AI systems' news citations are more concentrated than the outlets' actual traffic

media-concentrationai-searchgenerative-aiantitrust

Carragher and Carley (arXiv:2610.40164) compared how often seven generative-AI systems (ChatGPT, Copilot, Gemini, Google AI Overviews, Grok, plus Perplexity and GPT-4.0 via a second dataset) cite particular news domains against those same domains' SimilarWeb traffic estimates — 4,584 news domains across 199 countries, using Ahrefs AI-response-count data, tested with a Wilcoxon signed-rank test at p < 10⁻⁶ for all five systems in the primary panel.

Every system tested cites a narrower slice of outlets, relative to audience size, than readers actually visit: AI citation distributions run 1.23x–2.08x more concentrated than traffic distributions at the outlet level (Gemini is the strongest single-system ratio at 2.08x; Grok the weakest at 1.23x, per the paper's Table 2), and 1.11x–1.79x more concentrated at the owner level — the effect persists after affiliated outlets are grouped under one parent company (see claim-owner-level-news-concentration-130x-outlet-level for that grouping method), though it weakens somewhat once it does. In plain terms: whichever outlets an AI system favors, it favors them harder than readers' own browsing habits would predict.

The paper treats this as a measured association between citation share and traffic share within its sample window, not as evidence that traffic is shifting to follow citations over time — the authors say elsewhere that confirming such a causal mechanism "requires longitudinal data" they do not have.

Related

• entity-herfindahl-hirschman-index — the family of concentration measures both this and the owner-level note apply.
• question-corroborate-carragher-carley-2026-ai-overview-concentration-scenario — the paper's more speculative modeled-scenario finding, held at the question stage rather than promoted as a sibling claim-note, per the vault's single-source concentration cap.