AI Overviews reduced monthly search referrals to English Wikipedia by roughly 5% after Google made the feature the U.S. default in May 2024, according to a University of Washington working paper last updated September 2. The estimate translates to approximately 100 million fewer referrals per month, or 1.2 billion annually, and is already contested by Google over how the measurement was constructed.
How Did Researchers Calculate the 5% Figure?
The paper, by Mehrzad Khosravi and Hema Yoganarasimhan, applied a difference-in-differences model (a statistical method that compares a treated group against a control group over the same time window) to Wikimedia’s publicly released monthly clickstream data. Because Google rolled out the feature as a U.S. default in May 2024 but had not yet done so in Germany or France during the study period, the authors used German and French Wikipedia editions as comparison groups representing regions without the feature.
- Dataset: December 2023 to December 2024
- Article pairs: 499,927 English-German and 530,873 English-French matched article pairs (per the working paper)
- Key result: English Wikipedia referrals fell 5.45% relative to German and 4.82% relative to French after May 2024
- Scale: Roughly 100.27 million fewer search referrals per month; approximately 1.20 billion annually if the effect holds throughout the year
A supplementary English-Japanese comparison in the same paper shows a 16.53% decline, though the authors flag this as directional support only, given a shorter timeframe and a different control group.
Why Did the AI Overviews Estimate Shrink from 15% to 5%?
Earlier versions of the paper reported a roughly 15% decline, based on daily pageview data without language-edition controls. The revision, introduced in version 5 on August 26, switched to monthly search referrals and added German and French comparisons. That methodological shift produced the more conservative 5% figure. The paper has not yet been peer-reviewed.
Google disputes the analysis, telling UOL Tilt that Wikimedia’s data aggregates all external search engines together, making it impossible to isolate Google’s contribution specifically. Yoganarasimhan responded that other engines account for only a small fraction of Wikipedia’s external traffic, and that the timing of the U.S. rollout provides a credible natural experiment regardless.
What Does Google’s Own Traffic Data Show?
Google published a data summary on its blog stating that total organic click volume from Search to websites has been “relatively stable year-over-year” since the feature launched. The company reports that average click quality has increased, and that it is “sending slightly more quality clicks to websites than a year ago,” where quality clicks are defined as visits where users do not immediately return to the search results page.
Google attributes part of the shift to user behavior: people are asking longer, more complex questions, and the feature displays a wider set of links than classic results. Its blog post notes that traffic is moving toward sites with forums, videos, podcasts, and first-hand perspectives, as well as in-depth reviews and original analysis. Google’s developer documentation adds that both AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related sub-searches to build a response and surface a more diverse set of supporting links than a single classic search would show.
What This Means for AI-Search Visibility
Here is the clearest way to frame what the Wikipedia data shows: when AI gives someone a direct answer, fewer people click through to the source, even when that source is one of the most trusted sites on the internet. That is the underlying pressure every publisher now faces.
A few observations, as editorial analysis from Hingewise:
- The 5% is an average across all article types. Simple definitional queries, where Wikipedia once captured near-automatic clicks, are likely where the impact is steepest. Articles covering complex or evolving topics probably fared better, consistent with Google’s data showing click quality holding up for deeper content.
- Wikipedia’s position is not typical. It is one of the most-cited sources inside AI responses, which means it retains visibility as a cited source even as direct referrals fall. Most brands cannot assume the same. Being cited inside an AI response is not automatic; it reflects the credibility and structure of the underlying content.
- The revision from 15% to 5% is itself instructive. Early traffic impact estimates for new AI features are volatile. Studies published in the months after a feature launch often use incomplete data or insufficient controls. Treating any single estimate as settled, in either direction, is premature.
- Google’s “stable overall traffic” claim and the 5% Wikipedia finding are not necessarily contradictory. If the feature redistributes clicks from quick-answer content to deeper editorial content, aggregate totals can stay flat while individual content categories diverge sharply.
Hingewise’s read: the meaningful question is not whether total web traffic is up or down in aggregate. It is whether a given page serves a simple factual query (high exposure to AI substitution) or a complex, original, or experience-based query (more likely to benefit as AI directs users to richer sources). Publishers and brands that can map their top-traffic pages onto those two buckets will have a clearer picture of their actual exposure than any headline percentage provides.
What to Check Before Drawing Conclusions from Your Own Traffic Data
- Review Search Console’s Performance report under the “Web” search type, where AI Overview clicks are included per Google’s developer documentation
- Segment pages by query intent: informational or definitional versus comparative or experience-based
- Compare click-through rates for queries where the feature regularly appears versus those where it does not
- Track engagement metrics (time on page, scroll depth) alongside raw click volume, since click quality may be rising even if volume is flat or declining
- Avoid comparing traffic from May 2024 onward against pre-May 2024 baselines without accounting for the U.S. rollout as a structural break in the data
- Note which content formats are gaining traction: forums, first-hand accounts, in-depth reviews, and original analysis are specifically cited in Google’s blog data as benefiting from the current shift
The University of Washington paper remains a working paper and may be updated further. The effect of AI Mode on referral patterns at this scale is largely unmeasured. That makes Wikipedia’s monthly clickstream data an unusually clean signal in a space where most traffic data is far noisier. How these numbers evolve as AI Mode continues to expand will be worth tracking closely.
Sources: Search Engine Journal; University of Washington working paper by Khosravi and Yoganarasimhan (last updated September 2); Google Blog; Google Search Central developer documentation.
Sources
This article was drafted with AI assistance from the sources cited above and checked against them before publishing. The "What this means for AI-search visibility" section is Hingewise's editorial analysis, built on the frameworks and positions we use in our GEO work. How we work: Editorial policy.