If you ask an AI system a question from the Netherlands, does it mainly use Dutch sources? Or do you get an answer based largely on international websites?
That is what I wanted to investigate. I also looked at whether asking the question in Dutch or English makes a difference. After all, people in the Netherlands regularly use English when looking for information online or through AI. The Netherlands ranks first of 123 countries in the EF English Proficiency Index.
I tested this by asking four AI platforms 120 questions in both Dutch and English: ChatGPT, Claude, Gemini and Google AI Mode. Where possible, I set the location to Amsterdam. In total, I analysed 1,680 responses containing 4,346 citations.
For questions in Dutch, 91% of the cited sources are Dutch. For the same questions in English, that figure is 21%.
The differences between platforms are substantial. ChatGPT regularly cites Dutch sources for English questions. Claude and Gemini rarely do. For example, when asked in English about switching health insurers, Claude gave an answer about the US healthcare system, even though its location was set to Amsterdam.
On this page, I explain what I investigated, the differences in the results and what I take from them as an SEO consultant.
Key findings
Here, “Dutch” means more than a .nl domain. It also includes Dutch companies using other domains and Dutch versions of international websites. I explain this definition below.
Why I wanted to investigate this
When developing a strategy for GEO and AEO, I want to know which sources AI systems actually cite. It makes a considerable difference whether a platform compares Dutch providers for a question asked from the Netherlands or mainly uses information from the United States and the United Kingdom.
This study relates to Aleyda Solis’s research into local and international AI citations. My main question was to what extent AI systems cite Dutch sources when a question is asked from the Netherlands. Do Dutch sources dominate, or do international sources remain more prominent?
Within that question, I also investigated the role of language. Do the platforms use the same types of sources when you search in English from the Netherlands? Or does the answer also shift towards a different market?
How did I conduct the study?
To measure which sources the platforms cite for questions asked from the Netherlands, I set the location to Amsterdam wherever possible. I then asked 120 questions across ten topics. Each question had a Dutch and an English version, allowing me to investigate the difference between the two languages within this setup. Each topic included four informational questions, four recommendation questions and four questions about making a decision.
Examples of the Dutch questions, with translations:
- “Hoe werkt het eigen risico bij de zorgverzekering?” — “How does the health insurance deductible work?”
- “Welk merk e-bike is het beste?” — “Which e-bike brand is best?”
- “Moet ik nu een vast energiecontract afsluiten?” — “Should I take out a fixed-rate energy contract now?”
The questions did not explicitly mention the Netherlands. The English versions were natural translations of the Dutch questions.
I also distinguished between topics where the Dutch market or regulations play a clear role, such as health insurance and taxes, and topics with many international similarities, such as software and consumer electronics.
| Element | Study design |
|---|---|
| Questions | 120, each in Dutch and English |
| Topics | 10, with 12 questions per topic |
| Platforms | ChatGPT, Claude, Gemini and Google AI Mode |
| Measurement rounds | Two for ChatGPT, Claude and Gemini; one for AI Mode |
| Responses | 1,680 |
| Citations | 4,346 |
| Period | October 2026 |
| Location setting | Amsterdam, except for Gemini, where I could not set a location |
For ChatGPT, Claude and Gemini, I used the API with a search tool. I collected Google AI Mode results through SerpApi.
This distinction matters: these measurements use these specific settings. The results do not necessarily match what you see in the consumer apps. And for Gemini, I could not specify Amsterdam as the location.
When do I count a source as Dutch?
Looking only at .nl domains does not give a complete picture. A Dutch company may use a .com domain. An international website may also offer a separate Dutch version. I therefore used two definitions:
| Definition | What counts? | Dutch questions | English questions |
|---|---|---|---|
| Broad | .nl domains, Dutch companies using other domains, Dutch-language pages and Dutch sections of international websites | 91% | 21% |
| Strict | Only .nl domains | 79% | 18% |
Throughout this article, I use the broad definition unless stated otherwise. A Dutch source is therefore not necessarily a Dutch-language page. An English page published by a Dutch company can also qualify, for example.
The strict definition is comparable to the measure Aleyda Solis used. With it, 79% of citations for Dutch questions come from .nl domains. She found 24% to 41% for the UK, Spain and Brazil, though without separating languages.
The percentages are calculated per citation, across the measurement rounds combined. A source that is cited more often therefore counts more often. The overall figure is not a simple average of the four platform percentages.
1The language of the question changes which sources AI cites
Dutch questions: 91% Dutch sources. English questions: 21%.
Share of Dutch sources in the citations, per platform
Source: Dutch AI Citation Study 2026 · 4,346 citations in 1,680 responses · broad definition
For Dutch questions, all four platforms mainly use Dutch sources. The share ranges from 86% for ChatGPT to 97% for Google AI Mode.
For English questions, the picture changes. ChatGPT continues to cite Dutch sources regularly. Google AI Mode does too, but less often. For Claude and Gemini, Dutch sources almost disappear from the citations.
For Claude, this happens despite my having specified Amsterdam as the location. In this measurement, that location setting is therefore not enough to produce Dutch sources for English questions.
My main observation is that you cannot assume a location setting will result in local sources. The language of the question plays a major role in these measurements. Understanding exactly how language and location interact requires further research across multiple locations.
2An English question can produce an answer for a different market
One of the clearest examples came from the health insurance questions. I asked Claude the same question in Dutch and in English, with Amsterdam set as the location.
Same question, same location: two different health insurance systems
Claude, first measurement round
Asked in Dutch
“Wanneer kan ik overstappen van zorgverzekeraar?”
Claude searched for
overstappen zorgverzekering 2027 opzeggen 31 december
Sources cited (7)
- advies.alpina.nl
- vgz.nl
- overstappen.nl
- consumentenbond.nl
- cz.nl
- juridischloket.nl
- unive.nl
Answer about the Dutch health insurance system
Asked in English
“When can I switch health insurers?”
Claude searched for
open enrollment 2027 marketplace dates healthcare.gov
Sources cited (4)
- healthinsurancenetwork.org
- textdrip.com
- benefitchamps.com
- healthmarkets.com
Answer about the US healthcare system
Source: Dutch AI Citation Study 2026 · question T01-04, location set to Amsterdam
The question concerned the same topic, but the answer concerned a different health insurance system.
I also saw this difference in the recorded search queries. Claude never includes the Netherlands when searching for English questions. ChatGPT does so in 55% of its queries. Sometimes ChatGPT also used Dutch words, as in:
Netherlands best health insurance 2026 zorgverzekering comparison Consumentenbond
How much do the sources overlap?
I also compared the citations for the Dutch and English versions of the same question. For this comparison, I only included question pairs where both responses contained citations.
| Platform | Question pairs with citations in both | Same domain | Same page |
|---|---|---|---|
| ChatGPT | 148 | 42% | 24% |
| Google AI Mode | 69 | 15% | 8% |
| Claude | 123 | 9% | 8% |
| Gemini | 53 | 4% | 1% |
The percentages show what share of citations in the English response also appear in the Dutch response. For ChatGPT, there is still a reasonable amount of overlap between domains. For Claude and Gemini, switching languages produces an almost entirely different set of sources.
3The topic makes a difference too
Not every question requires Dutch information to the same extent. Dutch health insurance operates under different conditions from US health insurance. For software or electronics, the same products may be relevant in several countries. Even within these topics, however, the share of Dutch sources differs substantially between the two languages.
English questions: 31% Dutch sources for Dutch-market topics, 9% for universal topics
Share of Dutch sources in the citations, all platforms combined
Topics with a Dutch market context
Topics with many international similarities
Source: Dutch AI Citation Study 2026 · 4,346 citations · broad definition
This is particularly noticeable for software and online courses. Dutch sources rarely appear, or do not appear at all, in citations for English questions. That does not automatically mean the answers are unhelpful. It does show that Dutch sources are much less visible for these questions.
4The recommended brands change too
A citation tells you which source a platform refers to. For businesses, it is also useful to know which brands or providers the answer actually recommends. For the 40 recommendation questions, I therefore also analysed which brands, companies and products were recommended, and what share of them were Dutch.
Dutch brands: 46% of recommendations for Dutch questions, 9% for English questions
Share of Dutch brands among the recommended brands, 40 recommendation questions
Source: Dutch AI Citation Study 2026 · 4,667 recommended brands in 560 responses
For Dutch questions, almost half of the recommended brands are Dutch. For English questions, fewer than one in ten are. Two examples from Claude’s responses:
| Topic | Recommended for the Dutch question | Recommended for the English question |
|---|---|---|
| Cheapest energy supplier | Eneco, UnitedConsumers, Vandebron, Delta, Innova Energie | Octopus Energy, E.ON Next, OVO |
| Best accounting software for self-employed professionals | e-Boekhouden.nl, Moneybird, MoneyMonk, Jortt, Rompslomp | Zoho Books, QuickBooks, FreshBooks, Wave, Xero |
The energy supplier example makes it particularly clear why this matters. A recommendation for the British energy market does not help someone looking for a contract for a home in the Netherlands.
Of course, recommending a foreign brand is not inherently wrong. International software may be perfectly suitable for Dutch customers. Alongside a brand’s origin, you therefore need to assess whether the recommendation fits the market and the question.
5Where does the difference for English questions begin?
I wanted to know whether AI finds Dutch sources but then does not cite them, or whether those sources barely appear during the search process in the first place.
Sources found versus sources cited
For ChatGPT, Claude and Google AI Mode, I compared the share of Dutch sources in the available search results with their share of the final citations.
For English questions, the difference starts with the search query
English questions, step by step
Source: Dutch AI Citation Study 2026 · ChatGPT 186 and Claude 132 recorded search queries
For Claude, the share of Dutch sources is already very low in the retrieved results. Their share of citations is almost the same. For ChatGPT and Google AI Mode, Dutch sources actually account for a higher share of citations than of the retrieved results.
These figures suggest that the search queries play an important role. Dutch sources barely appear in Claude’s recorded search results. There is no indication here that Claude first found many Dutch sources and then ignored most of them.
How many Dutch pages have an English version?
I also investigated whether the Dutch pages cited for Dutch questions were available in English. Of the 1,402 pages examined, I could check 1,210. The remaining 192 blocked the check.
| Check | Share of pages that could be checked |
|---|---|
| English equivalent of the page found | 13% |
| English section found on the website | 24% |
| English version found through hreflang | 8% |
An English section on a website does not necessarily mean that the cited information is also available in English.
The most cited Dutch sources rarely have an English version of the cited pages
Citations for Dutch questions, and what the check for an English version found
Source: Dutch AI Citation Study 2026 · check of hreflang and language links, October 2026
This check has limitations. An English version may exist without the script finding it, for example on a separate domain without a clear language link. I therefore interpret these figures as English versions that were demonstrably found, rather than a complete inventory of all English content.
This part of the study also does not prove that missing translations cause the low citation percentages. To establish that, I would need to test what happens to citations after an English version is published.
AI can use Dutch sources for an English answer
An English question does not lead exclusively to English sources. For ChatGPT, 50% of the pages cited for English questions are in Dutch. For Google AI Mode, that figure is 13%; for Claude, 1%; and for Gemini, 0%.
ChatGPT in particular therefore regularly uses Dutch information to produce an English answer. In those cases, an English translation is not a requirement for being cited.
6What happens when you explicitly mention the Netherlands?
To test what explicit location context changes, I asked Claude and Gemini all 120 English questions again. This time, I added one sentence at the beginning:
I live in the Netherlands.
The rest of the question remained unchanged.
One sentence: Claude goes from 1.5% to 81% Dutch sources
First measurement round, 120 English questions per variant
Dutch sources in the citations
Dutch brands in the recommendations
Source: Dutch AI Citation Study 2026 · Claude and Gemini, 240 additional responses
For this comparison, I use the first round of the original measurement. Some percentages therefore differ slightly from the earlier charts, which combine both rounds.
The additional sentence changes the results substantially. Claude searched in Dutch again for the health insurance question. And for the energy question, Eneco and Vandebron returned instead of Octopus Energy and OVO.
Gemini also uses the search tool more often: for 72% of questions, compared with 37% in the first English measurement round without the extra sentence.
The cited pages do not have to be in English for this to happen. In the experiment, 72% of the pages Claude cites are in Dutch. For Gemini, the figure is 46%.
For me, this experiment mainly shows how much explicit market context can change. These platforms can find and use Dutch sources, but rarely do so for the original English questions.
7Which websites are cited most often?
The Consumentenbond is cited 184 times for Dutch questions, more than any source for English questions
Ten most cited domains, by language of the question
Dutch questions
- consumentenbond.nl
- rijksoverheid.nl
- anwb.nl
- zorgwijzer.nl
- belastingdienst.nl
- independer.nl
- overstappen.nl
- rabobank.nl
- kvk.nl
- tweakers.net
English questions
- youtube.com
- consumentenbond.nl
- reddit.com
- rijksoverheid.nl
- belastingdienst.nl
- rtings.com
- techradar.com
- forbes.com
- business.gov.nl
- facebook.com
Source: Dutch AI Citation Study 2026 · number of citations, both measurement rounds
The Consumentenbond is cited most often for Dutch questions: 184 times. That is more than twice as often as the runner-up, Rijksoverheid.nl. For English questions, YouTube and Reddit rank higher. International review and technology sites such as RTINGS and TechRadar also appear in the top ten.
The breakdown by website type shows the same difference:
Dutch questions lean on comparison sites and consumer organisations; English questions on blogs and review sites
Share of citations per website type. Darker bars: at least 4 points more than in the other language
Source: Dutch AI Citation Study 2026 · percentages rounded, so the categories shown do not necessarily add up to 100%
ChatGPT also stands out for its share of institutional sources. For Dutch questions, government bodies, regulators, public institutions and universities account for 38% of its Dutch citations. For the other platforms, the figure ranges from 5% to 11%.
These figures show which sources are frequently cited within my question set. They do not measure trust or provide a general ranking of the websites AI considers most important. The topics selected influence which websites appear most prominently.
8Searching, citing and repeatability are different things
A platform does not use its search tool for every question. And a recorded search does not automatically mean the answer contains citations.
Gemini answers two thirds of the questions without searching
Share of responses per outcome
Source: Dutch AI Citation Study 2026 · 240 responses per platform and language (AI Mode: 120)
Gemini uses the search tool for around a third of questions. Two-thirds of its responses contain no citations.
This also limits what I can conclude from the study. Without citations, I cannot establish which websites contributed to an answer. A brand mention without a source link is also different from a measurable citation.
Do you get the same sources for the same question?
Citations also change between measurement rounds. I examined what share of the domains in round two also appeared in round one.
| Platform | Dutch question | English question |
|---|---|---|
| Claude | 65% | 76% |
| Gemini | 62% | 41% |
| ChatGPT | 55% | 53% |
This was not measured for Google AI Mode, as I only conducted one round there.
This matters when tracking AI visibility. A single answer is a snapshot. To assess whether your visibility is changing, I would measure the same questions repeatedly and look at the pattern.
What do I take from this as an SEO consultant?
Dutch content remains important for the Dutch market
For Dutch questions, 91% of citations come from Dutch sources. Depending on the platform, 84% to 97% of the cited pages are also in Dutch.
If you want to reach Dutch customers, I would therefore start with relevant Dutch-language pages. This also applies to international companies. In this measurement, Qonto, Shopify, Udemy and Raisin, for example, are cited through Dutch-language content. The results give me no reason to rely exclusively on English content for a Dutch audience.
English content can be useful, with clear Dutch context
An English page can be relevant to people searching in English for information about the Dutch market. Think of software for Dutch business owners or explanations of Dutch regulations.
I would tailor the content of such a page to that audience. Which providers are available here? Which conditions apply in the Netherlands? And what local information does the reader need to make a decision?
That is my practical interpretation of the results. The study does not show how many additional citations such a page would generate. It does show that English questions without Dutch context bring a very different market into view on some platforms.
Look at sources beyond your own website
Comparison websites and consumer organisations together account for 34% of citations for Dutch questions. Looking only at your own website therefore gives you a limited view of the sources AI cites.
For each topic, I would investigate which websites appear regularly and how they describe your business. A relevant mention, comparison or review can be a place where potential customers find information about your business, including through AI.
Measure each language and platform separately
Strong visibility in Dutch responses tells you little about your visibility for the same questions in English. You also cannot simply apply ChatGPT’s results to Claude or Gemini.
For a business with a multilingual audience, I would therefore track which sources are cited and which providers are recommended for each language and platform. Repeat those measurements so you can distinguish occasional differences from a recurring pattern.
International & multilingual SEO
Want to know how well your website serves the countries and languages your customers search in?
Methodology and limitations
Questions and topics
The question set consisted of 120 questions across ten topics: health insurance, mortgages and housing, energy contracts, bicycles and e-bikes, self-employment and taxes, B2B software and SaaS, consumer electronics, fashion and online shopping, travel, and online courses and education.
Each topic included four informational questions, four recommendation questions and four decision questions. Each question was asked in Dutch and natural English, without explicitly naming a country.
This selection is not a representative sample of all questions people in the Netherlands ask AI. The results apply to the questions, platforms, settings and measurement period examined.
Platforms and settings
| Platform | Measurement method and settings |
|---|---|
| ChatGPT | OpenAI Responses API, GPT-5.5 (gpt-5.5-2026-04-23) with search tool, reasoning effort “low”, location Amsterdam |
| Claude | Anthropic Messages API, Claude Opus 5.5 with search tool, location Amsterdam |
| Gemini | Gemini API, Gemini 3.6 Flash (requested as gemini-3.5-flash) with Google Search grounding; no configurable location |
| Google AI Mode | SerpApi, Google Netherlands, location Amsterdam, interface language Dutch or English |
For Google AI Mode, the interface language changed alongside the language of the question. Gemini had no location setting. Conditions were therefore not identical across all platforms.
API results can differ from responses in consumer apps. Account settings, memory, available features and location handling may differ, among other things.
Measurement rounds and citations
ChatGPT, Claude and Gemini answered each question twice per language. Google AI Mode answered each question once. This produced 1,680 responses and 4,346 citations, collected in October 2026.
In this study, a citation is a source reference in the response output: a source link in the text for ChatGPT and Claude, a supporting grounding source for Gemini, and a source linked from the answer for AI Mode. These forms of source attribution are not entirely equivalent.
Unless otherwise stated, percentages are weighted by citation and combined across rounds. They therefore describe the distribution of citations, rather than the share of users or the share of responses containing a Dutch source.
Classification and manual review
Claude Haiku 5.5 was used to classify domains, page language and recommended brands. This included the organisation’s country, the website type and whether it was Dutch.
I manually checked the 100 most frequently cited domains and the 120 most frequently recommended brands. Inconsistent brand classifications were standardised.
International websites such as Reddit and Salesforce were not counted as Dutch in their entirety. Only Dutch-language pages and Dutch sections, such as /nl/ or a Dutch subdomain, could qualify. Automated classification may still contain errors outside the manually reviewed selection.
Checking for English versions
A script checked 1,402 Dutch pages cited for Dutch questions. Among other things, it looked for English hreflang references and language-switch links pointing to an English page.
Of these, 1,210 pages could be checked; 192 blocked the check. An English equivalent was found for 13% of the pages that could be checked. English content on a separate domain or without a recognisable language reference may have been missed. Blocked pages may also differ from those that could be checked.
Experiment with explicit location context
The 120 English questions were asked again in one additional round on Claude and Gemini, preceded by “I live in the Netherlands.” This produced 240 additional responses and 1,335 additional citations. These are separate from the 1,680 responses and 4,346 citations in the main measurement.
For the comparison with the experiment, I used the first original measurement round. The experiment tests the effect of this addition within the chosen setup; it does not compare different physical user locations.
Download the full dataset
Questions, responses, citations and classifications · 4 CSV files · 2.4 MB
Etienne Gozems is an international SEO consultant. He helps businesses improve their visibility in Google and AI search engines, focusing on the countries and languages their customers search in.