The AI Overview Effect: How Google’s Summaries Cut Into Wikipedia’s Search Referrals

As generative artificial intelligence transforms how humanity discovers information on the web, a foundational question has loomed over the digital ecosystem: What happens to the underlying sources when search engines answer queries directly on the results page?

A new, non-peer-reviewed working paper from the University of Washington provides one of the most rigorous attempts yet to answer that question. Authored by researchers Mehrzad Khosravi and Hema Yoganarasimhan, the study estimates that the rollout of Google’s AI Overviews reduced monthly search referrals to English-language Wikipedia by roughly 5%.

While a 5% decline may sound modest at first glance, scale changes everything. For a global public resource that relies heavily on search engines to connect human readers with volunteer-generated knowledge, that percentage translates to staggering losses: roughly 100.27 million fewer human search referrals every month, or approximately 1.2 billion annually.

Yet, the study is only one piece of a much larger, complex puzzle involving shifting user behavior, automated bot scraping, and a fundamental realignment of how major tech platforms interact with the open internet.

Main Facts: The Core Findings and Methodology

To measure the impact of Google’s AI Overviews—which dynamically synthesize web content at the top of search result pages—the researchers needed a reliable control group. Because AI Overviews became the default setting for U.S. searches in May 2024 but were not simultaneously deployed in the same manner across all global regions and language editions, Wikipedia offered a rare, natural laboratory.

The Research Design

The paper, initially published in February 2024 and significantly updated through late summer, analyzes monthly clickstream data released by the Wikimedia Foundation at the article level.

• The Scope: The dataset covers December 2023 through December 2024, treating May 2024 as the crucial post-treatment threshold.
• The Comparison: The authors matched 499,927 English-German article pairs and 530,873 English-French article pairs. Because default AI Overviews had not rolled out in Germany or French-speaking regions during the sample period in the same capacity, German and French editions served as the control groups.
• The Model: Using a Poisson pseudo-maximum likelihood difference-in-differences model, the authors isolated the relative dip in English referrals compared to German and French referrals for the same subjects.

The results point to a statistically significant decline: default AI Overview availability reduced monthly external-search referrals to English Wikipedia by 5.45% relative to German and 4.82% relative to French. A secondary check comparing English to Japanese data indicated a directional 16.53% decline, though under a shorter timeframe and different parameters.

Chronology: How the Data and Estimates Evolved

The path to these figures has been iterative, reflecting the fast-moving and complex nature of measuring generative AI’s footprint on web traffic.

• February 2024: The authors released their initial working paper, which leaned on daily pageview metrics and originally estimated a much steeper decline of roughly 15%.
• May 2024: Google officially switches AI Overviews on by default for U.S. search users, setting the stage for the structural traffic shift.
• August 26, 2025 (Version 5): The researchers revise their methodology, shifting away from raw daily pageviews to a more targeted focus on monthly external search referrals, while introducing German and French language editions as rigorous control groups.
• September 2, 2025 (Version 6): The working paper is updated to its current iteration, cementing the 5% estimate while explicitly noting that the research remains non-peer-reviewed.
• October 2025 – Present: Broader institutional warnings from the Wikimedia Foundation begin to compound the narrative, highlighting that traffic pressures are no longer viewed as short-term anomalies.

Supporting Data: Pew Research, Bot Traffic, and Revenue Hypotheses

The University of Washington study is contextualized by a growing body of independent research detailing user engagement with AI search features.

User Click Behavior

A study by the Pew Research Center analyzing nearly 69,000 U.S. Google searches revealed stark disparities in user engagement. When an AI summary was present on a search results page, users clicked on traditional organic links only 8% of the time, compared to 15% when no summary appeared. Furthermore, only 1% of visits involving an AI summary resulted in a click on a source embedded directly within the summary box.

The Bot Economy vs. Human Readers

Compounding the referral decline is a massive surge in automated machine demand. In April 2025, Wikimedia engineers reported that bandwidth for downloading images and media files had jumped 50% since January 2024, driven primarily by automated bots scraping Wikimedia Commons for AI training data rather than by regular human readers. By 2026, the Foundation noted it was actively blocking or throttling roughly 1.5 billion automated requests per day from non-compliant crawlers.

The Hypothetical Revenue Impact

The study also calculates a purely financial thought experiment: what would this lost traffic mean for a commercial, ad-supported website? At standard industry ad rates, the authors estimate that a comparable enterprise might lose between $10.82 million and $37.08 million annually. However, the authors stress that this is entirely hypothetical; Wikipedia does not run advertisements, and no funds are actually changing hands between Google and Wikimedia based on this metric.

Official Responses: Google vs. Wikimedia Foundation

Both tech giants and open-source advocates have weighed in on how these traffic trends should be interpreted.

Google’s Position

Google has consistently disputed whether broad referral metrics can effectively isolate the specific impact of AI Overviews. Representatives have argued that external search data lumps various engines together and fails to account for the nuanced ways users navigate the web. Additionally, Google executives—such as Liz Reid, head of Google Search—have maintained that overall organic click volume remains "relatively stable" compared to previous years, arguing that AI Overviews ultimately help users discover a broader and more diverse range of web resources.

Wikimedia’s Perspective

The Wikimedia Foundation views the situation through a lens of systemic transformation. In October 2025, Wikimedia Product Director Marshall Miller noted that overall human pageviews across all languages had dipped about 8% compared to prior years, attributing the trend heavily to generative AI and social media platforms.

In its draft fiscal plan for 2026–2027, the Foundation explicitly stated that declining pageviews and reduced search referrals from major engines like Google are structural realities that must be planned for, rather than temporary dips.

Implications: The Future of Open Knowledge and Enterprise Access

The implications of the Washington study stretch far beyond academic curiosity, striking at the economic and operational heart of how open-source platforms survive.

The Contributor and Donor Pipeline

Wikipedia’s model is entirely cyclical: users arrive via search engines, read articles, become inspired, and eventually evolve into financial donors or volunteer editors. If top-of-funnel search referrals drop by 5%—and overall human pageviews slide alongside them—the long-term viability of the volunteer ecosystem faces silent erosion. Fewer readers naturally mean fewer hands to write, edit, and fact-check the encyclopedia.

The Rise of Wikimedia Enterprise

To navigate this shifting landscape, the Wikimedia Foundation has increasingly leaned on Wikimedia Enterprise, a paid commercial service providing high-volume, structured API access, uptime guarantees, and support feeds to major tech entities.

While the service charges for infrastructure and access rather than content licensing, its partner roster has expanded dramatically. Alongside Google, enterprise partners now include Amazon, Meta, Microsoft, Mistral AI, and Perplexity. In the 2024–2025 fiscal year, the unit reported $8.3 million in revenue, capped at 30% of the Foundation’s overall revenue.

However, Foundation officials emphasize that Wikimedia Enterprise is designed to support large-scale machine reusers and data infrastructure costs—it is not intended to serve as financial compensation for organic referral losses suffered by human readers.

What Wikipedia Can—and Cannot—Prove

Ultimately, researchers and industry analysts alike acknowledge the inherent limitations of the study. Because Wikipedia is uniquely structured with massive multilingual parallel datasets, it makes for an ideal academic testing ground. Yet, whether news outlets, independent blogs, or e-commerce retailers experience the exact same 5% penalty remains an open question.

As generative AI continues to rewrite the user journey from "search and click" to "search and read," the University of Washington paper serves as a vital statistical baseline. It proves that when machines answer the questions first, fewer human feet walk through the virtual doors of the web’s foundational knowledge base.