Google's AI Overview feature cuts English Wikipedia pageviews by about 15%, substituting for many informational pages; culture-related articles suffer the largest traffic losses while STEM topics are less affected.

Main Finding

Google’s AI Overviews (AIO) causally reduced English Wikipedia article pageviews by about 15% after AIO was rolled out in the U.S. (March 22, 2024 → Aug 14, 2024). In the authors’ sample this implies roughly 11.5 million fewer daily visits (≈4.21 billion fewer visits per year, scaled to the sample). Effects are heterogeneous: declines are largest for Culture articles and much smaller for STEM topics, consistent with substitution being strongest when a short synthesized answer can satisfy the user’s informational intent.

Key Points

• Estimated effect: ≈15% average decline in daily pageviews to English Wikipedia articles following AIO exposure (DiD estimate).
• Absolute effect: ~−220 daily pageviews per English article (Model with article×language and date fixed effects).
• Sample: 52,262 English articles with matched versions in Hindi, Japanese, Portuguese, and Indonesian → 161,382 article–language pairs; ~46.5 million article–language–day observations (Oct 28, 2023 – Aug 14, 2024).
• Identification strategy: difference-in-differences exploiting Google AIO’s staggered U.S. rollout and Wikipedia’s multilingual editions (English treated; other languages as controls).
• Robustness: similar estimates with alternative post-period definitions, weekly aggregation, and alternative model specifications (PPML, weighted log) reported in appendices.
• Standard errors: two-way clustering by underlying article and calendar date.
• Important caveat: Wikimedia pageviews are not reported by reader country → the authors proxy exposure using language edition (English ≈ U.S. exposure during test/early rollout).
• Context: AIOs appeared on an estimated ~18% of searches according to third-party audits; Google expanded AIO internationally after the study window.

Data & Methods

• Data source: Wikimedia public pageview API (metrics/pageviews/per-article and metrics/pageviews/top) for daily pageviews, all-access types (desktop + mobile).
• Time window: Oct 28, 2023 – Aug 14, 2024. Pre-treatment: Oct 28, 2023 – Mar 21, 2024. Post-treatment: Mar 22, 2024 – Aug 14, 2024.
• Treatment definition: English Wikipedia articles treated (proxying exposure to U.S. AIO rollout); matched same underlying articles in hi/ja/pt/id as controls.
• Article selection: union of pages that appeared at least once in top-1,000 most-viewed lists (per-language) during July 1, 2023 – Mar 21, 2024; then restrict to only those with English + at least one control-language version.
• Empirical model (levels): viewsi,l,t = αi,l + δt + β · (Englishl × Postt) + εi,l,t. Authors use levels (not logs) to keep publisher-relevant additive interpretation (impressions → revenue), to avoid undue influence of low-traffic pages, and to avoid sign issues when treated/control magnitudes differ.
• Inference: article×language fixed effects absorb time-invariant cross-language differences; date fixed effects absorb common shocks; two-way clustering (article and date).
• Heterogeneity: analyzed by topical categories (Culture vs STEM etc.); substitution strongest in categories where short synthesized answers suffice.

Implications for AI Economics

• Attention and revenue reallocation: Generative-answer features embedded in SERPs can materially reallocate attention away from third-party informational publishers. For ad-supported publishers, the measured reduction in pageviews implies economically meaningful revenue losses (even though Wikipedia itself is non-advertising in the sample).
• Differential effects by content type: The stronger declines for culture/soft-information pages versus STEM indicate that AIO’s economic impact depends on query/intent type — search features that provide short factual summaries will substitute more for publishers than those requiring richer content.
• Platform design and incentives: The finding fuels the policy debate over whether platforms should share ad revenues or otherwise compensate content producers whose content is synthesized into on-platform answers. The authors note potential attribution-based approaches (e.g., Shapley valuation) as one candidate.
• Long-run content supply risk: If publishers’ revenues fall persistently, this may reduce the incentive to produce or maintain high-quality informational content, creating a negative feedback loop that could degrade the source material search engines rely on.
• Measurement and regulation: Results point to the need for finer measurement (country-level exposure to AIO, query-level AIO incidence) and possibly regulatory scrutiny of platform interface changes that materially shift ad impressions from publishers to the platform.
• Research agenda: Further work should (i) obtain user-level/country-level exposure data to refine causal estimates, (ii) quantify advertiser and publisher revenue transfers, (iii) study publisher responses (SEO, paywalls, API usage), and (iv) evaluate long-run effects on content investment and quality.

Limitations to keep in mind: the identification relies on language as a proxy for geographic exposure (Wikimedia does not provide country-disaggregated pageviews), the setting is Wikipedia (highly cited, open encyclopedia) so external validity to other publisher types may vary, and the study window predates some later AIO ad integrations and broader international rollout.