What Wikipedia Reveals About AI Overviews and Web Traffic

For years, the "zero click search" has been a looming threat for publishers. The idea is simple but devastating: if a search engine can answer a user's question directly on the results page, the user has no reason to click through to the source. This turns the web from a network of interconnected sites into a series of silos where the platform captures all the value. A useful companion note is LLM Traffic Converts Differently, because it looks at a nearby part of the same system.

We often talk about this in the abstract, but because Wikipedia is one of the few entities that shares massive amounts of granular traffic data, we finally have a way to measure the actual impact of AI Overviews. It is a glimpse into a future where the relationship between content creators and discovery platforms is fundamentally rewritten. This connects with AI Overviews Vs. Featured Snippets when the same signal needs a clearer operating decision.

Using Wikipedia as a Proxy for Search Trends

A recent working paper from the University of Washington suggests that Google's AI Overviews have had a measurable impact on traffic. Specifically, the researchers estimate that monthly search referrals to the English version of Wikipedia dropped by roughly 5% after AI Overviews became the default in the United States in 2024.

The study, authored by Mehrzad Khosravi and Hema Yoganarasimhan, is a working paper and has not yet undergone peer review. It is worth noting that this 5% figure is a revision; an earlier version of the paper published in February suggested a more significant drop. This adjustment shows how difficult it is to isolate a single feature in a complex ecosystem.

Google has contested these findings, arguing that the data doesn't specifically isolate AI Overviews from other search changes. Meanwhile, Wikimedia points out a broader reality: a vast amount of the work produced by its volunteers is being consumed via AI tools and search summaries without the user ever landing on wikipedia.org.

Expert Interpretation: The tension here is between "referrals" and "consumption." For a site like Wikipedia, which doesn't rely on ad revenue, a drop in referrals is a blow to the visibility of the community. For a commercial publisher, however, a 5% drop in referrals is a direct hit to the bottom line. The tradeoff is clear: the user gets a faster answer, but the source loses the opportunity to engage the visitor.

The Mechanics of the Traffic Test

Wikipedia is a unique test case because Wikimedia releases monthly clickstream data at the article level across multiple languages. This allows researchers to compare how the same information is accessed in different regions.

The researchers looked at the period between December 2023 and December 2024. They used a "difference in differences" model, comparing English Wikipedia traffic (which is heavily influenced by the U.S. market where AI Overviews were rolled out in May 2024) against German and French editions. Since AI Overviews weren't the default in Germany or France during the sample period, those language editions served as a control group.

By matching nearly half a million article pairs between English and German, and over 500,000 between English and French, the authors could see if English referrals fell specifically when the AI feature went live, while the other languages remained stable.

Expert Interpretation: This methodology is as strong as it can be given the available data, but it relies on the assumption that the U.S. market behaves as a proxy for English speakers globally. The decision to use language editions as regional proxies is a clever workaround, but it reminds us that we are seeing a fragmented rollout, not a global shift.

The Limits of the Data

There are significant caveats to these findings. First, Wikimedia's public data groups all external search engines together. It doesn't separate Google from Bing or DuckDuckGo. While the researchers argue that Google dominates the traffic share enough that the timing of the AI Overview rollout is the primary driver, it remains a point of contention.

this study measures referrals from human readers. If a user clicks a link from Google and then navigates through three more Wikipedia pages, it is recorded as a single visit. Because a large portion of English Wikipedia's audience is outside the U.S., the overall impact of the U.S. rollout is "watered down" in the final percentage.

Perhaps most striking is the hypothetical financial cost. The authors calculated that if a traditional ad supported website experienced a similar drop in traffic, the annual revenue loss could range between $10.82 million and $37.08 million. Since Wikipedia is non profit, this is a theoretical exercise, but it highlights the existential risk for commercial media.

Expert Interpretation: When analyzing these numbers, you have to distinguish between a "visit" and a "session." The fact that the drop is only 5% might seem small, but in the world of high volume search, a 5% shift in the baseline is massive. The real question for site owners is whether the remaining 95% of traffic is higher quality or if the "easy" informational queries are being completely cannibalized.

A Broader Decline in Human Pageviews

While the University of Washington paper focuses on search referrals, the Wikimedia Foundation has observed a more general decline. In October 2025, product director Marshall Miller noted that overall human pageviews across all languages dropped by about 8% compared to the previous year.

This discovery came after Wikimedia improved its bot detection. They found that a spike in traffic, particularly from Brazil, was actually driven by sophisticated bots designed to mimic human behavior to bypass detection. Miller attributed the genuine decline in human traffic to the rise of generative AI and the shift toward social media platforms for information.

The Foundation's internal planning for the 2026 to 27 fiscal year suggests this isn't a temporary dip. They expect decreasing pageviews and fewer Google referrals to be a continuing trend, noting that nearly 90% of their visitors have historically come from Google search.

Expert Interpretation: This suggests that the problem is larger than just one Google feature. We are seeing a structural shift in how humans seek information. The tradeoff is convenience versus depth. If users are moving toward social media and AI summaries, they are trading the complete, cited nature of a Wikipedia article for a condensed, potentially less accurate snippet.

The Rise of Machine Demand

While human traffic is dipping, machine traffic is exploding. Wikimedia engineers reported a 50% increase in bandwidth for downloading images and media files since January 2024. This isn't caused by people looking at pictures, but by bots scraping Wikimedia Commons to train AI models. The same pattern also shows up in 000x Human Traffic in 5 Years, where the practical question is how the signal becomes visible.

Bots now account for at least 65% of the most expensive traffic hitting core data centers and roughly 35% of all pageviews. To combat this, Wikimedia has been blocking or limiting about 30% of these requests, which amounts to roughly 1.5 billion blocked or throttled requests per day.

The goal isn't to stop the reuse of data entirely, but to force large scale users to identify themselves and use official channels rather than using proxies or imitating home internet connections to scrape data covertly.

Expert Interpretation: This is the great irony of the AI era: the very tools that are reducing human referrals to the source are simultaneously increasing the mechanical demand for the source's data. The "value" is being extracted at the infrastructure level (scraping) while the "visibility" is being erased at the user level (AI summaries).

The Click Through Gap

Data from Pew provides a sobering look at how users actually interact with AI summaries. In a study of nearly 69,000 searches from U.S. adults, Pew found that when an AI summary was present, users clicked on traditional search results only 8% of the time. When no summary was present, that number jumped to 15%.

Even more concerning is the "citation" rate. Only 1% of visits to pages containing an AI summary resulted in a click on a source cited within that summary. While Wikipedia, YouTube, and Reddit are frequently cited, the vast majority of users simply read the summary and leave.

This contradicts some of Google's internal narratives. Google has claimed that AI Overviews guide users to a broader range of websites, and some internal reports suggest organic click volume has remained "relatively stable." However, the Pew data suggests that while the *volume* might be stable, the *behavior* has shifted away from clicking the sources that provided the answer.

Expert Interpretation: This is the "attribution gap." Being cited in an AI Overview is not the same as receiving a visit. For a content creator, a citation without a click is a brand mention, not a lead or a reader. The decision for publishers now is whether to optimize for "LLM visibility" (hoping for the 1% click) or to build direct to consumer channels that bypass search engines entirely.

Why Wikipedia is a Unique Case

It is important to remember that Wikipedia is not a typical website. Its ability to provide matched article pairs across different languages allows for a level of scientific rigor that most publishers cannot replicate. A news site or an e-commerce store cannot simply compare a "German version" of a product page to an "English version" to isolate the effect of a Google update.

We don't know if news sites or retailers are seeing the same 5% drop or if they are experiencing something far more severe. Wikipedia has a unique lever: Wikimedia Enterprise. This is a paid service for commercial entities to access their data, providing a potential revenue stream that most independent publishers do not have.

Expert Interpretation: The "Wikipedia effect" is a baseline, not a ceiling. If a site with the global authority of Wikipedia is seeing a decline in referrals, smaller sites with less authority are likely feeling the impact more acutely. The lesson here is that authority does not protect you from a change in the fundamental delivery mechanism of the web.

The Evolving Landscape

The research we are seeing is a snapshot of a moving target. The University of Washington paper has been revised multiple times, and the figures are estimates, not finalities. As AI search evolves from a "summary" to a "reasoning engine," the way traffic flows will continue to shift.

The overarching trend is clear: the search engine is transitioning from a map that points you to a destination to a destination in itself. For those of us who build for the web, the challenge is no longer just about ranking in search, but about providing value that is so unique that a user is willing to leave the AI summary to find the original source.