The rise of generative artificial intelligence has fundamentally altered how humanity navigates information, shifting the paradigm from hunting across a network of web links to receiving instantaneous, synthesized answers. Now, new empirical research offers a clearer—though still heavily contested—window into the human cost of this transition for the web’s largest repository of human knowledge.
According to a revised working paper from the University of Washington, Google’s AI Overviews reduced monthly search referrals to English Wikipedia by roughly 5% following the feature’s default deployment in the United States in 2024. Authored by Mehrzad Khosravi and Hema Yoganarasimhan, the study—last updated on September 2—has not yet undergone formal peer review, yet it provides one of the most rigorous econometric attempts to date to measure the elusive "zero-click search" phenomenon.
While Google disputes whether the aggregated data can definitively isolate its specific interface changes, the findings underscore an existential reality for digital publishers: as search engines absorb the content to answer queries on the spot, the traditional funnel of web traffic is drying up.
Main Facts
The core finding of the University of Washington working paper is stark: the rollout of default AI Overviews in the U.S. in May 2024 precipitated a measurable decline in organic, human search traffic directed toward English-language Wikipedia pages.
- The Quantified Drop: Using a difference-in-differences econometric model, the authors estimate that default AI Overview availability reduced monthly external search referrals to English Wikipedia by 5.45% relative to German-language editions and 4.82% relative to French-language editions, where the feature had not yet rolled out during the sample window.
- The Scale of Loss: For the English edition as a whole, this translates to approximately 100.27 million fewer search referrals every month, accumulating to roughly 1.20 billion vanished visits annually.
- The Methodology: The researchers analyzed nearly half a million matched article pairs across languages between December 2023 and December 2024, leveraging Wikimedia’s monthly clickstream data to control for global trends.
- The Caveats: The data exclusively tracks direct human referrals from search engines, omitting bot traffic, internal link navigation, and the broader macro-decline in desktop and mobile web browsing. Furthermore, because Wikimedia aggregates all external search engines into a single data bucket, Google cannot be isolated independently within the raw logs, though the authors argue competing engines command a negligible market share.
Chronology of an Information Shift
To understand how researchers arrived at these figures, it is necessary to retrace the timeline of AI search integration and the evolving body of data surrounding it.
- May 2024: Google officially switches AI Overviews on by default for U.S. users, permanently altering the search results page (SERP) landscape by placing conversational answers above traditional blue links.
- February 2025: Khosravi and Yoganarasimhan publish the initial iteration of their working paper, which utilized daily pageview metrics and initially pointed toward a much steeper drop of approximately 15%.
- April 2025: Wikimedia Foundation engineers report an explosive 50% increase in media bandwidth consumption since January 2024, driven almost entirely by automated bots scraping content for AI training rather than human readers. Concurrently, independent field studies—such as analysis published by Search Engine Journal—begin documenting broader organic click losses across the web.
- July 2025: Pew Research releases a comprehensive behavioral study of nearly 69,000 U.S. Google searches, revealing that when an AI summary appears, user interaction with traditional organic results plummets, and clicks on sources within the AI summary itself remain statistically negligible at roughly 1%.
- August–September 2025: The Washington researchers revise their methodology in versions 5 and 6 of their working paper, pivoting from daily pageviews to monthly search referrals and introducing German and French language editions as rigorous geographical controls. This refines the estimated referral decline to the current ~5% figure.
- October 2025: Wikimedia product director Marshall Miller publicly discloses that overall human pageviews across all languages dropped by roughly 8% compared to the previous year, pointing the finger squarely at generative AI and social media shifts.
- January 2026: Wikimedia Enterprise announces an expansion of its corporate partnerships, welcoming tech giants Amazon, Meta, Microsoft, Mistral AI, and Perplexity alongside existing customer Google to manage high-volume API access.
- March–July 2026: Wikimedia infrastructure reports reveal that the foundation is actively blocking or throttling roughly 1.5 billion malicious or policy-violating automated crawler requests per day, highlighting the widening chasm between human readership and machine demand.
Supporting Data: What the Numbers Reveal
The econometric modeling behind the University of Washington paper relies heavily on natural experiments enabled by Wikimedia’s monthly clickstream data. Because Wikimedia releases article-level clickstream files categorizing external search traffic, researchers were able to exploit a geographical discrepancy: while U.S. users experienced default AI Overviews starting in May 2024, users accessing German and French editions largely resided in regions where the feature had not yet launched.
By analyzing 499,927 matched English-German article pairs and 530,873 English-French pairs using a Poisson pseudo-maximum likelihood model, the authors isolated the treatment effect. The resulting figures—a 5.45% relative drop against German controls and 4.82% against French controls—provide a statistically robust baseline.
However, auxiliary data paints an even starker picture of user behavior:
- The Pew Research Insight: Pew’s tracking of 900 U.S. adults found that users click on traditional search results 15% of the time when no AI summary is present. That figure drops by nearly half—to 8%—when an AI overview occupies the screen. Crucially, only 1% of users click the citation links embedded within the AI summaries.
- The Revenue Hypothesis: Translating traffic loss into financial terms, the authors calculate that a comparable ad-supported web property would suffer between $10.82 million and $37.08 million in annual losses at standard advertising rates. Because Wikipedia operates on a non-profit donation model, this figure is strictly hypothetical, but it illustrates the devastating commercial vulnerability ordinary publishers face.
- The Bot Onslaught: While human referrals are dropping, machine demand is surging. Wikimedia reports that automated crawlers account for up to 35% of total pageviews and at least 65% of high-cost core data center traffic, forcing the foundation to throttle 1.5 billion unauthorized bot requests daily.
Official Responses and Industry Perspectives
Unsurprisingly, the intersection of proprietary AI models and open-web data has generated intense friction between tech conglomerates and non-profit stewards.
Google’s Defense
Google has consistently pushed back against external studies attempting to measure the downstream economic impacts of its search features. Responding to the Washington paper via industry reports, Google argued that Wikimedia’s public data lumps all external search engines together, making it impossible to scientifically isolate Google’s specific interface modifications from other variables.
Furthermore, Google executives—including Head of Search Liz Reid—have maintained that AI Overviews ultimately benefit the broader web ecosystem. In public statements and podcast appearances, Reid asserted that total organic click volume has remained "relatively stable" year-over-year, and argued that AI summaries often guide users toward a more diverse array of niche, high-quality websites they might otherwise have overlooked.
Wikimedia’s Pragmatic Pivot
The Wikimedia Foundation has taken a measured, dual-pronged approach: publicly voicing alarm over declining human engagement while pragmatically monetizing machine access through enterprise channels.
While Wikimedia leadership acknowledges that generative AI and social media are starving the encyclopedia of traditional human traffic—noting that nearly 90% of historical visitors arrived via Google search—they have simultaneously doubled down on structured institutional access. Through Wikimedia Enterprise, the foundation offers high-volume, reliable API feeds, uptime guarantees, and support to major technology developers.
Lane Becker, who oversees the Enterprise program, clarified that the service charges for technical access and infrastructure support rather than content licensing. The unit generated $8.3 million during the 2024–2025 fiscal year, with revenues capped at 30% of the foundation’s total annual budget. By expanding its partner roster in early 2026 to include Amazon, Meta, Microsoft, Mistral, and Perplexity, Wikimedia has established a formalized toll road for the AI engines that are simultaneously eroding its organic human readership.
Implications for the Open Web
The implications of the Khosravi-Yoganarasimhan working paper extend far beyond a single digital encyclopedia. They strike at the foundational social contract of the internet: that search engines drive traffic to publishers in exchange for indexing their content.
- The Death of the Traffic Funnel: If generative AI answers queries comprehensively on the search results page, the economic incentive for creators, journalists, and researchers to publish open-access content evaporates. Why fund original reporting or deep-dive article curation if the traffic never materializes?
- The Threat to Volunteer Ecosystems: Wikipedia relies on a self-sustaining cycle—readers become enthusiastic editors, and editors become financial donors. As organic discovery bottlenecks, the pipeline of new human contributors risks drying up, threatening the long-term vitality of crowdsourced knowledge.
- The Rise of Two-Tiered Data Access: The Wikimedia Enterprise model points toward a future where open web scraping is aggressively blocked (as evidenced by Wikimedia’s 1.5 billion daily throttled bot requests), while heavily capitalized AI corporations pay for clean, structured data pipelines. Smaller publishers lacking the legal and technical leverage of Wikimedia may find themselves squeezed out entirely—their content harvested for free by AI models while their human audiences vanish behind zero-click search walls.
As the University of Washington working paper undergoes further peer review and refinement, its core warning remains clear. The transition to conversational AI search is not a zero-sum reshuffling of web traffic; it is a structural contraction that places the architects of human knowledge at the mercy of a handful of platform gatekeepers.
