AI Information Map: Spain

AI Information Map · Spain

The Spain that AI systems describe

Ten assistants, two languages, one bank of questions. These pages record which Spanish outlets the machines put in front of a reader, which they vouch for, and which they warn against.

Contents

Research on this country is continuing. Most visibility results below are expressed as percentages; where counts appear, they are clearly labelled. Figures are revised as further collection rounds are folded in. The underlying dataset can be requested from MJRC.

AI Information Map · Spain

Behind the numbers

What was asked, of which assistants, how many times, from where, and by what rules every source they produced was catalogued.

The systems and the questions

Ten widely used assistants were placed under identical conditions: Perplexity, ChatGPT, Gemini, Copilot, Claude, DeepSeek, Meta AI, Grok, Mistral Le Chat and Google AI Overviews. Each received a fixed bank of questions about how Spaniards find out what is happening in their country. The bank covers five territories, namely media and information sources, politics and governance, economy and public services, society, rights and environment, and culture, sport and local life. Every question was posed twice, once in English and once in Spanish.

Only the free tiers were examined, the versions a person reaches without paying. Subscription and enterprise tiers may behave differently. What follows therefore describes the experience of an ordinary user rather than a paying one.

Scale, repetition and geography

The 20 prompts published beside these findings are illustrative rather than exhaustive. Each underlying question was reformulated into roughly 11,000 phrasings, which yielded on the order of 220,000 prompt runs and approximately 400,000 archived answers for this batch. Requests were dispatched from 72 separate IP addresses so that the effects of geography and session state could be observed, over a window running from April 2025 to March 2026. A concluding round of validation prompting was carried out in July 2026.

The coding scheme

Each answer was stored together with its prompt, the system that produced it, the language, the date, and whatever links, citations or source cards were on screen. Researchers then lifted out every source the assistant put forward and gave it a role:

• Used or cited, drawn upon to construct the answer
• Recommended, named outright as trustworthy
• Recommended alternative, put forward to widen or counterbalance a reader’s diet
• Cautioned, marked as requiring care
• Disinformation-related, identified as manipulation or falsehood

Sources are gathered under a single brand, so that Wikipedia’s language editions register once, El País’s Spanish and English properties are treated as one title, European Commission sub-sites are pooled, and alternate spellings are reconciled. Prominence is counted per outlet, never per address.

On the disinformation category

How the results are expressed

Since the mapping is continuous, visibility results are expressed as percentages, meaning an outlet’s portion of the visibility available within whichever slice is under discussion, be that the whole corpus, a single assistant, one language, one subject area or one role. Running totals of source mentions are withheld, because they climb with each fresh round of collection. Where counts do appear on these pages, they describe the design of the study rather than its results: the number of assistants, languages and IP addresses, the approximate scale of the run, the count of distinct outlets behind a percentage, and how many assistants took a given position. These are labelled wherever they occur.

The work is not finished. Spain’s database remains open and the figures presented here are refreshed at intervals as new rounds arrive. Shares will move. Read every finding as the state of the evidence today rather than a settled verdict.

Boundaries of the exercise

This batch charts where sources appear and what roles they are given. It passes no judgement on whether an answer was factually right, and it does not sort outlets into public, private, state-held or captured. Ownership analysis belongs to a separate strand of MJRC’s work. You.com, Brave, DuckDuckGo and Poe fell outside this round.

How visibility is distributed

A shallow summit and a deep field

No Spanish outlet commands the machines’ attention. A newspaper, an encyclopaedia, a radio network and a ministry recur above the rest, while four fifths of all visibility scatters across nearly two thousand other names.

The arithmetic of dominance is a reasonable place to start, if only because there is remarkably little of it. El País, the outlet the machines lean on more than any other, supplies barely 4.9% of the sources behind answers about Spain. Add the next two and the three together reach only 10.2%. The remaining 77.6% is parcelled out among 1924 separate outlets. Set beside the winner-takes-most economics of search and social feeds, that is a strikingly flat distribution.

The names themselves tell a second story. Behind the newspaper at the top comes Wikipedia, then the radio network Cadena SER, and then La Moncloa, the prime minister’s own website. Reuters and the European Commission follow close behind. Only one Spanish newspaper outranks an encyclopaedia and a government department. These systems tend to consult the record and the state before they consult the press.

How far down the press sits is the sharper point. RTVE holds around one percent. elDiario.es holds half of that. La Vanguardia barely registers, and El Mundo and ABC, two of the country’s most-read papers, are all but absent from the citation record. The picture is less one of exclusion than of demotion: Spanish journalism remains available to these systems, but it sits behind reference works and official communication in the queue of what an assistant reaches for first.

Both halves of that picture deserve weight. The summit is narrow and institutional, and it warrants scrutiny. The field beneath it is wide and genuinely varied, taking in provincial dailies, statistical agencies and specialist publications that a more concentrated system would never trouble to surface. Spain’s AI information layer is neither captured nor healthy. It is top-light and bottom-heavy at once.

The sources assistants draw on most

Everything outside this table, some 77.6% of all citations, is divided among 1924 further outlets. Those appearing most often include Idealista, Euronews, PSOE (party site), YouTube, BBVA Research, INE (National Statistics Institute), CIS (Sociological Research Centre), Moncloa.com (news site), elDiario.es, Ministry of Inclusion, El Español, BOE (Official State Gazette), Marca, EFE, AFP, El Confidencial, CaixaBank Research, Ministry of Health, La Razón, Viator, Elcano Royal Institute, Spain.info (official tourism), The Local Spain, El Debate, Reddit. A complete listing can be obtained from MJRC.

Every bar shows one outlet’s portion of all sources used to assemble answers, pooled across ten assistants and both languages.

One question, ten answers

Ten assistants, ten reading lists

Run an identical prompt bank through ten systems and each settles into its own habit of sourcing. Some reach for the newsroom, some for the ministry, and citing a great deal turns out to be quite separate from citing widely.

Ask the same thing of ten assistants and ten answers come back, drawn from ten different Spains. There is nothing random about the divergence. It follows from how each system was built: whether it queries the live web, which index sits behind it, and how much authority it grants a ministry over a masthead.

Perplexity and Google AI Overviews, both wired directly into search, reach for working Spanish newsrooms and platforms. Their source lists are the ones a Spanish reader would find familiar. Gemini and ChatGPT pull toward the settled and the official, so that La Moncloa, Wikipedia and the wire agencies shoulder much of the load. Claude ranges more idiosyncratically still, leading with Infobae and the polling institute CIS. Meta AI and Mistral cite more heavily than anyone, scattering links across their answers by the dozen.

The practical upshot is blunt. Which assistant a person opens determines which Spain they are shown. Perplexity leads with El País and RTVE, so a reader arrives by way of the newsroom. Gemini leads with La Moncloa, so the same reader arrives by way of the government. Neither is malfunctioning, but source visibility is plainly not a characteristic of artificial intelligence as such. It belongs to each product individually, and pluralism has to be assessed one product at a time.

What each assistant reaches for

Quantity of citation against variety of citation

How much an assistant cites and how widely it ranges turn out to be separate characteristics. Meta AI is the clearest illustration. It packs roughly twenty-seven links into an average answer, yet those links resolve to only about fifteen distinct outlets, the poorest ratio of any system. It returns to the same small circle again and again. Grok inverts this entirely, offering around five links per answer that lead to some nine separate outlets, and so ranges further on a fraction of the volume.

What follows from this is that an answer thick with links may still be drawn from very few places. Abundance and variety have to be measured apart, since one gives little indication of the other.

Volume and range are indexed against the heaviest citer, Mistral, set to 100. “Outlets per link” divides distinct outlets by total links: above 1.00 an assistant spreads its sourcing, below 1.00 it repeats itself. The ratio flatters assistants that cite very little, so Gemini’s high figure reflects sparse citation rather than breadth. Copilot, Gemini and Perplexity surfaced fewer visible sources, so their shares move more easily. How often an assistant shows no source at all is treated separately in Finding 08.

What the question’s language decides

Switch the language, change the country

The same twenty questions, asked in English and then in Spanish, return two separate source lists. One leans on governments, agencies and wire copy; the other on the newsrooms that cover the country.

The language in which a question is posed turns out to shape the answer’s foundations. Both layers open with the same title, El País, so the shift lies in what follows it. Ask in English and the sources ranged behind that newspaper are largely institutional: La Moncloa, the European Commission, the OECD, alongside wire copy from Reuters and Euronews and analysis from the Reuters Institute. Several of these are Spanish, La Moncloa most obviously so. What unites them is not their nationality but their character. They are governments, agencies and international bodies describing Spain in their own words, rather than newsrooms reporting on it.

Ask in Spanish and journalism moves up the order. El País roughly doubles, from 3.0% to 6.2%. Cadena SER climbs past threefold to 3.7%. RTVE, Diario AS and Infobae enter the leading ten, while the European Commission and the OECD recede. La Moncloa remains present in Spanish, though it slips from second place to fourth and takes a smaller share. The government has not been displaced so much as overtaken by the outlets that cover it.

Two titles make the shift unusually legible. The Olive Press and The Local Spain, both written in English for foreign residents, each account for around 0.6% of English-prompt visibility and for virtually nothing in Spanish. Where a Spanish-speaking reader is handed Cadena SER, an English-speaking one is handed a newspaper for expatriates. The Spanish press does not merely thin out in English; part of its place is taken by publications written about Spaniards rather than by them.

The consequences fall on identifiable people. A correspondent filing from Madrid, a Brussels official preparing a brief, a member of the diaspora reading in English, a prospective resident researching a move: each asks in English, and each is steered toward official accounts and international summaries rather than toward the newsrooms covering the country day to day. No assistant has misled them. They have simply been shown the Spain that institutions describe, while those who ask in Spanish are shown the Spain that its journalists report.

Bars record each outlet’s portion of sources within its own language layer, and are scaled independently in each column.

Subject by subject

Subject matter decides who gets heard

Ask about money or government and a ministry replies. Ask about a festival, a match or a neighbourhood and the Spanish press is finally allowed to speak.

Who speaks for Spain has no fixed answer, since it depends entirely on what is being asked. On the economy or the government, the reply is assembled from La Moncloa, the INE, the BOE, the Bank of Spain, the OECD and the European Commission. On precisely the questions where citizens might most value independent scrutiny, the assistant tends to relay official and supranational reporting, and does so more insistently when the question arrives in English.

On culture, sport and neighbourhood life the picture reverses. Spanish and regional titles come forward, joined by Wikipedia, Instagram, YouTube and tourist boards, for the simple reason that no ministry publishes on these matters. It is in the ordinary, located, everyday subjects that Spanish journalism becomes most visible to a machine, and this is the only terrain where social platforms enter the record in any quantity.

The most self-revealing category is the one about the press itself. Asked to appraise Spanish media, the assistants turn to the Reuters Institute’s Digital News Report more readily than to any Spanish source except El País. The report is a well-regarded piece of comparative research and an entirely reasonable place to begin. What stands out is how much weight rests on it. A single comparative study carries much of the burden when assistants describe Spanish journalism, with the country’s own media scholarship, regulators and press bodies appearing far less often.

Beneath every subject the language split persists. Even within economics or politics, a Spanish-worded question draws in more domestic outlets than an English one. Pluralism is therefore patchy across the terrain: thinnest for hard public-affairs questions asked in English, richest for cultural and local questions asked in Spanish.

Leading sources in each subject, split by language

Politics & governance

Economy & public services

Society, rights & environment

Culture, sport & local life

Media & information sources

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

What the machines say to trust

The reading list versus the working list

Invited to name reliable sources, the assistants recommend broadsheets and the national broadcaster. Left to answer anything else, they reach instead for an encyclopaedia and a government website.

Drawing on a source and vouching for one are different acts, and the assistants perform them differently. Pressed to name outlets a reader can rely on, they abandon the encyclopaedia and the ministry and reach for the press. El País leads, trailed by RTVE, La Vanguardia, El Correo, El Mundo and the regional Diario de Sevilla, with the English-language Olive Press among them.

Placed side by side, the two lists diverge sharply. Wikipedia and La Moncloa, first and fourth among the sources the assistants actually use, are recommended essentially never. El Mundo, which barely appears in the citation record at all, is the sixth most-recommended title in the country. These systems appear to know which outlets a reader ought to consult, while consulting rather different ones themselves.

Endorsement is also spread thinly. The leading title accounts for roughly 6.6% of all recommendations and the leading five for about 18.7%, with the balance dispersed across national, provincial and investigative mastheads. Regional papers, El Correo, Diario de Sevilla and Deia among them, surface here far above their usage share, as though the systems grasp the principle of regional plurality while defaulting in practice to whatever Madrid publishes.

Two reservations. Outlets published in English for foreign residents are commended to English-speaking users at a level no Spanish reader would accept as representative of their press. And because endorsement lists are short and shift with the phrasing of the request, not one title is recommended by every assistant. Guidance on reliability turns out to depend on which product a person happens to open, and in which language.

The outlets assistants vouch for

Everything outside this table, some 60.5% of endorsements, is divided among 190 further outlets. Those appearing most often include sevilla.abc.es, Onda Cero, elDiario.es, ABC, Europa Press, 20minutos, Ara, SUR in English, spainenglish.com, Catalan News, valenciaplaza.com, Euro Weekly News, Time Out, Reuters, elcorreoweb.es, Maldita.es, BBC, Euronews, laSexta, Newtral, The Guardian, COPE, sevillaactualidad.com, bilbao.eus, EFE Verifica. A complete listing can be obtained from MJRC.

Endorsement split by the language of the request

Where the assistants disagree

Titles offered as alternatives

Distinct from outright endorsement, assistants also put certain outlets forward as alternatives, suggested to broaden or offset a reader’s existing diet rather than as first recommendations.

Chips list the titles most often endorsed inside each assistant’s own recommendation set, so shares are internal to that system.

What the machines say to avoid

What the systems tell you to distrust

Combative digital titles, messaging platforms, and the public service broadcaster. Warnings are issued freely, yet the word disinformation is never once used of a Spanish source.

What the assistants did warn about arranges itself into three groups. The first is a body of combative digital titles. OKDiario draws more warnings than anything else, at 6.5% of the total, with Periodista Digital, Mediterráneo Digital, Libertad Digital, El Debate, El Diestro and El Toro TV behind it. These sit, almost without exception, on the political right. Nothing resembling an equivalent grouping on the left appears anywhere in the warned set. That imbalance is what the assistants produced. Whether it reflects Spain’s media terrain or the assistants’ own priors is not something this dataset can adjudicate.

The second group consists of platforms rather than publishers. X, Facebook, WhatsApp, Telegram and TikTok are flagged as conduits, and the accompanying wording concerns forwarding, virality and the absence of checking rather than any editorial failing. What is being warned about here is a mechanism of circulation, and the assistants are careful to distinguish it from the newsrooms whose material passes through it.

The third is the one nobody would predict. RTVE, the public service broadcaster, attracts the second largest share of warnings, 4.5%, and does so at practically the same rate whichever language the question is asked in. The stated grounds concern closeness to government rather than any suggestion that it fabricates. ABC and La Razón on the right, and the commercial channels Telecinco, Antena 3 and Cuatro, meet much the same treatment. More than one of these titles also appears on the list of outlets the assistants recommend.

All of this should be read with some caution. Warnings are far scarcer than endorsements, so the warned set is small and moves easily between collection rounds. The absent disinformation category also merits reflection. It may be that Spain genuinely presents no such case to a machine, or it may be that the models will not level that charge at a Western European democracy. The evidence assembled here cannot separate the two possibilities, and records only that the charge was never made.

Sources, outlets and channels the assistants say to handle carefully

Everything outside this table, some 59.4% of warnings, is divided among 87 further entries. Named sources appearing most often among them include Telecinco, Instagram, El Debate, El Diestro, YouTube, Cuatro, laSexta, COPE, Telemadrid, 3cat.cat, laultimahora.es, aldescubierto.org, El País, El Mundo, Derecha Diario, Canal Red, Hispan Online. A further 7.1% of warnings named no source at all, describing instead a type of source such as “hyper-partisan pages” or “anonymous Telegram channels”. A complete listing can be obtained from MJRC.

Warnings split by the language of the request

Which assistant warns about which sources

Naming a source, or describing a type

A warning can take either of two forms. An assistant may name a specific source, as Claude does when it puts OKDiario at the head of its list, or it may describe a kind of source without identifying any instance of it: “hyper-partisan pages”, “anonymous Telegram channels”, “medios altamente polarizados”. The second form gives a reader far less to work with, since a description cannot be looked up, checked or contested.

Assistants differ considerably in how far they will commit. Claude and Gemini name a source every single time they issue a warning. At the opposite extreme sits Copilot, which never names one at all: every warning it produced about Spanish media described a type of publisher rather than a publisher. Perplexity and DeepSeek lie between, framing roughly two thirds of their warnings as categories. Only named sources are charted elsewhere on this page, since a description cannot meaningfully hold a share of visibility.

The portion of each assistant’s warnings aimed at a described category rather than a named source. A striped bar marks an assistant that always named its target. The warned set includes news outlets alongside platforms and messaging services such as X, Facebook, WhatsApp, Telegram and TikTok, which are flagged as channels of circulation rather than as publishers. Warnings form a small fraction of all source mentions and pick out repeat subjects rather than every troubling source. Appearing here records what an assistant said, and is not MJRC’s judgement of any source.

The proof behind the judgement

Who vouches for the verdict

Praise arrives with a link attached. Warnings frequently arrive with nothing. And such proof as exists tends to issue from the very newsrooms being judged.

Once an assistant has told a reader whom to trust and whom to doubt, a harder question follows. On what authority? Where a verdict came with a supporting link, that link is the assistant’s reasoning made visible. Those links lead, overwhelmingly, back to Spain’s established newsrooms. El País, RTVE, La Vanguardia, El Mundo and Cadena SER supply most of the proof.

The loop closes on itself. When an assistant sets out to establish which Spanish outlets deserve confidence, the evidence it produces is drawn from the outlets in question. A recommendation of El País underwritten by a link to El País verifies nothing. A small number of newsrooms end up furnishing both the answers and the appraisal of who should be answering, and no independent voice interrupts.

The imbalance in that evidence is the finding that matters. Endorsements arrived with a supporting link roughly 91.6% of the time. Warnings did so only 52.9% of the time. An assistant that praises a Spanish outlet nearly always shows its working. An assistant that warns against one withholds it almost half the time.

The asymmetry would be less troubling were it reversed. An unsupported recommendation costs a reader relatively little. A warning that an outlet is partisan or unreliable, delivered with nothing attached, cannot be examined, contested or traced, leaving the reader to accept it or discard it whole. Meanwhile the institutions built precisely to make such judgements auditable, among them the Reuters Institute, Media Freedom Rapid Response, verification desks such as Newtral and Maldita.es, and rulings by press bodies, appear only at the margins. The assistants have learned the language of media criticism without acquiring its machinery.

How often a verdict comes with proof

The portion of each kind of verdict that carried a link a reader could follow. Warnings are the least substantiated.

The sources cited in support of those verdicts

Everything outside this table, some 65.0% of supporting links, is divided among 208 further outlets. Those appearing most often include Deia, Newtral, elDiario.es, ABC, 20minutos, Onda Cero, sevilla.abc.es, Europa Press, Ara, EFE Verifica, Euro Weekly News, OKDiario, SUR in English, spainenglish.com, Catalan News, valenciaplaza.com, Time Out, Reuters, laSexta, BBC, Euronews, COPE, X (Twitter), La Razón, The Guardian. A complete listing can be obtained from MJRC.

How much proof each assistant offers

The portion of each assistant’s own endorsements and warnings that arrived with a supporting link. A striped bar marks a system that supplied none.

Where each assistant looks for authority

Bars give each outlet’s portion of all supporting links attached to source judgements, pooled across every assistant.

The citation without a link

Sources named, not linked

About one reply in seven names Spanish outlets in prose while offering no link to any of them. Two assistants are responsible for nearly all of it.

In 14.0% of all runs, an assistant delivered a complete and confident account of some aspect of Spain while showing the reader not a single source they could open. Roughly one reply in seven.

The shortfall is concentrated rather than general. Gemini and DeepSeek between them account for 87.5% of every instance, with ChatGPT and Google AI Overviews making up the remainder. The other six systems never once withheld their sources: Claude, Copilot, Grok, Meta AI, Mistral and Perplexity each showed a link on every run.

Those replies rarely fall silent about sourcing. They offer instead an attribution that cannot be checked. The assistant writes that El País reported something, that RTVE has covered it, and stops there. No link follows. To the reader, an outlet named without a link and an outlet invented outright look exactly alike. The sentence carries all the authority of sourcing and none of its accountability.

For the outlets themselves the arrangement is worse than obscurity. Their names are spent to lend weight to an answer while no reader is ever sent to them, no claim is ever traced back, and no credit accrues. The citation has become ornamental, borrowed authority with the borrowing concealed.

How often each assistant hides its sources

A striped bar marks zero. Six of the ten assistants always showed their sources.

What the omission looks like

Shares of all source-free replies; figures are rounded and may not total exactly 100%. The prevailing pattern is naming without linking, rather than an absence of the language of sourcing.

A single outlet, examined

The broadcaster the machines cannot place

RTVE is the only Spanish title to sit high on both the recommended list and the cautioned list at the same time. Five assistants recommend it and warn against it in the same breath.

Most names near the top of these findings explain themselves. El País is the paper of record. Wikipedia and La Moncloa are enormous, static and endlessly indexed. OKDiario draws warnings, and few in Spain would raise an eyebrow at that. RTVE is the one outlet these systems cannot seem to settle on.

The public service broadcaster stands second among the titles assistants endorse, taking 3.8% of all recommendations. It stands second among the titles they warn about, taking 4.5% of all cautions. In the entire Spanish dataset no other outlet occupies a high position on both lists at once.

Endorsed and warned about at once

The overlap is the heart of it. Perplexity, Gemini, DeepSeek, Meta AI, Mistral (Le Chat) each recommend RTVE as a source worth trusting and, elsewhere, mark it as a source to approach with care. Which verdict a user receives depends only on how the question was framed. ChatGPT, Copilot, Google AI Overviews and Grok endorse it without qualification. Claude alone warns about it and never recommends it.

Assistants that call RTVE reliable

Assistants that warn about RTVE

Shares fall within each assistant’s own set of endorsements or warnings. DeepSeek is the most emphatic on both counts, devoting 18.2% of its warnings to RTVE while also recommending it.

Why the machine cannot decide

Once the two questions are separated, the contradiction dissolves into something more ordinary. Asked which Spanish sources are dependable, an assistant names the public service broadcaster, much as a media-literacy handbook would: a funded newsroom with national reach, statutory duties and decades of output. Asked which Spanish sources warrant caution, it names the same broadcaster again, this time on the grounds of proximity to government, and never on any allegation that it invents.

Both propositions have serious advocates inside Spain, where the governance of RTVE, the appointment of its board and the independence of its newsroom have been argued over for years. The assistants reproduce both sides of that argument faithfully. What they never do is reconcile them, or signal to the reader that a disputed question is being settled on their behalf. Ask one way and RTVE is trustworthy. Ask the other and it is suspect. The reader is told nothing of the answer they did not receive.

Language offers no way out. RTVE is warned about at essentially the same rate in each (4.5% in English, 4.5% in Spanish), and recommended somewhat more readily in Spanish (4.7%) than in English (2.9%). The ambivalence is built in, not lost in translation.

The outlet

Origins

Television broadcasting from 1956; brought together as Radiotelevisión Española in 1973 and reconstituted as a state corporation in 2006

Standing

Spain’s national public-service broadcaster, whose principal channels carry no commercial advertising

Oversight

Its board is appointed through parliament. The mechanism has been rewritten more than once and remains a matter of political dispute

Reach

National television, the RNE radio network, a substantial Spanish-language digital newsroom, and services aimed abroad

Where it sits in this data

Second most endorsed. Second most warned about. Among the titles most often cited as proof when assistants pass judgement on other outlets

Details of the outlet come from RTVE’s public corporate documentation and the public record. Its presence among warned titles reports what assistants said, not MJRC’s view of the broadcaster. The full source-visibility dataset behind these figures can be requested from MJRC.