Where Does Your Chatbot Get Its Information From?

MJRC’s study of ten AI systems across four countries reveals which sources shape their answers, and how changing the chatbot or language changes what you read.

A reader asks an AI assistant about Germany in English. Another asks in German. The subject is the same, but the sources behind the answers can look remarkably different. In the first case, German broadcasters and newspapers supply 7.4% of citations. In the second, their share rises to 18.8%.

The gap is larger in Hungary. Hungarian-language domestic newsrooms account for 3.7% of citations generated by English questions and 20.3% when the questions are asked in Hungarian. A change in language brings much more of the country’s journalism into view.

These are two findings from the first four published country editions of the Media and Journalism Research Center’s AI Information Map. Together, the studies of Germany, Hungary, Romania and Spain suggest that an AI answer has an editorial dimension before anyone assesses whether its sentences are correct: somebody’s reporting, statistics or institutional account has been selected to explain the world. Findings from more countries will be published soon as MJRC’s AI Information Map expands its examination of the sources behind chatbot answers.

The emerging picture is more complicated than a handful of dominant publishers swallowing all the attention. Across each country’s pooled results, the leading source accounts for less than 5% of citations. Yet that broad distribution conceals sharp differences between languages and assistants, alongside a recurring gap between the journalism systems recommend and the sources they actually use.

The identity of the leading source still varies. Spain puts El País first, at 4.9%; Germany puts Reuters first, at 4.3%. Hungary’s pooled EU institutional entry leads on 4.7%, while the European Commission heads Romania’s list on 4.0%. Only in Spain does a domestic newspaper take the top position. An apparently similar level of concentration can therefore conceal quite different kinds of authority.

The language of the question changes the newsroom

In Spain, El País receives 3.0% of citations when prompts are in English and 6.2% when they are in Spanish. The European Commission and other EU institutional sources move the other way, falling from 2.4% to 0.8%. The language switch changes the balance between a newspaper reporting on Spain and institutions describing it.

Germany offers a similar contrast. Tagesschau’s share rises from 0.8% in English to 3.4% in German; ZDF’s increases from 0.2% to 2.7%. In Hungary, Telex rises from 1.1% to 3.4%. These are shares of the sources appearing in each language layer, not percentages of users or answers.

Romania shows how an outlet can benefit from the reverse movement. Romania Insider, which publishes in English, accounts for 4.6% of citations from English prompts and 2.2% from Romanian ones. It ranks third overall, behind the European Commission and OECD. A mainstream news broadcaster, Digi24, meanwhile, leads the Romanian-language source list at 4.1%.

For a publisher, this raises a question that audience size alone cannot answer: in which language does its work become visible to an assistant? For readers, it means that choosing English for convenience may change the selection of sources they encounter.

But translation is not the only thing happening.

Nor does a national-language question automatically produce a less official account. In Germany, official and government-linked sources also gain share when prompts switch to German. More domestic journalism and more institutional material can appear together. The relevant question is which voices enter the answer, and in what proportions.

Ten assistants, different reading lists

Switching assistants can alter the source mix just as sharply. In Spain, ChatGPT’s leading source is El País, at 13.1%. Gemini’s is La Moncloa, the Spanish government’s website, at 4.5%. DeepSeek’s is the news agency EFE, at 3.5%. The same country can be introduced through a newspaper, a government portal or a wire service.

Romania offers another contrast. Claude gives Romania Insider 9.0% of its citations, while DeepSeek gives Radio România Internațional 14.5%. Digi24 leads for both Perplexity and Google AI Overviews, at 3.2% and 4.2%, respectively. A pooled country ranking smooths over these differences between the products people actually choose.

In Hungary, pooled EU institutional sources account for 16.2% of ChatGPT’s citations; Google AI Overviews puts Telex first, at 4.3%. In Germany, official and government-linked sources together provide 42.9% of ChatGPT’s citations. Claude’s leading German source is Wikipedia, while Google AI Overviews leads with YouTube.

None of these figures establishes which assistant produces the best answer. A statistical office may be appropriate for a statistical question. But a ministry’s account and a newsroom’s scrutiny of that ministry serve different purposes. Counting sources without examining their roles can obscure that distinction.

The comparisons describe samples collected over different periods. They reveal where visibility goes, but cannot establish whether a pattern results from retrieval, available material, a commercial arrangement or the wording of a question.

Recommended does not mean used

One of the most revealing distinctions in the project is between using an outlet to construct an answer and explicitly recommending it. These are different forms of visibility, measured against different totals. An endorsement tells the reader where to go; a citation identifies a source brought into the answer itself.

In Germany, Süddeutsche Zeitung receives 3.7% of recommendations but 0.5% of citations. Der Spiegel receives 2.7% and 0.3%, respectively. Hungary’s Telex takes 6.0% of endorsements and 2.4% of citations. These are not conversion rates, but the contrast is revealing: a newsroom can occupy a much larger place in an assistant’s advice than in its sourcing.

That distinction matters for how publishers interpret AI visibility. Being named on a list of trustworthy outlets does not establish that a newsroom’s reporting routinely informs answers on politics, schools or public services. Nor does either measure show how many readers click through. The project measures appearances and roles, not referral traffic or revenue.

The editorial implication is that recognition and use need separate attention. A system may be able to name a range of credible publications when directly asked, yet draw on a narrower or more institutional set of sources for everyday questions. A reassuring reading list does not settle the pluralism question.

Question design matters, too. In Germany, 96.9% of recommendations of regional newspapers came from one question asking for local sources outside Berlin. That result is useful evidence of what assistants can surface when prompted. It also cautions against treating a pooled recommendation list as a portrait of what users would encounter across every topic.

Who gets a warning?

The stakes change when an assistant moves from choosing sources to warning readers against them. The project separates a caution from the stronger claim that a source is associated with disinformation. These are records of what the assistants said, not independent verdicts by MJRC.

Bild heads Germany’s caution list, accounting for 17.5% of cautions. Hungary’s pooled public-media entry, MTVA and its associated services, leads there with 11.2%. România TV heads Romania’s list at 10.8%, followed by Antena 3 CNN and Realitatea Plus/TV, both on 8.9%.

Spain’s largest share of cautions goes to OKDiario, at 6.5%, followed by the public broadcaster RTVE at 4.5%. Platforms also feature: X accounts for 3.9% and Facebook for 3.2%. The warning lists therefore mix publishers with channels through which information circulates.

The stronger flags produce a different map. In Germany, Doppelgänger-style impersonation sites account for 15.7% of disinformation flags, ahead of RT DE and its mirrors on 13.3%. Hungary’s public-media grouping leads that category too, at 12.6%, followed by Origo at 10.7%. In Romania, Sputnik and ActiveNews each account for 11.1%.

These percentages describe each source’s share of warnings in that category, not the proportion of its output that is false. Spain has no recorded disinformation labels for Spanish sources. That absence describes the assistants’ behaviour in the sample; it does not establish that disinformation is absent from the country.

The Hungarian findings carry an additional qualification. The project warns that many judgements refer to ownership and political arrangements from the Orbán period and may not describe the outlets today. An assistant can carry an established reputation into an answer after the circumstances behind it have changed.

The evidence supplied with the warnings is often thin. In Hungary, 16.1% of cautions and 15.5% of disinformation flags came with a supporting citation or link. Romania’s corresponding figures are 43% and 17%. In both countries, fewer than one in five of the more serious accusations arrived with a reference.

A link is only a starting point: it does not establish the quality of the justification. Without one, a reader has even less to examine and a newsroom less to challenge. The question is not merely which sources the assistant distrusts, but how a user can assess its reasons.

Whose work reaches the reader?

All four studies examine free versions of ten systems, using questions in English and the principal national language across news, politics, public services and everyday life. Their published shares come from coded material. The much larger archive totals described in the methodology should not be mistaken for the denominator behind every chart.

The country batches span different collection and validation windows across 2025 and 2026. The databases continue to change. Institutional websites are sometimes pooled, while media outlets are generally counted by brand. These choices matter when comparing a group of government websites with one newspaper. The findings also do not constitute an accuracy test of the answers.

Within those limits, the first four countries point towards a practical test for AI-mediated news: whose work appears when the user asks an ordinary question, rather than explicitly requesting a diverse reading list? The answer needs to be checked across languages and products, with links that allow readers to follow the evidence.

The concern is not simply whether an assistant knows that a newsroom exists. It is whether that newsroom’s reporting gets a place in the account of public life that the assistant delivers.