Who Owns Digital News, and Who Gets Cited (Carragher & Carley, arXiv:2610.40164)

media-concentrationgenerative-aiai-searchnews-ownershipstate-mediaantitrustoutside

Capture from a fresh, vault-external surface: arXiv:2610.40164, "Who Owns Digital News, and Who Gets Cited: Measuring Media Concentration Across 77 Countries and Seven AI Systems," by Peter Carragher and Kathleen M. Carley (Carnegie Mellon University), submitted 29 Oct 2026 to cs.CY. Both the abstract page and the PDF were fetched directly (TLS verified on the PDF fetch); no safety signals fired on either.

The paper builds a global news-ownership dataset (9,761 domains, 77 countries with reliable coverage) from Wikipedia/Wikidata, cross-validates it against independent US (Medill) and UK (PINF) directories and against Eli Noam's prior cross-national study, and then asks whether citation patterns from seven generative-AI systems (ChatGPT, Copilot, Gemini, Google AI Overviews, Grok, plus Perplexity and GPT-4.0 via a second dataset) track or amplify existing media concentration.

Claim: Owner-level ownership concentration is systematically higher than outlet-level concentration (median 1.30x)

Grouping digital news outlets by their ultimate owner, rather than counting each outlet as an independent competitor, raises measured market concentration. Across the dataset's countries, owner-level concentration indices are a median factor of 1.30x higher than the equivalent outlet-level indices — i.e., treating affiliated outlets as independently competing understates how concentrated a national news market actually is.

• source_url: https://arxiv.org/abs/2610.40164 (PDF: https://arxiv.org/pdf/2610.40164)
• source_sha: 33521fd5dc14fbd653efcf1482cf3bbd35bd9ea16d212128846c8b0cea92e1f6
• source_author: Peter Carragher; Kathleen M. Carley
• source_date: 2026-10-29
• source_title: "Who Owns Digital News, and Who Gets Cited: Measuring Media Concentration Across 77 Countries and Seven AI Systems"
• source_venue: arXiv:2610.40164v1 [cs.CY]
• source_tier: 1
• source_quote: "outlet-level measures understate concentration, with owner-level indices a median 1.30× higher"
• claim type: quantitative — Tier 1 (own arXiv preprint, their own measurement) — clears the floor.

Claim: AI citation distributions are substantially more concentrated than traffic distributions, both by outlet and by owner

Comparing citation counts from each AI system against SimilarWeb traffic estimates over the same outlets (Ahrefs AI-response-count data, 4,584 news domains across 199 countries, tested against a Wilcoxon signed-rank test at p < 10⁻⁶ for all five systems), the paper finds that every AI system's citations are more concentrated among fewer outlets than the audience traffic to those same outlets. The effect persists after outlets are grouped by owner, though it weakens somewhat at that level.

• source_url: https://arxiv.org/abs/2610.40164 (PDF: https://arxiv.org/pdf/2610.40164)
• source_sha: 33521fd5dc14fbd653efcf1482cf3bbd35bd9ea16d212128846c8b0cea92e1f6
• source_author: Peter Carragher; Kathleen M. Carley
• source_date: 2026-10-29
• source_title: "Who Owns Digital News, and Who Gets Cited: Measuring Media Concentration Across 77 Countries and Seven AI Systems"
• source_venue: arXiv:2610.40164v1 [cs.CY]
• source_tier: 1
• source_quote: "we find that AI citation distributions are 1.23×–2.08× more concentrated than traffic distributions at the outlet-level, and 1.11×–1.79× more concentrated at the owner-level."
• claim type: quantitative — Tier 1 — clears the floor. The strongest single-system ratio (Gemini, Ahrefs data) reaches 2.08x at outlet level per Table 2; Grok is the weakest at 1.23x.

Claim: A hypothetical full shift of US organic search to Google AI Overviews would raise US media concentration roughly 3.4x more than the Paramount–Warner Bros. Discovery merger did

Using the actual Paramount Skydance acquisition of Warner Bros. Discovery (closed 6 Oct 2026, ~$110bn deal) as a real-world baseline, the paper calculates that combining those two companies' outlets under one owner raised their traffic-based estimate of US media concentration by +9.9%. By contrast, in a scenario where all US organic search traffic is redistributed in proportion to Google AI Overviews' current citation patterns (holding total traffic constant), their estimate of US media concentration rises by +33.3% — about 3.4 times the merger's effect. The paper frames this as evidence that AI-mediated discovery, which no media regulator currently has authority over, may now be a larger lever on market concentration than the mergers-and-acquisitions review process that does receive regulatory scrutiny.

• source_url: https://arxiv.org/abs/2610.40164 (PDF: https://arxiv.org/pdf/2610.40164)
• source_sha: 33521fd5dc14fbd653efcf1482cf3bbd35bd9ea16d212128846c8b0cea92e1f6
• source_author: Peter Carragher; Kathleen M. Carley
• source_date: 2026-10-29
• source_title: "Who Owns Digital News, and Who Gets Cited: Measuring Media Concentration Across 77 Countries and Seven AI Systems"
• source_venue: arXiv:2610.40164v1 [cs.CY]
• source_tier: 1
• source_quote: "In this scenario, with attention redistributed in proportion to Google AI Overview's citations, US media concentration rises by +33.3%, roughly 3.4 times the Paramount merger's effect."
• claim type: quantitative + counterfactual/modeled — Tier 1 for the paper's own stated figure, but this is explicitly a scenario estimate (not an observed outcome): the paper itself says elsewhere that confirming the underlying mechanism (traffic actually shifting to follow AI citations) "requires longitudinal data" it does not have. Treat the 3.4x figure as the authors' own modeled projection, not a measured historical effect.

Claim: When prompted in a country's own native language, AI systems cite that country's state-controlled media outlets at a higher rate relative to traffic than matched commercial outlets receive

Using the AI Search Arena dataset (non-English prompts, ~40% of its prompt set), and restricting to users located in a given country prompting in that country's own language about that country's own outlets, state-controlled media outlets are cited 1.65x [95% CI 1.46–1.87] more often than commercial outlets matched to the same traffic share. In absolute terms, state-controlled outlets are cited at 1.01x their traffic share versus 0.61x for the matched commercial comparison group, across 2,961 citations spanning 16 countries. The paper attributes this effect primarily to Russian- and Mandarin-language prompts, which draw markedly more citations to Russian and Chinese state media respectively than English-language prompts about the same outlets (5.3x and 2.5x more, respectively, per the paper's Table 8). The same preference did not appear in the paper's separate English-language Ahrefs panel, which the authors attribute to that panel's English-only prompt construction rather than evidence against the effect.

• source_url: https://arxiv.org/abs/2610.40164 (PDF: https://arxiv.org/pdf/2610.40164)
• source_sha: 33521fd5dc14fbd653efcf1482cf3bbd35bd9ea16d212128846c8b0cea92e1f6
• source_author: Peter Carragher; Kathleen M. Carley
• source_date: 2026-10-29
• source_title: "Who Owns Digital News, and Who Gets Cited: Measuring Media Concentration Across 77 Countries and Seven AI Systems"
• source_venue: arXiv:2610.40164v1 [cs.CY]
• source_tier: 1
• source_quote: "Restricting to users in a country, prompting in its national language, and citing that country's own outlets, state-controlled media are cited 1.65× [1.46, 1.87] more often than commercial outlets matched to the same traffic share"
• claim type: quantitative — Tier 1 — clears the floor. Note the authors' own caveat that the Arena dataset is "underpowered" for some sub-tests and that statistical claims on it are "difficult to defend" for the owner-level cut specifically (appendix, Limitations section) — the 1.65x headline figure itself rests on the better-powered 16-country/2,961-citation cut, not the thinnest slice.
• Related framing: this finding builds explicitly on prior work the authors cite as showing LLMs "exhibit pro-government valences when prompted in the languages of countries with lower media freedom" — see Entity candidates below for the foundational prior study this result is checked against.

Further leads

• The paper's own self-citations [11][12] (Carragher, Williams & Carley 2024/2025 on misinformation-source detection and webgraph-based search interventions) connect this work to a broader line on search/webgraph manipulation — possible bridge to vault material on information ecosystem manipulation, not pursued here.
• Journalists report more owner/manager pressure in countries with higher ownership concentration (Spearman ρ = +0.318 for HHI, +0.359 for largest-owner share, across 55 countries, Worlds of Journalism Study data) — supporting/validation finding, not one of the paper's AI-specific results, flagged here rather than worked up as a core claim.
• State-controlled media's share of national news traffic correlates negatively with RSF press-freedom scores (ρ = −0.467, n = 40 for media trust; separate RSF correlation also reported) — a validation result for the ownership dataset itself.
• The dataset and code are released at https://doi.org/10.5281/zenodo.22782826 per the paper's own citation — not independently verified or fetched this session.
• Press Gazette (Dominic Ponsford, 2026) is cited in the paper's own references [37] as trade-press coverage of "which journalists and news outlets are most cited by AI answer engines" — a Tier 3 secondary source on the same general topic, not fetched independently this session.
• The paper models two underlying mechanisms by which AI systems affect news economics: citation allocation (measured directly here) and traffic generation/click-through, citing a separate finding that "roughly one AI citation link in twenty is clicked, well below click-through in conventional search" [Agarwal & Sen 2026] — a quantitative claim from a different paper, not verified independently this session, flagged as [unsourced -- needs verification] since it was only read secondhand via this paper's citation.

Entity candidates

• Eli M. Noam — person (media economist, Columbia University) — concept/term — author of Who Owns the World's Media? (Oxford University Press, 2016) and the earlier Noam concentration index (2009); this is the foundational prior work the present paper explicitly measures itself against (reproduces his four-firm newspaper concentration ranking at ρ=+0.465, and his HHI ranking at ρ=+0.407, across the 29-30 countries his 2008-2013 study covered) and whose national-press/broadcast-census methodology it says it is "translating... to attention-weighted measurement of digital news." Flagging first per the instructions' note that the foundational ancestor tends to be under-caught.
• Peter Carragher — person — concept/term — first author, CMU PhD researcher; also behind cited prior work on misinformation-source detection via webgraphs (refs [11],[12] in this paper).
• Kathleen M. Carley — person — concept/term — senior/corresponding author, CMU professor (Carnegie Mellon's Center for Computational Analysis of Social and Organizational Systems is her long-running affiliation, not independently confirmed this session).
• Waight, Yang, Yuan, Messing, Roberts, Stewart & Tucker (2026), "State media control influences large language models," Nature — the specific prior study [51] this paper's native-language state-media finding (H4) is explicitly built on and compared against ("a finding we build on in this work"); this is the direct methodological ancestor for the state-media claim, distinct from Noam's ancestry for the concentration-index claims.
• Herfindahl–Hirschman Index (HHI) — concept — the standard antitrust concentration measure (adopted by the FCC in 2005 per the paper) that underlies every concentration figure in this capture; likely warrants its own definitional note.
• Generative Engine Optimization (GEO) — concept — the paper's term (borrowed from Aggarwal et al. 2024) for SEO-style strategies aimed at increasing AI-citation rates; connects to the paper's broader argument about publishers reacting to AI-mediated discovery.
• State Media Monitor (SMM) — concept/term — the Media and Journalism Research Center's hand-compiled census of state-administered media that the paper relies on for its state-media ownership labels (298 domains, 80 countries) rather than its own derived Wikidata labels, which it reports have higher precision but much lower recall.
• Paramount Skydance / Warner Bros. Discovery merger — historical/event — the real-world ~$110bn deal (closed 6 Oct 2026, contested by twelve state attorneys general and the Writers Guild of America) used as the paper's empirical anchor for comparing merger-driven vs. AI-driven concentration shifts.