When a facilities director, an infection preventionist, or a design team asks an AI assistant a clinical or plant question, whose document does the answer cite? Increasingly, it cites the source itself — and that shift has a direct consequence for what hospitals and the firms that build for them should publish.
The measurement comes from Ahrefs’ Brand Radar citation data for the US, comparing 2.4 million ChatGPT answers in March 2026 with 3.8 million in September 2026, ranked across the 200 most-cited domains. In that cut, PubMed climbed from 64th to 4th among the most-cited domains, and the FDA entered at 8th. The CDC, MedlinePlus, DailyMed, and the NIH rose with them. The sites that fell were the summarizers: health explainers such as WebMD and Verywell Health lost ground, and general reference and forum sources fell sharply in the same window — Reddit from about one citation in five answers to about one in 43, Wikipedia from one in seven to one in 35. Ahrefs checked the movement against roughly 25,000 prompts present in both months; every tracked site moved the same direction on identical questions.
What counts as a source document in a hospital
A source document is the version a professional would accept in the room: the standard itself, the regulator’s own page, the manufacturer’s instructions for use, the peer-reviewed paper. In facility work it is also the owner’s own primary documents — the water management program written to ASHRAE 188, the infection control risk assessment process the construction team actually follows, the ventilation positions the plant holds against ASHRAE 170. These are documents with an author who is accountable for them, stated values, and a date.
The summarizer layer — the explainer sites, the aggregators, the marketing pages that restate a standard’s conclusions without its numbers — is exactly the layer losing citation share. An answer engine that has learned to prefer PubMed over a health explainer will, given the choice, prefer the facility’s own stated program over a vendor’s summary of what such a program usually contains.
The habit that now loses
Much of a hospital’s specification knowledge still lives where machines read worst: scanned PDFs of submittals, binder documents behind a project portal, values stated in drawings but never in text, program descriptions written as brochure copy (“we take water safety seriously”) with no temperatures, no control limits, no monitoring frequencies. None of that is citable. An engine cannot quote a number it cannot read, and it will cite the nearest readable source instead — which is how a summarizer ends up speaking for your plant.
The fix is publishing discipline, not marketing volume. State the actual values your program runs to. Name the standard and the edition. Date the page and keep it current when the program changes. Put it on the domain of the organization accountable for it — the hospital’s own site, the firm’s own site — in plain text a machine can lift without excavation.
Where to start
Water safety. Your ASHRAE 188 program is the model: control locations, temperature and disinfectant limits, monitoring cadence, corrective actions. We cover the working structure in ASHRAE 188 water management plans for hospitals and the construction-side discipline in Water ICRA: infection control for hospital water work. A published program page at that level of specificity is a source document; a paragraph saying the hospital “follows best practices” is not.
Infection control in construction. The ICRA process — who assesses, what the classes and precautions are, how containment is verified — is asked about constantly by project teams, usually through an AI intermediary first. Publish the process as you run it, with the matrix values in text, and your version becomes the citable one in your market.
For design and construction firms. Your project and capability pages are source documents too, or they are brochures. The firms whose pages state system types, standards worked to, and project parameters in plain text give the engine something to cite when an owner asks who builds this kind of space. Adjectives do not get cited; specifications do.
Read the instrument honestly
One study, one engine, free logged-out US answers — this is not a census of clinical search, and the decimals should not be treated as gospel. The defensible claim is the direction, confirmed on matched prompts: answers are being built from primary documents more, and from summaries of them less. For organizations whose specifications carry life-safety weight, that direction is enough to act on. Publish the document, not the summary of it.
The operator’s reading of the companion dataset — how often the engine searches at all, and what it searches for when it does — is on Tygart Media: Six in Seven AI Answers Never Look Anything Up.
Source: Ahrefs, “How ChatGPT Does Query Fan-Out” (October 12, 2026), Brand Radar US citation comparison — 2.4 million answers (March 2026) and 3.8 million answers (September 2026), ranks among the 200 most-cited domains; matched-prompt check on ~25,000 prompts. Figures from the separate fan-out dataset in the same article are not used here.