Who do AI assistants cite for health questions?

Four findings

1. The assistant cites institutions, not companies or crowds

60% of cited links go to public health bodies such as the NHS, sundhed.dk and gesundheitsinformation.de. Patient and professional organisations take 23%. Pharma company sites received 0 citations, Wikipedia 0 and forums 0, out of 747. For a company, this means its own website is not how it reaches these answers. The sources the assistant trusts are.

2. Sources are local to the market

The same question gets different sources in each language. In Germany, 49% of citations go to domains in the national top-level domain, led by gesundheitsinformation.de (26 of 241). In Denmark, 64% of citations go to domains in the national top-level domain, led by sundhed.dk (85 of 233). In United Kingdom, 71% of citations go to domains in the national top-level domain, led by nhs.uk (93 of 273). A visibility standard set in one market does not transfer to the next. Each market has its own short list of sources that decide what the assistant says.

3. The question type changes the source mix

For diagnosis questions, public health bodies take 73% of citations and research takes 3.2%. For lifestyle questions, public health bodies take 37%, patient and professional organisations 37%, and research 23%. Reporting one blended number per brand hides this. Prompt clusters need their own numbers.

4. Company text reaches answers through regulators

Pharma company sites got 0 citations. But 30 links in 14 answers went to official medicine documents: EMA product information and assessment reports, UK patient leaflets on medicines.org.uk, and product pages on pro.medicin.dk. All of them came from obesity questions. By market: Germany 12, Denmark 15, United Kingdom 3. The product information in these documents is drafted by the company and approved by the regulator. So when a company’s own words reach an answer, they arrive through a regulator’s page, not the company’s site. Those documents are where accuracy has to be right first.

Method

• Questions: 50 consumer health questions on diabetes, weight, heart health and general health, in six intent clusters. Each was written in English and translated to Danish and German. 45 are covered in this release.
• Collection: each question was sent through the vendor API with web search switched on and the user location set to the market. Models: ChatGPT (gpt-5.5). Run date: 2026-09-19.
• Repeats: every question was asked twice per model. For ChatGPT, the two runs shared 44% of their cited domains on average (Jaccard overlap). One run is a sample, not a fixed answer.
• Extraction: each cited URL is reduced to its domain and matched against a rule list of source types. After one review pass, 1.2% of links remain unclassified. They are shown, not dropped.
• Medicine documents: a cited link counts as an official medicine document when its URL matches a short pattern list in the study config (EMA EPAR pages and product information, medicines.org.uk leaflets, pro.medicin.dk product pages). No product names are counted.
• Company mentions: answers were searched for 13 pharma and device company names. None was named in any answer. Product names were not searched.

Limits

• An API answer with web search is close to, but not identical to, what a logged-in user sees in the chat app.
• 32 of 238 answers cited nothing. Those answers still shape what people read, and this study cannot say where their content came from.
• Source types are assigned by domain. A press release hosted on a public body’s domain counts as that public body.
• The translations were written by one person and not reviewed by native-speaking clinicians.
• This is one run on one date. Shares will move as models and search indexes change.

Data

Download citations.csv (747+ rows: question, language, market, model, sample, cited URL, domain, source type) · summary.json · live dashboard · how we measure