Google AI Overviews Linked to Significant Decline in Wikipedia Referral Traffic

A recent working paper from the University of Washington has reignited the debate over the impact of generative AI on the open web, estimating that Google’s AI Overviews feature reduced monthly search referrals to English Wikipedia by approximately 5% following its widespread U.S. rollout in 2024. Authored by researchers Mehrzad Khosravi and Hema Yoganarasimhan, the study suggests that the shift in how search engines present information—prioritizing summarized answers over direct links—is fundamentally altering the traffic patterns that have sustained Wikipedia for over two decades.

The paper, which was last updated on September 2 and remains pending peer review, utilizes a sophisticated difference-in-differences model to isolate the effect of AI Overviews. By comparing traffic to English Wikipedia against traffic to German and French language editions during a period when AI Overviews were not yet the default in those regions, the authors have attempted to quantify the "AI effect." According to their findings, the introduction of the feature led to an estimated loss of 100.27 million search referrals per month for the English-language edition, which equates to roughly 1.20 billion fewer visits annually.

A Chronology of the AI Search Shift

The integration of AI into search engines represents one of the most significant architectural changes in the history of the internet. For years, the primary function of search was to act as a discovery engine—a gateway directing users to destination websites. The current transition toward "answer engines" marks a departure from that model.

The trajectory of this shift can be tracked through several key milestones:

• December 2023: The period begins for the University of Washington’s longitudinal data collection, serving as the baseline before the broad deployment of AI-integrated search features.
• May 2024: Google officially switches AI Overviews to the default setting for U.S. users, marking the "post-treatment" phase for the study.
• August 2024: The researchers update their methodology to include German and French controls, refining their estimate from a previous, higher figure of 15% down to the current 5% range.
• October 2025: The Wikimedia Foundation reports an 8% year-over-year decline in human pageviews across all languages, signaling that the trend is not confined to a single market or language.
• April 2026: The Foundation’s fiscal year planning documents formally acknowledge that declining referrals from search engines and the rise of AI-driven consumption are anticipated to be long-term structural challenges rather than temporary fluctuations.

Dissecting the Methodology and Google’s Rebuttal

The study’s findings hinge on the researchers’ ability to control for variables. By utilizing Wikipedia’s monthly clickstream data, Khosravi and Yoganarasimhan mapped hundreds of thousands of matched article pairs across different language versions. The model assumes that English, German, and French users behave similarly when searching for the same topics, allowing the researchers to attribute the divergent traffic trends in the U.S. to the presence of AI Overviews.

Google, however, has contested the study’s conclusions, arguing that the combined referral metrics used in the paper are insufficient to isolate the impact of AI Overviews from other variables in the search ecosystem. Google maintains that AI Overviews are designed to guide users toward a diverse range of high-quality sources, effectively expanding the utility of the search results page. In official communications, Google representatives have noted that total organic click volume remains relatively stable, though the company has not released granular data to support this claim or to verify how those clicks are distributed between Wikipedia and other publishers.

The Broader Economic and Operational Impact

While the 5% reduction in referrals may appear incremental, the implications for the Wikimedia Foundation and the broader digital ecosystem are profound. Wikipedia is a unique entity; it does not display advertising, and its financial model relies on donations and, increasingly, paid enterprise access. The foundation has noted that nearly 90% of its traffic has historically originated from search engines, primarily Google.

The financial estimates provided by the researchers—suggesting an ad-supported site of similar scale might lose between $10 million and $37 million annually—serve as a hypothetical indicator of the value being extracted from the open web. Because Wikipedia is non-profit, these figures do not represent direct revenue losses for the foundation, but they do highlight a "value gap." As users receive the information they need directly on the search engine results page (SERP), the incentive to click through to the source diminishes.

This behavior is supported by independent data from the Pew Research Center, which found that users are significantly less likely to click on traditional links when an AI summary is present. In tests conducted in 2025, only 1% of users clicked on the sources cited within an AI summary, compared to much higher interaction rates with organic blue-link results.

The Rise of the Machine-to-Machine Web

Parallel to the decline in human traffic is an unprecedented surge in automated traffic. Wikimedia engineers have reported a 50% increase in bandwidth consumption for image and media files since early 2024, the vast majority of which is driven by scrapers and crawlers training AI models.

This creates a paradox: while human readers are visiting Wikipedia less frequently via search, the platform’s content is being ingested by bots at an exponential rate. The Foundation has responded by implementing stricter rate limits and promoting the Wikimedia Enterprise API, a service that provides high-volume, reliable access to Wikipedia’s data. While this service generated $8.3 million in the 2024-25 fiscal year, it is designed to facilitate sustainable data use rather than to compensate for the loss of human referral traffic.

Implications for the Future of Information

The ongoing decline in referral traffic raises existential questions for the open-web model. If the primary reward for creating free, high-quality information is visibility, and that visibility is being cannibalized by the platforms that host the information, the incentive structure for content creation may collapse.

The Wikimedia Foundation’s leadership has been vocal about the need for a new social contract between AI developers and content creators. In their recent public statements, they have urged developers to provide clear attribution, arguing that proper credit is essential to maintain the "virtuous cycle" of the internet—where readers are converted into editors and donors. Without this cycle, the community-driven model that built Wikipedia could face significant strain.

As the University of Washington paper continues to be refined, it serves as a critical diagnostic tool for a digital landscape in flux. Whether the current 5% decline stabilizes or accelerates remains to be seen, but the data suggests that the era of the search engine as a simple traffic conduit is ending. In its place, a new, more opaque ecosystem is emerging, one where the distinction between the creator of information and the aggregator of information is increasingly blurred. For the time being, the academic community and the Wikimedia Foundation remain in a state of close observation, monitoring whether these shifts in traffic patterns will eventually translate into a decline in the quality and breadth of the world’s most extensive public knowledge base.