures are the latest estimates rather than final numbers. If your organic traffic is drifting and you want to know how much of it is AI answers, and what to do about it, our content marketing team can help you work out which pages to measure first.
