We analyzed 96,785 sources cited by ChatGPT, Gemini, Google AI Overviews, and Perplexity against the same 150 questions, asked in Spanish from Spain and in English from the United States. The conclusion: the famous rankings of "sources most cited by AI" describe the world in English, not yours. Only 12% of the domains appear in both languages.
When you read a study that says "these are the 10 sources that AI cites the most", there is one question that almost nobody asks: does it cite everyone, or only those who ask in English from the United States?
The two benchmark studies in the industry, from Ahrefs and Semrush, analyzed millions of citations. Neither of them states in what language they asked or from where. And decisions about content in Spain and Latin America are being made every day based on those rankings.
So we measured it. We wrote 150 identical questions in intent, written natively in Spanish and in English, and ran them against the 4 engines with controlled localization: Spain for Spanish, the United States for English. We then classified each of the 96,785 cited sources.
AI does not have a top 10 sources. It has a top 10 by language, and the Spanish one looks almost nothing like the one published by American studies.
Design matters as much as the result, because that is exactly what is missing in the studies that circulate.
The hypotheses, metrics, and the 300 prompts were frozen before launching the first query. The complete dataset is available at the end of the article.
This is what the AI cites when you ask it in Spanish and when you ask it in English. The percentage indicates how many responses each domain appears in.
% of responses where each domain appears. Scaled bars within each chart.
Take a moment to look at them. The English top consists of American editorial giants that appear in all studies. The Spanish top consists of financial communities (Finect, Rankia), a comparator (HelpMyCash), a consumer organization (OCU), press (El País), and a local tech outlet (Xataka).
It's not that AI cites "less" in Spanish. It cites something else.
Of the top 10, only 3 coincide: YouTube, Reddit, and TechRadar. And even those 3 weigh differently depending on the language you ask in:
A fact to calibrate all of the above: we repeat each measurement several times and the rankings of each language remained stable across runs (8 or 9 domains from the top 10 repeat). The difference between languages is twice as large as the natural noise of the engines themselves. This is not chance.
We classify the most cited domains by source type. The distribution tells a story that no one is telling:
% on the citations from the main domains classified by type. Bars to scale within the chart.
In English, the dominant source type is specialized editorial (Forbes, Healthline, TechRadar, Tom's Guide). There is a giant review publishing industry that AI uses as a trust layer.
In Spanish, that editorial layer is weak, and the gap is filled by the brands' own content. HubSpot's blog in Spanish, Brevo's pages, banks, insurers. AI cites the brand directly because it finds no one else talking about the topic with authority.
In Spanish, your own content can be the source. In English, you compete against Forbes for the citation. In Spanish, many times the citation is vacant.
The generalist press is the other contrast: in Spanish it weighs 7 times more (El País, Cinco Días, or El Mundo appear as authority on consumer and finance topics). In English, the generalist press is almost irrelevant compared to niche editorial.
How many domains from the top 10 in Spanish match the top 10 in English? It depends heavily on the topic:
Top 10 domains in Spanish that are also in the top 10 in English, by category.
The pattern makes sense:
Health converges because major American authorities invested in Spanish: Mayo Clinic, MedlinePlus, and the US health institutes have official versions in Spanish, and AI uses them. In fact, in Spanish, half of health citations go to institutional and government sources.
Finance is another planet. Products, regulation, and communities are local. The place occupied by NerdWallet and Bankrate in English is taken by Rankia, HelpMyCash, and Finect in Spanish. One shared domain out of ten.
| Category | Dominant source type | Weight |
|---|---|---|
| B2B Software | Brand content | 57% |
| Health and wellness | Government and institutions | 50% |
| Digital marketing | Brand content | 44% |
| Personal finance | Comparators and forums | 43% |
| Travel | Brand + institutions | 61% |
| Consumer product | Editorial + retailer | 38% |
Your GEO strategy cannot copy that of the American market or another category. First, look at who has the citation in YOUR topic and in YOUR language.
When you ask in Spanish, does the AI respond with sources in Spanish? Almost always, but not always.
The data replicates what we found in Experiment 1 with a single vertical, now confirmed in six. The asymmetry is revealing: the Spanish-speaking user consumes Anglo content translated on the fly by the model. The Anglo user practically never reads a Spanish source.
And the main responsible party has a name. This is how the percentage of sources written in English within responses in Spanish is distributed, engine by engine:
% of sources written in English when answering questions in Spanish, by engine.
ChatGPT cites pages written in English in 1 out of every 3 responses in Spanish. Perplexity, in 1 out of every 25. If your market uses ChatGPT, you are also competing against Anglo content.
Where English is filtered the most follows the pattern of the previous section: digital marketing (27%) and health (22%) at the forefront, finance (7%) at the end. Where the local ecosystem is strong, the AI stays local. Where it is not, it fills the gap with Anglo content.
A finding we weren't looking for: we recorded when Google shows AI Overview and when it doesn't, in each language, for the same questions.
For the English user, Google almost always responds with AI. For the Spanish user, one in five times it still shows the classic results page. Two readings:
In June 2025, Ahrefs' study crowned Wikipedia as the number 1 source of AI: 16.3% of mentions in ChatGPT. In our July 2026 data:
Did we measure incorrectly? We checked manually, and the explanation is better than the data. We asked ChatGPT "What is inflation?" and "Who invented penicillin?": it answered perfectly, without activating the search and without citing a single source. AI no longer consults Wikipedia because it knows it by heart. Wikipedia trained the models that today left it without citations.
AI citations no longer refer to encyclopedic knowledge (that's already within the model). They refer to what the model DOES NOT know: current prices, comparisons, opinions, local recommendations. That's where the competition to appear takes place.
We did not conduct this experiment to debunk them, but to complement them. Their studies measure on a huge scale and are valuable, but they answer a different question: how AI behaves in English. By controlling the language and localization, several of their conclusions change in scope:
The underlying lesson is not that their numbers are wrong. It is that no ranking of AI sources is universal: it depends on the language, the location, and the week you measure.
Because it retrieves content from the web ecosystem of each language, and those ecosystems are different. In English, there is a huge editorial layer (Forbes, Healthline) that is weak in Spanish. That gap is filled by brands, comparators, general press, and institutional sources. Additionally, localization influences: asking from Spain, the engines prioritize sources relevant to that market.
They are not wrong, they are incomplete. They measure AI behavior in English and present it as universal. Their rankings are valid for those competing in the Anglo market. For the Hispanic market, they describe a reality that is not yours.
Ask the engines what your client would ask, in your language and several times at different moments, and record the sources. That’s exactly what CreceRank does automatically: it monitors your prompts, extracts the citations, and tells you where you need to be.
Because counting isolated citations inflates the results: a response with 20 citations from the same domain is not worth 20 times more than one with just one. The presence rate answers the important question: what percentage of responses does this source appear in?
The weapon in Spanish was measured from Spain, so the specific domains (OCU, El País) are from the Spanish market. But Experiment 1 already showed that the underlying pattern is repeated in Argentina and Uruguay with different local actors. The structural lesson (own ecosystems, brands as sources, universal YouTube) applies to the entire Spanish-speaking world.
Yes. The complete dataset (runs, citations, language classification, and type of source) is available in CSV under CC0 license. Use it, replicate the analysis, and let us know what you find.
We don't ask you to believe us: check the data. We publish the complete dataset under CC0 license (public domain): the 10,800 queries, the 96,785 citations with their original URL, the language classification of the content, and the 150 paired intents. Replicate the analysis, look for errors, or use it for your own study.
4 CSV files + reading guide · CC0 License · If you use it, we appreciate the link.Álvarez, S. and Cardozo, S. (2026). Experiment 4: sources cited by AI in Spanish vs English. CreceRank. 10,800 queries to ChatGPT, Gemini, Google AI Overviews, and Perplexity, July 2026. Available at crecerank.com
This study maps the general landscape. Your specific case (your brand, your competitors, your prompts) is another experiment, and it is the one that matters. CreceRank runs it for you every day.
