The short version
- In an audit of 615 cited sources by Jacques and colleagues, ChatGPT drew 75.7% of its health citations from institutions such as hospitals, government agencies and Wikipedia.
- Commercial health publishers earned 12.4% of citations, and they stated a medical review on 71.1% of their cited pages.
- Of those commercial pages, 86.8% carried schema markup, the machine-readable labels that describe a page to software, and 68.4% ran past 1,500 words.
- Government health pages, by contrast, stated a medical review only 13.4% of the time, and only 31.1% of them were that long.
Which kinds of health sources does ChatGPT cite?
Mostly institutions with built-in authority; commercial publishers take about one citation in eight.
Jacques and colleagues (opens in a new tab) took a random sample of 100 questions from HealthSearchQA, a set of 3,173 consumer health questions that Google Research built from real search suggestions. One researcher entered each question into ChatGPT 5.2 Pro, in a fresh account and a separate chat. Every question asked ChatGPT to include sources with links. All answers were collected on January 11, 2026.
The team then coded every cited page. Of 615 usable sources, 75.7% came from organizations with inherent institutional authority. Medical institutions led, followed by government resources, Wikipedia, professional associations and journals.
The rest split almost evenly. Commercial health information platforms took 12.4% and small professional or practice websites took 11.9%. The commercial group was led by names such as Healthline, WebMD and Medical News Today, with 76 citations spread over 28 organizations.
What do the cited commercial sites have in common?
Three signals appear again and again: a stated medical review, schema markup and long, comprehensive content.
The authors describe these as “compensatory credibility signals”. A commercial publisher has no hospital or government name behind it, so its pages show their vetting openly instead.
| Signal on cited pages | Commercial health platforms | Government resources | Medical institutions |
|---|---|---|---|
| States a medical review | 71.1% | 13.4% | 49.5% |
| Comprehensive content (over 1,500 words) | 68.4% | 31.1% | 31.9% |
The study reports that cited commercial platforms implemented schema markup in 86.8% of cases. Government sources and journals sat below the sample’s typical rate on schema, at 61.3% and 42.9%.
One signal the commercial group did not lead on was freshness. Only 18.4% of cited commercial pages showed a date from 2024 to 2026. Small practice websites took a different route: 61.6% of their cited pages carried a recent date.
How do commercial sites compare with the whole cited set?
They beat the cited set as a whole on review, schema and length, but not on freshness.
Across all 615 cited sources, 70.4% did not state any medical review at all. Schema markup appeared on 74.3% of all cited pages, below the commercial rate. Comprehensive length was common too, but the commercial group still ran above the overall share.
The cited set was also strong on classic web authority. Sources had a median Domain Authority of 89, a 0 to 100 score from the SEO tool Moz that estimates the strength of a site’s links. Commercial platforms that were cited had a median score of 73, well above the small practice sites.
In other words, the commercial publishers that ChatGPT cited were large, established sites that also invested in visible quality markers. A new site that copies the markers would not copy the years of links and reputation behind them.
Does this mean those signals earn the citation?
No. The audit only looked at pages ChatGPT cited, so it cannot say what the signals cause.
The study has no comparison group of commercial health pages that ChatGPT passed over. If uncited commercial pages also state a medical review and use schema, the signals would not explain anything. The authors present their findings as a baseline to monitor over time, not as a ranking formula.
Two more cautions matter. The coding was largely automated: pages were scraped and labeled by an AI model working from the researchers’ codebook, with missing fields checked by hand. And the split between commercial publishers and small practices partly used a Domain Authority cutoff of 45 when the type was unclear.
What do other studies say about schema, length and review?
They suggest schema alone does little, while substance and evidence on the page matter more.
Our study of pages cited by AI Overviews took 486 US searches and compared cited pages with the uncited pages ranking beside them. Once pages on the same search were compared, no schema type held up as a reliable predictor of citation. Position in Google came first: 41.7% of pages ranked 1 to 3 were cited. Our study also notes a controlled test by Ahrefs in which adding schema did not raise citations. Our guide to on-page signals linked to AI citations covers the rest of that evidence.
Length looks more promising, but only with structure. Zhang and colleagues (opens in a new tab), three independent researchers, analyzed ChatGPT, Google and Perplexity citations. On their measure of how much a cited page shaped the answer, the average rose with length, from very short pages to pages longer than 3,000 words. The pages that did best were also better organized and closer to the question, so length worked as a container for evidence, not as a target. Whether reorganizing a page alone helps is covered in content structure and AI citations.
Finally, evidence beats tone. In a 252,000-trial controlled test by Vishwakarma and colleagues (opens in a new tab) at Sprinklr, a software vendor, AI models preferred pages that backed claims with evidence. That is close to what a stated medical review promises a reader.
What should you do about it?
If you publish health or other expert content without institutional backing, show your vetting and depth plainly.
- State who reviewed each page, with their qualifications and the review date, where a qualified reviewer really did the work. A named author on its own is not required; see whether cited pages need author bylines.
- Write pages that answer the full question a patient or buyer has, with sections, definitions, figures and references, rather than short, thin posts.
- Add accurate schema markup as basic hygiene, but do not expect it to earn citations on its own.
- Keep important pages dated and current; the cited commercial pages lagged here, and small practices did not.
- Keep building search ranking and links, since the cited sources were overwhelmingly high-authority sites.
- Track which of your pages ChatGPT actually cites for your core questions, and compare them with pages it skips.
If you want a structured way to run that tracking, see our generative engine optimization service.
What does the research not tell us yet?
The audit is a careful first map, but it is one engine, one day and one subject.
- It covers ChatGPT 5.2 Pro on January 11, 2026, with a prompt that asked for sources. Other assistants and ordinary prompts may cite differently.
- It used 100 of 3,173 questions; the authors say a different sample might change the results.
- It describes cited pages only, with no uncited pages to compare against.
- Health is a special case. Results may not carry over to finance, software or local services.
- No study here changed a real page and measured its citations before and after.
Frequently asked questions
Does a medical review statement help a page get cited by ChatGPT?
It is common on cited commercial health pages, but no study has shown it causes citation. In the ChatGPT audit, 71.1% of cited commercial pages stated a medical review, against 13.4% of cited government pages.
Does schema markup help health sites appear in AI answers?
There is no reliable evidence that schema alone does. Most cited commercial health pages used it, but our AI Overview study found no schema type that held up once pages on the same search were compared.
How long should health content be to be cited by AI?
Length went with citation only when it carried real substance. Most cited commercial health pages ran past 1,500 words, and a separate study found the most-used pages were long and well structured, not merely long.
Which health websites does ChatGPT cite most?
Institutions dominate. In the audit, Wikipedia, Mayo Clinic and Cleveland Clinic led, and 75.7% of citations went to institutional sources.
Sources
- Jacques, Datuowei, Jones, Basch, Vanderpool, Udeozo and Chapa (2026), Authority Signals in AI Cited Health Sources: A Framework for Evaluating Source Credibility in ChatGPT Responses (opens in a new tab), arXiv:2601.17109.
- Zhang, He and Yao (2026), From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms (opens in a new tab), arXiv:2604.25707.
- Vishwakarma, Kumar and Jamidar (2026), What Gets Cited: Competitive GEO in AI Answer Engines (opens in a new tab), arXiv:2605.25517.
- Underneath (2026), What pages cited by AI Overviews have in common: 3,096 pages