What Healthcare Clients Should Know About ChatGPT Citations
How we walk healthcare marketing teams through Promptwatch's October 2026 citation data, and why the work is about content quality and source standing, not medical claims.
The NIH earned 5.46% of all healthcare citations in ChatGPT, spread across six nih.gov subdomains. That is the headline from Promptwatch's data report on healthcare citation share, published on October 9, 2026. The window was September 9 to October 8, 2026, and the prompts were organic healthcare questions. We open healthcare client briefings with this figure because it resets expectations fast. A hospital group or a clinic brand will not be competing with a competitor down the road for these citations. It will be competing with PubMed Central (2.38%) and PubMed (2.33%).
A note on scope before anything else. We do not give medical advice, and nothing in this post tells a clinic what to say to patients. Our work with healthcare clients is about content quality, source standing, and measurement. Clinical accuracy stays with the clinicians and the compliance team.
What the numbers say
According to the report, the NIH takes more of the citations than the FDA, the NHS, CMS and SAMHSA combined. The FDA sits at 1.76%, including its accessdata.fda.gov drug database, and the NHS at 1.1%. The top non-government name is the American Society of Plastic Surgeons at 0.64%, followed by the American Academy of Dermatology at 0.54%. Mayo Clinic gets 0.49% and Cleveland Clinic 0.4%, and those are the only hospitals in the chart. Doctify and WebMD doctor profiles get 0.23% each, the only review or directory sites charted. Boots is the lone pharmacy at 0.28%.
The 20 domains in the chart add up to 15.54% of all healthcare citations. By source type, the report puts government and public health at 7.06%, NIH and research databases at 5.46%, medical associations at 1.39%, hospitals and clinics at 0.89%, reviews and directories at 0.46%, and pharmacy at 0.28%. Two associations, AAD and ASPS, together reach 1.18%, which is more than Mayo Clinic and Cleveland Clinic combined.
What it does not say
This is one vendor's dataset, one 30-day window, one engine, and organic healthcare prompts only. It does not cover other engines, other languages, or other markets, and we would not assume it does. It also shows correlation. A domain with a large share is cited often, but the report cannot tell us why, and it certainly does not show that publishing a certain kind of page causes a citation.
The 20 charted domains hold about 15.5% of citations. The rest is a long tail that the chart does not name. Plenty of citations go to pages nobody lists, and a client page could be among them. We say that out loud in briefings, because "you are not on the chart" is a different statement from "you are invisible."
What we do with it, no tools required
The report's own suggestions are modest, and we pass them on as the report's suggestions, not ours. It points to backing claims with peer-reviewed studies linked on PMC or PubMed, aligning pages with FDA, CDC, NHS and NICE guidance, working with medical associations, and keeping Doctify and WebMD profiles complete.
In practice, here is what that turns into on a content audit:
- Check every factual claim for a primary source. Not a source we find convenient, but one the clinical reviewer would sign off on. Where a claim has no good citation, it goes back to the client's reviewers before any marketing edit.
- Compare pages against the public guidance they should be consistent with. If a page contradicts current FDA, CDC, NHS or NICE material, that is a content problem regardless of any AI answer. We flag it for the clinical team and do not rewrite it ourselves.
- Map the client's relationships with associations. Many clients already take part in a professional body. Whether that body publishes anything worth linking to, and whether the client's pages reference it accurately, is a conversation for the client's team.
- Audit third-party profiles. Doctify and WebMD doctor profiles are the review and directory entries the report mentions. A profile with outdated specialties or a wrong address is cheap to fix and has nothing to do with AI.
None of this promises a citation. A page that is well sourced and consistent with public guidance is simply a better page, and if answer engines pick it up, good. If they do not, the page still serves patients and compliance reviewers.
The measurement step
The part that is hard to do by hand is knowing which sources ChatGPT actually draws on for the client's prompts, and whether that mix looks anything like the report. For client programs we run this in Promptwatch, where we track a set of agreed prompts and read the citation analytics at page and domain level. We want to see whether NIH, FDA or association pages appear next to the client's own domain in answers to the prompts that matter to that client, and which third-party pages sit in the same responses.
Two features do most of the work here. Domain-level citations show which sites appear for the tracked prompts. Page Tracker follows specific URLs, yours or a third party's, and shows which responses cite them. For healthcare clients that matters when a profile page or an association page is the one being cited, because the fix is then outside the client's own site.
If the client has crawler logs connected, Agent Analytics adds the access side: whether OpenAI's bots fetch the pages at all, and whether anything returns errors. Crawler logs start on the Professional plan at $245 a month. Essential has none, so we say that before the client picks a tier. Brand Narrative, available on every plan, pulls the claims AI makes about the client out of tracked responses and files them under themes such as Pricing, Features, Support and Reputation. For healthcare we use it as a read-only check on how the brand is described, and any claim that looks wrong goes to the client's reviewers, not to us.
The honest takeaway for a healthcare client
The report suggests that in healthcare, authority sits with public bodies, research databases and professional associations, and that the commercial and local players share what remains. We read that as a prompt to get the basics right: sourced content, consistent with public guidance, accurate third-party profiles. Then we measure what ChatGPT does with it, one window at a time, and we keep the claims about results as small as the data allows.