Key takeaways
- Of the 615 sources ChatGPT provided in response to health questions, 75.7% came from established organizations such as healthcare institutions, government agencies, and encyclopedias.
- The top 10 sources accounted for 52.8%, showing that citations were concentrated among a small number of organizations.
- Sources that named both an author and their credentials made up 7.0%. This study alone cannot tell us which attribution practices lead to citations.
Study findings - What sources did ChatGPT cite?
U.S. researchers published this paper on arXiv in January 2026. They entered 100 randomly selected questions from a set of 3,173 consumer health questions into ChatGPT 5.2 Pro and classified the 615 sources cited in its answers.
- Established institutional sources (75.7%): Healthcare institutions 30.6% · Government 19.3% · Encyclopedias 10.7% · Professional associations 8.6% · Academic journals 5.7% · News and media 0.8%
- Other sources (24.3%): Commercial health information sites 12.4% · Individual, small-practice, and specialty sites 11.9%
- Most-cited sources: Wikipedia 10.7% · Mayo Clinic 9.9% · Cleveland Clinic 9.8%
- Attribution practices: Author name and credentials listed 7.0%; no author listed 64.7% · References listed 39.2% · No medical-review statement 70.4%
- Small-practice websites (73 sources): Dates shown for 2024–2026 61.6% (healthcare institutions 41.0%) · Structured data 78.1%
The Institute’s interpretation - Where should hospitals outside major institutions begin?
This paper was published on arXiv; we could not confirm whether it had undergone peer review. It examines one version of ChatGPT at a single point in time and does not include Korean-language questions or hospitals in Korea. It also does not identify what causes a source to be cited, so the implications for hospitals are the Institute’s interpretation.
We could not find data on citations to Korean hospitals. However, a domestic survey found that official websites - grouping hospitals, government, and public institutions - had the highest trust rating at 93.1%. Organizing official hospital information so AI can read and cite it is a starting point for GEO.
Citations clustered around well-known organizations, and this study does not show that attribution details alone lead to citations. Hospitals need to prepare both clear attribution and useful content.
- Competing on the same terms may be difficult. The top 10 sources accounted for more than half of citations. Rather than covering every health topic, it may be more practical to start with questions our hospital can answer well.
- Attribution is no guarantee of citation. This study also included many sources without a named author or medical-review statement. Even so, identifying the author and supporting sources is a basic part of helping patients assess information.
- Smaller sites can stand out through freshness. Small-practice sites accounted for 11.9% of sources, and 61.6% displayed dates from 2024–2026, compared with 41.0% of healthcare institutions. This was not established as a cause of citation, but it is something hospitals without broad name recognition can consistently maintain.
Which questions should our hospital answer first?
- Have we identified patient questions our hospital is best placed to answer in its main areas of care?
- Are author names, references, or revision dates missing from any treatment guidance?
- Does our guidance answer patients’ questions fully rather than offering only a brief introduction?
Hospitals cannot immediately match the name recognition of established organizations. But any hospital, regardless of size, can write clear answers to the questions it knows best.
Sources
Jacques et al., “Authority Signals in AI Cited Health Sources” (arXiv, 2026.1.23)
Original paper: arXiv 2601.17109
Korea Press Foundation <미디어 서베이> 2026 Issue 1: Foundation materials · 기자협회보 · Yonhap News Agency (2026.4.15)