AI Information Map: Germany

AI Information Map · Germany

Germany, as the assistants source it

20 questions, 10 assistants, English and German, some 350,000 archived answers. These pages show which German sources the systems draw on, which they recommend, and which they warn against.

Contents

Research on Germany is continuing. Visibility results are given as percentages of the slice under discussion; counts describe the design of the study rather than its results and are labelled where they occur. Figures are revised as new collection rounds are added. The full list of sources can be requested from MJRC.

AI Information Map · Germany

How the study was run

What was asked, of which systems and in which languages, and how every source an assistant put on screen was logged and coded.

The assistants and the prompt bank

10 widely used assistants were tested using the same prompt bank and coding protocol: Perplexity, ChatGPT, Gemini, Copilot, Claude, DeepSeek, Meta AI, Grok, Mistral Vibe (formerly Le Chat) and Google’s AI Overviews. Each received a fixed bank of 20 questions about how people find out what is happening in Germany, covering five areas: media and information sources, politics and governance, economy and public services, society, rights and environment, and culture, sport and local life.

The bank contains three kinds of question. Some ask for a general overview, such as the main political issues or the biggest sports stories. Some ask directly which sources to trust or to treat with care. The rest are written from the point of view of a user with a practical need: a voter comparing party plans on corruption and public services, a German considering a return home, a family relying on public healthcare, a visitor travelling outside Berlin. Each question was asked in English and in German.

Only the free tiers were used, the versions anyone can access without paying. Paid and enterprise tiers may answer differently, so the findings describe the experience of an ordinary user.

Scale, repetition and geography

The 20 prompts published with these findings are illustrative rather than exhaustive. Each underlying question was reformulated into roughly 9,000 phrasings, so that no single wording could drive a result, which yielded on the order of 180,000 prompt runs and approximately 350,000 archived answers for this batch. Requests were issued through 65 IP addresses inside and outside Germany, to reduce dependence on a single network location and to observe possible geographic variation. The collection window ran from May 2025 to June 2026, and a concluding round of validation prompting was carried out from August to October 2026.

What the figures are based on

AI answers change over time, which is why the research is carried out continuously and the archive keeps growing. The figures on these pages come from a coded sample that summarises the archive for the whole study period: the full set of 20 prompts, run through all 10 systems in English and German and coded source by source. Every source an assistant used or cited was coded, as was every explicit judgement it made about a source. Recommendations make up 83.0% of those judgements, cautions 10.0% and disinformation-related flags 6.9%. Each judgement is counted once per source or channel in each answer. Of the sources used or cited, 90.4% carried a usable web address. Shares will move as further rounds are coded and added.

How sources were coded

Each answer was stored with its prompt, the assistant, the language, the date, and any links, citation markers or source panels shown. Every source the assistant produced was then extracted and given one role:

• Used or cited: drawn on to build the answer
• Recommended: named to the reader as worth following
• Cautioned: named as something to avoid or treat with care
• Flagged as disinformation-related: named as propaganda, manipulation or falsehood

Sources are counted by brand, not by web address. Wikipedia’s English and German editions count once, as do the English and German sites of Deutsche Welle, Der Spiegel and Die Zeit. Tagesschau is counted as ARD’s national news service, and ZDFheute together with ZDF. ARD’s regional broadcasters, including NDR, WDR, BR and MDR, are counted separately because each runs its own newsroom. ARD-faktenfinder and dpa-Faktencheck are counted on their own when an assistant names the fact-checking desk. Links to a portal whose entries are labelled as AFP planning notes are counted as AFP, following the label the assistant displayed.

Three pooled groups appear in the charts. “Federal ministries (pooled)” combines the websites of individual federal ministries; the Federal Government’s own website, the Bundestag and federal agencies such as Destatis and the Umweltbundesamt are counted separately. “EU institutional sources (pooled)” combines Commission, Council, Parliament and agency pages. “City portals (pooled)” combines the city portals of Berlin, Hamburg, Munich, Cologne, Leipzig and other cities. Pooled groups are broader than individual media brands and are not equivalent units: a pooled group ranking above a newspaper means that a whole set of institutional websites outweighs that newspaper, not that a single website does.

Two grades of warning

The assistants distinguish between advising a reader to treat a source with care and describing a source as propaganda or manipulation. All 10 systems did both in the German sample. Finding 06 reports the two separately because they differ in gravity.

Collection limits

Some systems were harder to capture than others, and the gaps have been recorded rather than repaired. ChatGPT’s exported answers kept the names and domains shown on its source cards but lost most of the underlying article links, so its link rates understate what a live user would have seen and are marked as not comparable in Finding 07. Perplexity’s free tier required several exports to capture in full. Some German Gemini answers contained restart artefacts; repeated sources in them were counted once. No missing link was reconstructed or looked up, and 2.5% of coded runs remain flagged for review because sources were refused, hidden or could not be recovered.

Systems and language layers were not always collected at the same time, so differences in current-affairs sourcing may partly reflect changes in the news cycle. Sequential prompts within a session may also carry conversational context forward.

How the results are expressed

Because the mapping is continuous, visibility results are given as percentages: an outlet’s share of the visibility within the slice under discussion, whether the whole coded sample, one assistant, one language, one subject or one role. Running totals of mentions are not reported because they rise with every round. Where counts appear, they describe the design of the study, such as the number of assistants, languages and IP addresses, the scale of the run, the number of distinct outlets behind a percentage, or how many assistants took a given position, and are labelled as such.

Overall source shares weight each source appearance equally. Assistants that produce more citations therefore contribute more to pooled percentages than assistants that cite sparsely. Grok and Claude, for example, each contribute more than three times as many source appearances as Gemini.

The work is not finished. Germany’s database remains open and these figures will be updated as new rounds arrive. Each finding describes the state of the evidence now.

What this study does not do

It records where sources appear and what the assistants say about them. It does not assess whether an answer was accurate, balanced or complete. The study does not make formal assessments of ownership, editorial independence or political capture. For descriptive purposes, Finding 01 groups sources by broad institutional type: official source, public broadcaster, private press, wire agency or platform. Appearing on any list in these findings reports what an assistant said, not MJRC’s view of the outlet.

Where the visibility goes

Official Germany answers first

Reuters is the most-cited single source on Germany. Taken together, official and government-linked sources account for roughly as much of what the assistants cite as the country’s broadcasters and newspapers combined.

No single source dominates. Reuters, the most-cited brand, accounts for 4.3% of all sources used to answer questions about Germany. The pooled websites of the federal ministries follow with 3.0% and Wikipedia with 2.4%. Together, the top three reach 9.7%. Beyond the 15 entries shown below, 73.5% of citations are spread across 1,308 other outlets.

Institutions dominate the top of the list. Seven of the 15 leading entries are German official or government-linked sources: the federal ministries, deutschland.de, the Federal Statistical Office (Destatis), the Federal Government’s website, the Bundestag, the Federal Environment Agency (Umweltbundesamt) and the city portals. The pooled EU institutions form an eighth institutional entry. As a group, official and government-linked sources account for 14.0% of all citations, public broadcasters for 7.3% and the private press, national and regional, for 5.2%.

The grouping is indicative, and the balance depends on where the line is drawn. Counting deutschland.de, which a private publisher produces in cooperation with the Federal Foreign Office, and the city portals as official sources tips it towards the state; leaving them out tips it the other way. On either reading, official and government-linked sources weigh about as much in the assistants’ answers as all of Germany’s broadcasters and newspapers put together.

Public broadcasters are the strongest journalistic presence. Tagesschau ranks fourth overall with 2.0%, Deutsche Welle sixth with 1.5% and ZDF 11th with 1.3%. Newspapers rank lower. Die Zeit, with 1.0%, is the only newspaper in the top 15. Süddeutsche Zeitung has 0.5%, and Der Spiegel and the Frankfurter Allgemeine 0.3% or less. Reuters alone outranks every German newsroom, and AFP is just behind Tagesschau.

German journalism is part of the picture, and the long tail of the data includes regional dailies, research institutes, statistics offices, foundations and specialist titles. In answers to everyday questions, however, the assistants tend to cite a ministry, an agency or a wire service before they cite a German newsroom.

The sources assistants draw on most

Who supplies the sourcing, by type of source

Each bar in the first chart is one outlet’s share of all sources used to build answers, pooled across 10 assistants and both languages; entries marked “pooled” are groups of websites, not single brands. The second chart groups outlets by broad institutional type for descriptive purposes. “Official and government-linked sources” covers federal ministries, the Federal Government’s website, the Bundestag, federal agencies, the KMK, deutschland.de and city portals; “public broadcasters” covers ARD, Tagesschau, ZDF, Deutschlandfunk, Deutsche Welle and ARD’s regional broadcasters.

Same questions, different answers

Each assistant builds its own Germany

The same 20 questions produce very different source lists. One assistant takes more than two-fifths of its citations from official websites; two others rely heavily on a single wire agency each.

ChatGPT relies most on official sources. Federal ministries, the Federal Government’s website and Destatis top its list, and official and government-linked sources make up 42.9% of everything it cites, nearly three times the average across the 10 systems. It also draws on the narrowest range of outlets of any assistant, less than a third of Claude’s.

Two assistants depend heavily on a wire agency. Meta AI takes 27.0% of its sources from Reuters, across all five subject areas. Mistral Vibe takes 18.9% from AFP. More than half of those AFP citations point to a portal whose entries are labelled as AFP’s German-language daily planning notes (Tagesvorschau), and all of them appear in answers to questions asked in English. Mistral and AFP announced a content agreement in January 2025; the pattern coincides with that integration, although this study cannot establish causation.

Claude sits at the other end of the range. It does not directly cite a wire-agency domain, leads with Wikipedia (6.9%) and cites a wider range of outlets than any other system. Grok has the highest share of German newspapers, 13.3% of its sources, and is the only assistant that cites the Swiss weekly Weltwoche, often on party politics. DeepSeek relies on Deutsche Welle and Deutschlandfunk and accounts for most of the study’s citations of Turkey’s state news agency, Anadolu. Google AI Overviews takes 16.0% of its sources from platforms such as YouTube, Facebook and Reddit, the highest share of any system.

Which assistant a person opens strongly shapes which sources they meet. ChatGPT explains Germany mainly through its ministries, Meta AI through Reuters and Grok through newspapers. Source pluralism therefore has to be assessed product by product, not for AI systems as a whole.

What each assistant reaches for

How much they cite against how widely they range

Volume and range are separate measures. Grok produces the highest volume of citations, at 0.45 distinct outlets per citation. Claude cites almost as much and has the widest range, at 0.61. ChatGPT returns to the same sources most often, at 0.36. A long list of sources can still draw on a small number of outlets.

Volume is indexed to the heaviest citer, Grok, and range to the widest-ranging system, Claude, each set to 100. “Outlets per citation” divides distinct outlets by total source appearances: the higher the figure, the less an assistant repeats itself. The ratio favours assistants that cite little, so Gemini’s high figure reflects sparse citation rather than breadth. ChatGPT’s sources are counted from the names and domains on its source cards; see the collection limits in the methodology.

What the language of the question decides

The newsrooms arrive with the German language

Asked in German, the assistants take 18.8% of their sources from German broadcasters and newspapers. Asked the same questions in English, they take 7.4%.

German-language prompts bring German journalism into the answers. Tagesschau rises from 0.8% of English-prompt citations to 3.4% of German ones, ZDF from 0.2% to 2.7% and Deutschlandfunk from 0.3% to 1.1%. Die Zeit and Süddeutsche Zeitung at least double their shares. Public broadcasters and the private press together account for 18.8% of German-prompt citations and 7.4% of English-prompt citations, a ratio of about two and a half to one.

English-prompt answers lean on international and institutional sources. Reuters leads with 5.0%, followed by Wikipedia with 3.5%, more than three times its share in German. AFP, the EU institutions and YouTube also rank higher in English than in German. English-language titles about Germany, such as The Local Germany, IamExpat and Expatica, take 2.7% of English-prompt citations and 0.9% of German ones. Clean Energy Wire, a Berlin-based English-language service on energy policy, appears only in English answers. Deutsche Welle is the only German broadcaster with a higher share in English (2.0%) than in German (1.0%), in line with its international remit.

Official sources also gain in German. Their share rises from 12.1% of English-prompt citations to 16.4% of German ones. The Bundestag goes from 0.5% to 2.1%, and the federal ministries from 2.3% to 3.9%. Answers in German are both more journalistic and more official.

The gap affects identifiable groups of users: foreign correspondents in Berlin, EU officials, researchers who do not read German, families considering a move to Germany. When they ask in English, they receive wire services, Wikipedia and English-language explainers. Tagesschau ranks 14th in the English layer and ZDF is outside the top 60. English-language users see Germany mostly through international and institutional sources, with little from German newsrooms.

Asked in English

Asked in German

Bars give each outlet’s share of citations within its own language layer, and each chart is scaled independently.

Subject by subject

Each subject has its own gatekeeper

Questions on rights, the environment and public services are answered mainly from federal agencies. Political questions bring in the Bundestag and the parties’ own websites. Culture questions are answered from Wikipedia and city portals.

Society, rights and environment is the most official subject area. Official and government-linked sources supply 22.6% of its citations, led by the Umweltbundesamt (4.6%), the federal ministries (4.1%) and Destatis (3.7%). Economy and public services follow the same pattern, with the ministries alone at 8.0%. In English, the EU institutions, Clean Energy Wire and Human Rights Watch move up these lists; in German, the ministries and the Federal Government’s website lead.

Politics is where German journalism and German institutions meet. Reuters and Wikipedia lead overall. In the German layer, the Bundestag comes first with 7.0%, followed by ZDF with 5.3%. Party websites account for 9.9% of all politics citations, almost all of them in answers to two questions: which parties have credible plans on corruption and public services, and which has the most progressive education agenda. When asked to compare parties, the assistants often rely on the parties’ own descriptions of themselves. Die Linke’s website is cited most, followed by those of the SPD and the Greens, a ranking that partly reflects the wording of the education question. The CDU, the largest party in the governing coalition, is cited much less, and the AfD’s website hardly at all.

Culture, sport and local life are answered from reference works and municipal websites. Wikipedia leads with 5.2%, ahead of the pooled city portals with 4.2%, the national tourist board, the Frankfurter Buchmesse and the Bundesliga. In German, Tagesschau and Süddeutsche Zeitung enter the top five; in English, the Buchmesse and the Bundesliga’s own website rank higher.

On media questions, the assistants rely first on wire agencies and research. Reuters (8.4%) and AFP (6.0%) lead, followed by Tagesschau, the Reuters Institute’s Digital News Report (4.5%) and deutschland.de (4.5%). The Leibniz Institute for Media Research, which produces the German report for the Digital News Report, adds another 3.8%. A large share of what the assistants say about German journalism comes from one annual survey, cited through two institutions.

Leading sources in each subject, split by the language of the prompt

Bars give each outlet’s share of citations within that subject. The tags beneath show which sources lead once the subject is split by prompt language.

What the machines say to trust

Vouched for, seldom used

All 10 assistants recommend Deutsche Welle, Tagesschau, ZDF, Süddeutsche Zeitung, Die Zeit and Der Spiegel. Most of these titles are rarely cited in the answers themselves.

Recommendations are dominated by German journalism. Deutsche Welle leads with 4.9% of all recommendations, followed by Süddeutsche Zeitung (3.7%), Tagesschau and Reuters (3.1% each). Six titles are recommended by all 10 systems: Deutsche Welle, Tagesschau, ZDF, Süddeutsche Zeitung, Die Zeit and Der Spiegel. Public broadcasters together receive 22.0% of recommendations, almost three times their share of citations.

The gap between what the assistants recommend and what they use is large. Süddeutsche Zeitung receives 3.7% of recommendations but only 0.5% of citations. Der Spiegel receives 2.7% and 0.3%, the Frankfurter Allgemeine 2.1% and 0.2%. In the opposite direction, the federal ministries and Wikipedia, second and third among the entries actually used, receive no recommendations at all. The assistants recommend the press when asked whom to trust, but build their answers from ministries, wire agencies and an encyclopaedia.

Deutsche Welle is Germany’s international broadcaster and is aimed mainly at audiences abroad. Even so, it is recommended slightly more often in German (5.1%) than in English (4.7%). The assistants treat it as a general German news source.

Regional newspapers score well, largely because of the prompt design. Titles such as Kölner Stadt-Anzeiger, Hamburger Abendblatt, Leipziger Volkszeitung and Münchner Merkur together receive 13.1% of recommendations, but 96.9% of these come from one question asking for local sources outside Berlin. That question alone produced 39.7% of all recommendations. Several local tabloids, including Abendzeitung München, Express, tz and Hamburger Morgenpost, were recommended as local sources and cautioned once each in other answers.

Recommendations are widely spread: the leading title has 4.9% and the top five together 17.9%, with the remainder shared among 314 other names. The Local Germany, an English-language site for foreign residents, is recommended by nine systems. Fact-checkers also feature: Correctiv is recommended by eight systems, alongside ARD-faktenfinder, dpa-Faktencheck and Mimikama.

The outlets assistants vouch for

Recommendations split by the language of the request

What each assistant recommends

Bars in the cards show the titles most often recommended within each assistant’s own set of recommendations, so shares are internal to that system.

What the machines say to avoid

Warnings that point to Moscow and the fringe

All 10 assistants issue both cautions and disinformation flags. The flags go to Russian state media, impersonation campaigns and far-right alternative outlets; no mainstream national newsroom receives one.

Cautions make up 10.0% of all judgements about sources and disinformation flags 6.9%. Each of the 10 systems issued both. The two lists overlap in places but differ in composition.

Most disinformation flags target operations rather than individual outlets. Doppelgänger-style sites that imitate established German media account for 15.7% of all flags and are named by eight systems. RT DE and its mirror domains take 13.3%, Sputnik 7.2%, and the Storm-1516 and Matryoshka networks a further 8.4% between them. The Federal Government formally attributed the Storm-1516 campaign to Russia in December 2025, and European authorities and independent researchers have linked Doppelgänger and Matryoshka to Russian actors. Foreign state media and impersonation campaigns together account for 54.2% of all disinformation flags.

The second group consists of German-facing alternative and fringe outlets: Compact (6.0% of flags), PI-News, Epoch Times, AUF1, Junge Freiheit and the conspiracy channels KenFM and Apolut. Platforms, mainly Telegram, TikTok and X, are flagged as channels through which such material spreads rather than as publishers. DeepSeek also flags Grokipedia, the AI-generated encyclopaedia launched by xAI, citing a warning from the German Journalists’ Association.

Cautions follow a different pattern. Bild alone accounts for 17.5% of all cautions and is cautioned by all 10 systems (see Finding 08). Platforms account for another 21.7%, flagged for virality and unverified content. Most of the remaining cautions concern right-wing alternative titles, including Compact, Tichys Einblick, Junge Freiheit, AUF1, NIUS, Apollo News and Reitschuster, as well as the generally left-associated NachDenkSeiten. The satirical site Der Postillon is cautioned once.

The warnings fall almost entirely outside the mainstream press. Across all cautions and flags, the only public broadcaster or major national daily cautioned apart from Bild is ZDF, in a single generic caution from Mistral Vibe. The daily nd (Neues Deutschland) receives a single caution, as do a few local tabloids that were otherwise recommended. The named outlets are overwhelmingly on the right or in the conspiracy milieu; NachDenkSeiten and nd are the only left-associated titles. Whether this reflects the German media landscape or the way the systems were trained and designed cannot be determined from this data.

Sources the assistants say to treat with care

Sources flagged as disinformation-related

Cautions split by the language of the request

Cautions and disinformation flags are counted separately and each chart is scaled to its own maximum. Variants of the same operation, such as RT DE’s mirror domains or the various descriptions of Doppelgänger-style imitation sites, are pooled. Warnings are a small share of all source mentions. Appearing on either list records what an assistant said; it is not MJRC’s assessment of the outlet.

What accompanies the judgement

Linked facts, bare verdicts

Nine in 10 ordinary citations come with a link. Judgements about sources come with a linked reference far less often, and most links attached to a recommendation lead to the recommended outlet’s own website.

Ordinary citations are almost always linked. 90.4% of the sources used to build answers had a usable web address. Excluding ChatGPT, whose exported answers lost most article links, the figure is 96.4%. Only 2.0% of coded runs showed no links at all, three-quarters of them from DeepSeek and the rest from Meta AI.

Judgements about sources are a different matter. 57.2% of recorded judgements came with a citation or link. Recommendations are much more likely to arrive with a citation or link than either kind of warning: recommendations did so 64.7% of the time. Cautions are least likely to do so, at 15.8%. Disinformation flags are somewhat better documented, at 27.7%, but most still arrive without a linked reference.

A link is not the same as evidence. In 70.1% of the linked recommendations, the link leads to the recommended outlet’s own website, and in 47.2% it is simply a homepage. A recommendation of Deutsche Welle with a link to dw.com shows where to find the broadcaster, not why it can be trusted. Warnings follow the same pattern: 54.8% of linked cautions and flags point to the outlet being warned about. Independent assessments, from bodies such as the Federal Agency for Civic Education, NewsGuard or the German Journalists’ Association, appear in only a handful of cases.

The systems differ widely. Mistral Vibe attached a citation or link to 97.5% of its judgements and Copilot to 90.2%. Google AI Overviews did so for 18.8%, DeepSeek for 32.5% and Claude for 35.9%. ChatGPT cannot be compared, because the export used for collection did not preserve its source-card links.

How often a judgement came with a citation or link

Share of each assistant’s judgements with a citation or link

The sources linked alongside those judgements

The second chart shows the share of each assistant’s own recommendations, cautions and flags that came with a citation or link; ChatGPT is omitted because its source-card links were not preserved in the export. The third chart gives each outlet’s share of all links attached to judgements; most belong to the outlets being recommended. Systems that made few judgements move more easily between rounds.

One title, one verdict

The paper no assistant vouches for

Bild is Germany’s best-selling daily newspaper. All 10 assistants caution against it, none recommends it as a national news source, and it accounts for 0.15% of their citations.

Bild is the only source cautioned by every assistant in the study. It is also the most-cautioned source overall, with 17.5% of all cautions, three times the next entry. Some cautions also cover its Munich and Leipzig editions. No assistant recommends Bild as a national source. Its only recommendations come from Meta AI, concern its Hamburg and Leipzig local editions, and appear in answer to the question about local news outside Berlin. No assistant flags Bild as disinformation; Copilot states explicitly that it is a real news outlet that requires extra verification.

The reasons given are consistent across systems and languages: tabloid style, sensationalism, selective framing and, in several answers, Bild’s position at the bottom of the trust rankings in the survey data the assistants cite. Gemini also mentions reprimands from the German Press Council. Bild accounts for 21.7% of cautions in English answers and 13.3% in German ones.

Bild hardly appears among the sources the assistants actually use. It accounts for 0.15% of them, two-thirds in answers about sport. A user asking an assistant about Germany will rarely be sent to Bild; a user asking about Bild will be told to treat it with caution.

This contrasts sharply with Bild’s audience. According to its publisher, Bild sold about 980,000 copies a day in mid-2025 and reached around six million readers. In the AI layer, the German daily with the largest readership appears almost only as a warning. The assistants’ treatment of Bild follows the trust surveys they cite and the views of media critics, and departs sharply from what Germans actually read.

Share of each assistant’s cautions aimed at Bild

Shares fall within each assistant’s own set of cautions; systems that issued few cautions, such as Meta AI and Google AI Overviews, show high shares on a small base. Circulation and readership figures are Bild’s own, from its publisher’s 2025 media data (IVW Q2/2025; ma 2025). Bild’s place on this page reports what the assistants said and is not MJRC’s assessment of the paper.

A single source, examined

deutschland.de, the official explainer the assistants keep reaching for

A public-diplomacy portal produced in cooperation with the Federal Foreign Office is the joint seventh most-cited source in the study. The assistants use it as routine background, particularly on media and information questions, and recommend it to readers abroad.

Most of the sources near the top of these findings are familiar: a wire agency, a ministry, an encyclopaedia, a public broadcaster. deutschland.de is different. It is a multilingual portal that presents Germany to audiences abroad. It is published by Fazit Communication, a Frankfurt company owned by the Frankfurter Allgemeine Zeitung group, in cooperation with the Federal Foreign Office.

It is the joint seventh most-used source in the study, level with Destatis at 1.5% of all citations and ahead of the Federal Government’s website and ZDF. Nine of the 10 assistants cite it; only Mistral Vibe does not. Copilot and Claude use it most.

Where it is used

Its use is concentrated on media and information questions, which account for 43% of its citations. Asked how to follow Germany from abroad, the assistants cite it for 9.5% of their sources, second only to Reuters. Asked which German sources to treat with care, they cite it for 4.5%, level with Tagesschau and behind only Correctiv. The pages they link to include an overview of press freedom and the media in Germany, an article on the federal government’s response to fake news, a briefing on disinformation and cyberattacks, and a summary of the coalition agreement.

How it is described

It is recommended by seven of the 10 systems, almost always to readers abroad, and never cautioned. In over 90% of its recommendations, the answer identifies it as official, government-supported or run in cooperation with the Foreign Office. DeepSeek describes it, inaccurately, as the Foreign Office’s own portal. The assistants therefore recognise the portal’s official character. They nonetheless present it next to Deutsche Welle and The Local Germany as a standard way to follow German news.

Why it matters

Recommending an official portal for background on Germany is reasonable in itself. What matters is how often it is used, and for what. When the assistants explain press freedom, disinformation or the coalition agreement, part of their account comes from a public-diplomacy portal produced in cooperation with the Foreign Office, cited alongside independent journalism as a routine explanatory source. The same pattern appears in Finding 01, where official and government-linked sources weigh about as much as the country’s newsrooms combined; deutschland.de is the most visible single case, and the entry that determines which way that balance tips.

The portal

Publisher

Fazit Communication GmbH, Frankfurt am Main, wholly owned by Frankfurter Allgemeine Zeitung GmbH

Partner

Produced in cooperation with the Federal Foreign Office in Berlin

Audience

Readers abroad, in German, English and a range of other languages including Spanish, French, Russian and Chinese

Where it sits in this data

Joint seventh most-cited source; recommended by seven assistants and identified as official or government-linked in over 90% of those recommendations; never cautioned or flagged

Details of the portal are drawn from its own publisher and ownership information. Its place in these findings reports how the assistants used it and is not MJRC’s assessment of deutschland.de.