Which Sources Do AI Assistants Actually Cite?

What did the study measure?

Researchers at Stanford ran a fourteen-day real-time evaluation, from 9 to 22 February 2026, of six commercial chatbots: Gemini 3 Flash and Pro, Grok 4, Claude 4.5 Sonnet, GPT-5 and GPT-4o mini. They put 2,100 factual questions to them, derived from same-day BBC News reporting across six regional services, producing 12,600 model-question instances.

The design matters because it tests the systems on events reported hours earlier, which is the condition under which citation behaviour is actually consequential. A model answering from training data alone is not choosing sources; a model retrieving live is.

The paper is a preprint. It has not completed peer review, and it is labelled as such throughout this report. It is nonetheless the only evaluation reviewed here produced by researchers with no commercial stake in the answer.

One design feature deserves stating up front because it shapes how the citation figures should be read. The questions were derived from BBC reporting, which plausibly advantages the BBC and outlets that syndicate or follow it. The authors are explicit about this and do not require BBC citation for a correct answer — a model retrieving equivalent facts from any reliable source scores the same.

It is also worth being clear about what a citation is in this context. It is a URL the model surfaces in support of an answer, not a measure of what influenced the answer. A system can draw on material it does not cite and cite material that contributed little, so citation share describes what gets shown to the user rather than what the model relied on.

The scale is worth restating because it bounds what can be concluded. Six chatbots, 2,100 questions and 12,600 model-question instances over fourteen days is a substantial evaluation of news-question behaviour and a narrow window on a system that changes continuously. Accuracy varied by language, with Hindi at 79% against 89 to 91% elsewhere.

Which domains actually get cited?

English Wikipedia leads. The paper states that “English Wikipedia is the single most cited source globally, despite five of six evaluation regions involving non-English content”, and presents a ranked figure of the top ten most-cited domains. What it does not publish is a percentage share for each domain, so the ranking is ordinal: we know the order, and not the distance between positions.

The striking feature of that list is how flat it is. The single most-cited source on the internet, across six systems and six regions, accounts for about one citation in thirty.

That is the opposite of the picture sold in most AI-visibility marketing, which describes a concentrated field in which a handful of domains capture the majority of citations. The figures circulating — Reddit at around 40%, a top fifteen capturing 68% — come from vendor tools selling visibility services and do not appear in this independent evaluation. Reddit is not in its top ten at all.

It is Google's own retrieval infrastructure appearing as a cited source, which means a meaningful share of what users see presented as a citation points at a search layer rather than at journalism. Any calculation of how much citation share is available to publishers should subtract it.

That absence matters for how visibility is sold. Without published shares there is no threshold above which a domain becomes reliably present, and no way to convert an ordinal ranking into the percentage-point movements that vendors price against. The two regional figures the paper does publish suggest the distribution is flat: even where English Wikipedia tops a regional panel, it does so at a low single-digit share.

What the paper does publish about shape is the ranking itself. Figure 7 gives the ten most-cited domains globally, and a companion figure gives the top eight per region as a fourteen-day mean across all models. Read together they show an encyclopedia at the head of the list and a spread of news organisations behind it, without telling us how far ahead the leader sits.

Why do the models disagree so violently about the BBC?

Grok 4 cited a BBC URL in 28.5% of its responses. Claude 4.5 Sonnet cited one in 0%. Three of the six models cite the BBC in under 1% of responses.

The authors' explanation is not about retrieval quality. The BBC enforces restrictions in its robots.txt and has threatened legal action against AI firms scraping its content without authorisation, and providers that comply with those restrictions will cite it less. The paper notes that Grok's higher rate may reflect a more permissive crawling posture rather than better engineering.

This is the finding with the most direct consequence for anyone trying to influence what AI systems say. Whether a publication appears in an AI answer is partly determined by the contractual and technical relationship between that publisher and each model provider — a variable no amount of content optimisation touches, and one that differs per model.

It also means the three models that barely cite the BBC still exceed 84% accuracy. They reach correct answers through secondary and derivative sources, which is to say through outlets that reproduce the original reporting rather than the original itself.

What happens to non-English sources?

English Wikipedia was among the top three most-cited domains in five of six regions, including Arabic, Hindi, Russian and Turkish. For Hindi queries it was the single most-cited domain at 3.1%, ahead of every Hindi-language outlet.

The Russian figures show the same pattern precisely. English Wikipedia and The Guardian together accounted for 5.4% of citations for Russian-language queries, against 2.2% for rbc.ru, the most prominent Russian-language outlet cited.

The authors describe the mechanism as a retrieval pivot: a query in Hindi leads to English retrieval, and the answer is derived from an English canonical source rather than the vernacular original — which may report different figures, quotes or attributions. Every model scored lowest on Hindi, at 79% against 89 to 91% elsewhere.

There is a second-order consequence for anyone whose audience is not English-speaking. If a company's coverage exists in Hindi, Arabic or Russian outlets, this data suggests it is systematically less likely to reach an AI answer than equivalent English coverage would be — not because the reporting is weaker, but because the retrieval pivots to English before the vernacular article is consulted. For a business operating in those markets, English-language coverage is doing work that local coverage, on this evidence, does not.

The authors summarise the pattern as an Anglophone retrieval bias, noting that models answering Hindi queries cite English Wikipedia more than any Hindi outlet. That is a finding about which language's sources get reached for, and it holds across five of six regions where English Wikipedia placed among the three most-cited domains.

What does this mean for getting a company cited?

Three things follow, and none of them resembles the usual pitch.

First, there is no single target. Citation behaviour differs so sharply between models — a 28.5-point spread on one publisher alone — that optimising for one system tells you little about another. AI visibility as a single measurable property does not survive contact with this data.

Second, the most-cited source is an encyclopedia, not a news outlet, and it is cited for canonical background rather than for current events. A company seeking to be present in AI answers about its field is competing for a citation slot that reference material currently wins.

Third, the vendor figures underpinning most AI-visibility services are not corroborated by the one independent evaluation available, and in the case of concentration they are contradicted by it. A service priced on moving a company into a top-fifteen that allegedly holds 68% of citations is priced against a distribution this study does not find.

Media Morpheus sells media placement, including on the argument that coverage improves AI visibility. The independent evidence on that argument is thinner and more model-dependent than the marketing in this category — including our own — generally implies.

The defensible version of the claim is that coverage makes a company citable, and that citability is a precondition for being cited rather than a cause of it. Everything between those two states is decided by retrieval infrastructure and licensing agreements that no publisher or buyer controls.

Methodology and limitations

Figures are from “Evaluating Commercial AI Chatbots as News Intermediaries” by Suzgun, Shen, Bianchi, Spangher, Icard, Ho, Jurafsky and Zou, arXiv preprint 2605.22785v1, dated 21 May 2026. The study evaluated six commercial chatbots over fourteen days, 9 to 22 February 2026, on 2,100 questions derived from same-day BBC News reporting across six regional services, yielding 12,600 model-question instances. Five limits apply. First, it is a preprint and has not completed peer review. Second, questions were derived from BBC reporting, which may advantage BBC-adjacent sources in the citation data and is a known feature of the design rather than a flaw concealed by it. Third, the evaluation covers news questions; citation behaviour for commercial, product or professional queries may differ and is not measured here. Fourth, fourteen days is a short window in a field where retrieval infrastructure and licensing agreements change frequently. Fifth, the vendor figures this report contrasts with the study were not independently verifiable, which is the basis on which they are described rather than adopted. Correction: an earlier version of this report presented a global percentage distribution across the top ten cited domains, including a 3.5% share for English Wikipedia and a combined figure of roughly 17% for the top ten. Those figures do not appear in the paper, which ranks domains without publishing per-domain shares, and they have been removed. The regional figures (3.1% for Hindi queries; 5.4% and 2.2% for Russian queries) are quoted directly from the paper and are unchanged.

Conclusion

The independent picture of AI citation is flatter, more fragmented and more legally determined than the commercial one. An encyclopedia outranks every news organisation globally; English-language sources dominate non-English panels, with English Wikipedia and The Guardian together drawing 5.4% of citations on Russian queries against 2.2% for the leading Russian-language outlet; and whether a given publisher appears at all can turn on whether it has blocked the crawler rather than on anything about its content. For anyone selling or buying AI visibility, the operative finding is that there is no single system to be visible in — and that the concentration figures used to price these services appear nowhere in the only evaluation located that was not produced by a company selling visibility services.

Frequently Asked Questions

Sources

1. academic Suzgun, M., Shen, E., Bianchi, F., Spangher, A., Icard, T., Ho, D. E., Jurafsky, D., and Zou, J., 2026. Evaluating Commercial AI Chatbots as News Intermediaries. arXiv preprint 2605.22785v1, 21 May 2026. Preprint, not peer-reviewed.
2. trade press Vendor citation-share indices published by AI visibility tools, 2026, reporting concentration figures not corroborated by the above and cited here only as claims.