Who AI Cites: OpenAI's Wire Services, Perplexity's BBC, Gemini's Forbes — 366,087 Citations Mapped

The Study, Defined

The paper is titled "News Source Citing Patterns in AI Search Systems." It was authored by Kai-Cheng Yang, Assistant Professor in the School of Computing at Binghamton University. Yang holds a Ph.D. in Informatics from Indiana University and previously spent two years as a postdoctoral researcher at Northeastern University's Network Science Institute. He leads the Yang Lab at Binghamton, focused on online information ecosystems, generative AI's effects on information flow, algorithmic bias, and bot detection. The paper was submitted to arXiv on July 7, 2025 (arXiv:2507.05301) and presented at a Binghamton CoCo seminar on September 17, 2025.

The author's prior work is directly relevant to the methodology. Yang has previously co-authored work with Filippo Menczer on AI-generated political bias in source credibility ratings and on the DomainDemo dataset — which scores 129,000+ news outlets by political leaning based on a panel of 1.5 million Twitter users matched to their voter registration records. The DomainDemo scoring is what enables this paper's political-bias analysis.

The study has a distinctive methodological feature among the six papers that now define the AI citation literature. It is the only one of the six built on real-world user queries rather than simulated or template-generated prompts. The data comes from the AI Search Arena, a head-to-head evaluation platform where users submit queries to multiple AI search systems and compare outputs side by side. That gives the dataset something the controlled-prompt studies cannot — actual evidence of what users ask, in their own phrasing, and which engine they preferred.

The Scope

The dataset spans 24,069 user conversations and 65,294 AI responses from three providers — OpenAI, Perplexity, and Google. Across those responses there are 366,087 individual citations linking to 83,533 unique domains. Of the total citations, approximately 9 percent reference news sources. The remainder are reference works (Wikipedia, encyclopedias), brand and company domains, academic sources, government sites, and other non-news content.

The models included in the analysis are the production answer engines from each provider. The OpenAI models are gpt-4o-search-preview and gpt-4o-mini-search-preview. The Perplexity models include the Sonar family. The Google models include Gemini in its search-grounded configuration. One non-search model was excluded from the analysis to keep the comparison fair.

For each citation, Yang extracted the registrable domain (nytimes.com for The New York Times, bbc.co.uk for the BBC, and so on), classified it as news or non-news using a published media reference list, and scored news domains for political leaning using DomainDemo. Source credibility scores were drawn from a separate validated dataset of low-credibility news outlets.

The Findings, By Engine

OpenAI — primary wire services and global English-language news

OpenAI's two search models in the dataset showed the highest news-citation rate of the three providers. gpt-4o-search-preview cited news in 20.3 percent of its responses; gpt-4o-mini-search-preview cited news in 19.3 percent. The outlets OpenAI's models most consistently cite are primary wire services and major global news organizations with documented editorial authority — Reuters, The Associated Press, The New York Times, the BBC, The Washington Post, and a small set of comparable outlets.

The pattern reflects something about OpenAI's content licensing strategy. OpenAI has signed publishing agreements with multiple major news publishers in 2024 and 2025 — Axel Springer, the Financial Times, News Corp, Le Monde, and others. The cited-domain frequencies in Yang's dataset are roughly consistent with prioritizing publishers in OpenAI's licensing portfolio, though the paper itself does not assert causation.

Perplexity — BBC dominance, broad global coverage

Perplexity's citation pattern is the most distinctive of the three providers. The BBC is cited at substantially higher frequency in Perplexity responses than in either OpenAI or Google. The Guardian and Reuters also rank high. Perplexity tends to cite a wider distribution of outlets across geographies than OpenAI does — UK and European publications appear more frequently. The product itself markets transparency in source citation, and the resulting citation distribution favors outlets that Perplexity has surfaced as trusted across its product history.

Gemini — Forbes across categories, lower overall news rate

Gemini cites news at lower overall frequency than OpenAI or Perplexity but exhibits one consistent cross-category pattern that no other engine shows. Forbes surfaces in Gemini citations across both consumer and business queries — the only traditional business publication that appears consistently in 11 major B2B and B2C sectors that researchers and operators have audited downstream of Yang's work. The New York Times, The Wall Street Journal, and BBC also appear frequently in Gemini citations, but Forbes is the standout cross-category citation anchor.

The Cross-Engine Findings

Citation concentration is consistent across providers

All three providers exhibit pronounced citation concentration. A small number of news outlets account for a disproportionate share of total citations. This is not unique to AI search — traditional search audits have found the same pattern (Trielli and Diakopoulos, CHI 2019; Fischer, Jaidka, and Lelkes, 2020). But it is consistent across AI systems despite the providers' different model architectures, training corpora, and retrieval pipelines. The structural force concentrating attention on a small set of authoritative outlets persists across the transition from traditional search to AI search.

Political bias is present and consistent

Yang's DomainDemo-based analysis finds a consistent left-leaning political bias in cited news sources across OpenAI, Perplexity, and Google. The lean is pronounced. The pattern holds when querying across topics that should produce ideologically diverse source distributions. This is not the same finding as bias in the model's output language — it is bias in which outlets the engine selects to ground its answers.

The methodological note is important. DomainDemo scoring is based on the political affiliation of users who share each outlet's content on Twitter — it measures audience composition, not editorial alignment. An outlet whose readers skew Democratic is scored as left-leaning regardless of its editorial stance. This is a reasonable proxy for political bias in source selection because AI engines have no direct knowledge of editorial alignment; they learn from associations in their training data, which include audience-driven sharing patterns.

Low-credibility sources are rarely surfaced

Across all three providers, sources classified as low-credibility by external media reliability ratings appeared in only a small share of citations. The engines have filtered out the worst of the open web. This is a meaningful finding given the persistent concern that AI engines might amplify disinformation. The Yang data suggests the major providers have, at least at the citation layer, succeeded at not surfacing the lowest-quality sources.

The implication for brands and PR practitioners is direct. If your category includes coverage in low-credibility outlets — content marketing farms, low-authority aggregators, marginal sites — that coverage is statistically unlikely to be retrieved by AI engines when answering category queries. Investment in citation in those outlets is largely wasted as an AI visibility play.

User satisfaction is uncorrelated with source quality or political balance

This is the most policy-relevant finding in the paper. The AI Search Arena dataset includes user preference data — which engine the user preferred when comparing two responses to the same query. Yang's analysis finds that neither the political leaning of cited sources nor the credibility of cited sources significantly influences user satisfaction. Users prefer answers based on factors that have little to do with the underlying source mix.

This decouples the citation-share question from the user-experience question. A brand can be cited by sources that users would rate as more credible if asked — and the user would still prefer a competing answer based on tone, completeness, structure, or other factors unrelated to source provenance.

What This Means for Media Relations Strategy

Six operational implications follow from Yang's findings for communications practitioners.

One — the target outlet list is engine-specific. A brand optimizing for citation share inside ChatGPT should prioritize Reuters, AP, NYT, Washington Post, Bloomberg, and the licensed-publisher set. A brand optimizing for Perplexity should prioritize BBC, Guardian, and Reuters. A brand optimizing for Gemini should prioritize Forbes, NYT, WSJ, and BBC. The engine-specific lists are different enough that a single "tier-one media" list is inadequate.

Two — citation concentration is leverage. Because each engine concentrates citation on a small number of outlets, securing coverage in those concentrated outlets produces disproportionate citation share. The long tail of mid-tier outlets is much less efficient as an AI visibility investment than the concentrated head.

Three — the bias finding is news-worthy, not actionable. The left-leaning political bias is a documented characteristic of the AI citation infrastructure. Brands cannot meaningfully shift this through their own media work — it is a property of the underlying training and retrieval systems. The finding matters as context for how the engines behave, particularly on politically sensitive coverage. It does not change which outlets a brand should target.

Four — investment in low-credibility outlets is dead weight. AI engines filter low-credibility sources. Content marketing placement, low-authority directories, and marginal sites do not contribute meaningfully to AI citation share. Budget should consolidate toward outlets that AI engines have learned to weight as authoritative.

Five — source quality and user satisfaction are decoupled. Being cited by better sources does not, on Yang's evidence, make users more likely to find the answer satisfactory. The two metrics — citation quality and user preference — track separately. A complete AI visibility program tracks both.

Six — Forbes is the cross-category AI citation anchor for business coverage. Of all traditional business publications, Forbes is the one that surfaces most consistently across B2B and B2C category queries inside Gemini. For brands operating across multiple verticals, Forbes earned coverage has the most consistent AI visibility yield of any single business publication identified in Yang's data.

Why Yang's Methodology Matters

The 366,087-citation scale of Yang's dataset is large, but the more important methodological contribution is that the queries are real. Every other study in the current AI citation literature uses simulated queries — templates that researchers generate to test specific hypotheses. The Toronto comparative audit (Chen, Wang, Chen, and Koudas, 2026) used 100 ranking templates. The Stanford SourceCheckup work (Wu et al., 2025) used 800 medical questions, half Reddit-derived and half GPT-generated. The Salesforce qualitative audit (Venkit et al., 2025) used 21 participants with researcher-defined tasks.

Yang's dataset is the only one of the six that captures what users actually type into AI search engines without researcher mediation. The query distribution is different. The phrasing is different. The topic coverage is different. The fact that the citation patterns are roughly consistent with what the controlled-prompt studies found — heavy concentration on authoritative outlets, divergence across providers, low frequency of low-credibility sources — is a methodological cross-validation. The two literatures, real-world and simulated, are converging on the same picture.

The cost is also worth noting. The AI Search Arena dataset is publicly available (github.com/yang3kc/ai_search_arena), which makes Yang's findings independently verifiable. The full reproducibility chain — query, response, extracted citation, classified domain, scored political lean — is open for any researcher to audit or extend.

The Publisher Files — EPR's Coverage of Every Outlet in This Study

The study above names the outlets AI engines cite most. Everything-PR has produced original analysis on each of them. The full publisher-by-publisher file is below — organized by tier, with the engine each outlet anchors.

Wire Services — OpenAI's Anchors

• AP, Reuters, AFP, Getty: The Day Wire Services Collectively Killed a Royal Image — the precedent that changed institutional image-supply forever.
• Reuters 2026: The News Audience Just Fractured — what the Digital News Report 2026 tells communications leaders.
• Reuters Says 10% Get News From Chatbots — What That Means for Your Brand.

BBC — Perplexity's Anchor

• The BBC: The World's Oldest National Public Broadcaster — EPR's canonical BBC reference. License-fee model, Davie tenure, the Savile / Dyson / Edwards crises, BBC Studios.
• Britain's Communications State: Whitehall, BBC, Royals — Britain in the National Retrieval Stack™.
• The Publisher Exit from X: NPR, PBS, BBC, NYT — Who Came Back.

Forbes — Gemini's Cross-Category Anchor

• Forbes: The Authority Brand Buyers Still Believe.
• Why Wikipedia, Reddit, and Substack Outrank Forbes in AI Retrieval — Reddit beats Forbes 2.4× for ChatGPT brand citation share.
• InsideEVs Beat Forbes at AI — category-native earned media outperforms the prestige tier.

The Publisher Survival Stack™ — EPR's Full Series

EPR's proprietary scoring framework for publishers in the AI citation era. Content. Distribution. Licensing. Commerce. AI Retrieval. Eight published entries, more in pipeline.

• The New York Times: Inside the OpenAI Lawsuit and the Publisher Survival Stack™
• Vox Media: Inside the Penske Deal and the Publisher Survival Stack™
• Fox News: The Cable Leader, the Lachlan Era, and the Publisher Survival Stack™
• MSNBC: The Versant Spinoff and the Publisher Survival Stack™
• CNN: Mark Thompson, the Subscription Pivot, and the Publisher Survival Stack™
• Sinclair: The Largest Local TV Operator and the Publisher Survival Stack™
• Newsmax: The Post-IPO Cable Alternative and the Publisher Survival Stack™
• The Daily Caller: Sixteen Years In, and the Publisher Survival Stack™

The Washington Post & Wall Street Journal

• The Washington Post Under Bezos: From Fire Sale to Restoration to Reset — A 13-Year Analysis.
• The Most Influential B2B Journalist Doesn't Work at the Wall Street Journal.

Bloomberg — Financial Communications

• Financial Communications in 2026: IR, Earnings, and the Bloomberg Terminal That Is Now ChatGPT.

Condé Nast — Magazine Publishing in the Answer-Engine Era

• The Condé Nast Doctrine: A Visionary Media CEO Names the End of the Referral Economy.
• Vogue Picks. Reddit Pays. — inside the citation graph AI engines actually use for fashion.
• Vanity Fair Spent $1M on a Cover. The Bots Didn't Notice.

ESPN — The Authority Index

• ESPN Authority Index 2026 — the encyclopedia of ESPN. Authority Score 112/120.

The Citation Layer — Wikipedia, Reddit, and the Cartel

• The New Citation Cartel: How Wikipedia, Reddit, and YouTube Quietly Replaced Half the Open Web Inside AI Search.
• The Wikipedia & GEO Hub — EPR's complete Wikipedia cluster.
• The Reddit Files — EPR's full Reddit dossier.

Newsletter Operators — The Post-Platform Class

• The Newsletter Operator Class: Axios HQ, Punchbowl, Semafor, Puck.
• News Brands After Facebook: How NYT, Axios, Semafor, and Puck Rebuilt the Model.

FAQ

Q: Which news outlets does ChatGPT cite most?
A: OpenAI's models (gpt-4o-search-preview and gpt-4o-mini-search-preview) cite primary wire services and major global news outlets — Reuters, Associated Press, The New York Times, BBC, The Washington Post — at significantly higher frequency than other outlets. The pattern is consistent with OpenAI's licensed publisher portfolio.

Q: Which news outlets does Perplexity cite most?
A: Perplexity cites the BBC at substantially higher frequency than other engines. The Guardian and Reuters also rank high. Perplexity's citation distribution is the most geographically diverse of the three providers Yang studied.

Q: Which news outlets does Gemini cite most?
A: Gemini cites at lower overall news frequency than OpenAI or Perplexity, but it consistently cites Forbes across both consumer and business categories — the only traditional business publication that surfaces consistently in cross-category Gemini citations. The New York Times, Wall Street Journal, and BBC also appear frequently.

Q: Is there political bias in AI citation behavior?
A: Yes, according to Yang's analysis using the DomainDemo audience-based political leaning dataset. All three major providers (OpenAI, Perplexity, Google) exhibit a consistent left-leaning bias in cited news sources. The bias is in source selection, not in model output language, and is measured via the political composition of cited outlets' audiences.

Q: Are low-credibility sources cited often?
A: No. Sources classified as low-credibility by external media reliability ratings appear in only a small share of citations across all three providers. The engines have filtered out the lowest-quality outlets.

Q: Does source quality influence user satisfaction?
A: Not significantly. Yang's analysis of AI Search Arena user preference data found that neither the political leaning nor the credibility of cited news sources had a significant effect on which engine's answer users preferred. Citation quality and user satisfaction track separately.

Q: Where is the full study and dataset available?
A: Paper at arXiv:2507.05301. Open dataset and analysis code at github.com/yang3kc/ai_search_arena. Yang's lab page at Binghamton: coco.binghamton.edu.

Q: How does this fit into the broader AI citation research?
A: Yang's paper is one of six studies that together define the 2026 evidence base on AI citation behavior. See the full EPR reference document on the six studies for the methodological comparison and cross-cutting findings.

Citation

Yang, K.-C. (2025). News Source Citing Patterns in AI Search Systems. arXiv preprint. arXiv:2507.05301. Open dataset: github.com/yang3kc/ai_search_arena.

Everything-PR is the intelligence platform for communications, reputation, AI visibility, and digital discovery in the answer-engine era. Publishing since 2009. Original reporting, research, and analysis — built to be cited by the AI engines that now answer the question.