How AI Engines Decide Which Health Publisher to Cite

Which health websites does ChatGPT trust and cite most?

In the most carefully controlled study available, ChatGPT's top three cited sources for consumer clinical health questions were Wikipedia at 10.7%, Mayo Clinic at 9.9% and Cleveland Clinic at 9.8%, with ten organisations accounting for 52.8% of all citations. That study examined 615 sources cited across 100 consumer health questions drawn from a curated research dataset, using a fresh account and a separate session per question. A much larger study covering five engines and 824,997 citations gives a different picture because it covers a broader and more commercially weighted set of questions: there, the most cited single domain was PubMed Central at 4.08%, followed by YouTube at 3.27%. The honest answer to the question is that ChatGPT concentrates heavily on a small number of institutional sources when the question is clinical, and disperses across a much wider commercial surface when it is not. The word trust is doing unearned work in the question, incidentally. Citation is a retrieval outcome, not an endorsement, and none of these studies measured whether the cited content was accurate.

The concentrated view: ChatGPT on clinical questions

The academic study is the smallest of the three and the most carefully constructed. 100 questions randomly drawn from a curated set of 3,173 consumer health questions, one fresh session per question, 618 responses recorded and 615 usable after three broken links were dropped. Classification was done with an automated pipeline and manual verification.

Grouped as the study groups them, established institutional sources account for 75.7% of citations. On this evidence the conventional hierarchy looks broadly right, with one sharp exception: peer reviewed journals sit at 5.7%, below encyclopedias at 10.7%. An engine answering a consumer health question in plain language reaches for a plain language source, and the most cited single organisation in the entire study is Wikipedia.

Three studies, three different answers

Here are the three datasets side by side. Reading across the rows is the most useful ten seconds on this page.

Why they disagree

Four reasons, in descending order of how much they explain.

1. Query intent, which explains most of the gap. Study C used curated consumer clinical questions of the “what causes this” kind. Study B used explicitly commercial decisions of the “which provider should I use” kind. Study A sat between them and says so, noting its sample is weighted toward the categories brands choose to monitor. A clinical question retrieves clinical sources and a commercial question retrieves commercial ones. The engine is not applying a fixed trust hierarchy; it is answering the question in front of it.
2. Classification, which explains more than it should. Study A puts hospitals and providers at 1.0% and Mayo Clinic at 0.42%. Study C puts medical institutions at 30.6% with Mayo Clinic alone at 9.9%. Part of that is query intent and part of it is that the two studies drew the category boundary in different places. Study A's commercial and platform bucket at 75.4% includes general platform URLs such as google.com and play.google.com that are not health sources in any meaningful sense. Any citation statistic is partly an artefact of its taxonomy, and almost nobody publishing one says so.
3. Engine. Study A found the share of citations going to authoritative medical sources ranging from 24.1% on ChatGPT down to 15.1% on Google AI Mode. That is a real and large difference and Study C tested only the engine at the top of that range. It explains some of the gap and nothing like all of it.
4. Prompt framing. Study C explicitly asked each question with a request for sources and links. Asking an engine to show its sources changes what it retrieves. It is a defensible experimental choice and it is not how a member of the public asks a health question.

The practical rule this produces

Before you accept any statistic about AI citations, ask three questions: which engine, which kind of question, and how was a domain classified. A figure missing any of the three cannot be compared with another figure, and most of the numbers circulating in this category are missing all three.

Apply this to our own numbers as well. Everything on this page is drawn from studies with stated methods, two of them vendor produced and one an academic preprint, and all three are described above so you can weigh them. If a competing page quotes a single clean percentage for which sources AI trusts in health, that page has either not read the underlying work or has chosen the study that suits it.

The measured citation tier table for health content

This is the table the cluster was promised. It takes the tier order that our own pillar and publisher guide published as a strategic position, and puts the measured numbers next to each tier. The two frequency columns come from the two studies that classified sources by type. Study B is left out of this table on purpose, because it published a domain ranking rather than a type breakdown, and forcing its domains into these tiers would be exactly the classification laundering this page is arguing against.

Three things in that table matter more than the rest. First, the two frequency columns do not agree on a single tier, and on hospitals they are out by a factor of thirty. Second, the conventional hierarchy is not merely wrong on frequency, it is close to inverted: the tier at the top of every strategy deck sits at 3.4% in the larger dataset and the tier at the bottom sits at 6.2%. Third, and this is the part almost nobody reports, the position column tells a completely different story from either frequency column.

Read the same data by position and the hierarchy reappears

Study A recorded not only how often each source type was cited but where in the answer it appeared. Lower numbers mean earlier. Sorted by position rather than by volume, the same dataset produces this:

Read that column top to bottom and the conventional hierarchy is almost exactly right. Government first, publishers second, research third, commercial and social last. The one break is hospitals, which sit sixth on prominence despite being the single largest category in the ChatGPT clinical study. Nothing in an authority based model predicts that, and it is the strongest single hint in the data that position is driven by retrieval behaviour rather than by a trust score.

So both answers are true at once. The tier hierarchy is a reasonable model of prominence and a poor model of frequency. Our two earlier pages published it without making that distinction, which is the correction this page owes them.

Report both metrics or report neither

A citation count on its own cannot tell you whether your visibility improved. A publisher that moves from 40 citations at an average position of 6 to 80 citations at an average position of 9 has probably lost ground, because the second half of a long AI answer is read by very few people and attributed by fewer.

Two numbers, every time: citation share and average position. Study A is the only one of the three that reports position at all, which tells you how young this measurement discipline still is. If your agency or vendor reports only a count, ask for position before you act on the count.

What this table cannot tell you

Four limits, stated plainly, because a table that looks this authoritative invites misuse.

• It is a snapshot of 2026 engine behaviour, not a law. Study A covers mid April to mid July 2026, Study B July to September 2026, Study C a single day in January 2026. Every one of these engines changed its retrieval stack at least once in that window.
• It is overwhelmingly United States data. Study A is explicitly a US panel. Nothing here tells you how an engine answers a health question in Hindi, or how it weights the Indian or Gulf health web. For a market with a different publishing landscape, the honest answer is that nobody has measured it yet.
• None of the three studies measured accuracy. A citation is evidence that a page was retrieved and used, not that it was correct. Study C in particular noted a very high citation share for sources with no author attribution, which is a quality question these datasets cannot answer.
• Shares are not addressable market. A tier holding 75.4% of citations is not an opportunity for a publisher who is not in it, because much of that bucket is platform URLs such as app stores and video pages. Read the tier table to understand the competition you are in, not to pick a tier to move into.

Query intent is the strongest variable in the data

If you only change one thing after reading this page, change the unit of analysis. Stop asking which sources AI engines cite in health and start asking which sources they cite for a given kind of health question. The three studies are effectively a natural experiment on exactly that question, because each one chose a different question set and each one got a different answer.

That table is the practical core of this page. A health publisher with a strong clinical reference library is competing in rows one and two, where it has a genuine structural advantage. The same publisher trying to win row four is competing with Amazon, YouTube and Reddit on terms that have nothing to do with clinical quality. Those are two different businesses and they should not share a single visibility target.

Which means the first deliverable of any serious AI visibility programme is not a tier list, it is a question set. Building that question set properly, segmented by intent and weighted by commercial value, is the subject of the prompt map page in this cluster, and the method for sampling and running it at scale sits in the measurement page. Both of them assume the segmentation above rather than restating it.

The engines behave differently, and the gap is large

The habit of saying AI search as though it were one surface is the second most expensive mistake in this category. Study A measured the share of citations going to what it classified as authoritative medical sources, engine by engine, across the same question panel.

The spread from 24.1% to 15.1% is a nine point gap on the same questions. Put another way, a clinically authoritative publisher is roughly sixty percent more likely to be cited on ChatGPT than on Google AI Mode for the same query. Study A also found that the Google surfaces cited social and user generated content at roughly four to five times the rate ChatGPT did, which is the same finding seen from the other side.

This has an uncomfortable commercial shape. The engine where authoritative health content performs best is not the engine with the most reach. Google AI Overviews appear above the organic results on an enormous share of health queries, and it is the surface least inclined to cite the sources a medical publisher is built to be. Study B, which measured only Google AI Overviews, is the sharpest illustration: no peer reviewed journal domain appears anywhere in its top 100.

The operating consequence is that engine coverage is a strategic choice, not a technical detail. A programme that measures only ChatGPT will report flattering numbers and miss the surface most of its audience is actually using. A programme that measures only Google AI Overviews will conclude that clinical authority does not matter, and will be wrong about a quarter of the market. Measure at least one of each.

The concentration question, and why it is better news than it looks

Two findings sit in tension, and both are worth holding.

That second row is the most hopeful number on this page and it is routinely reported as bad news. A surface where no brand holds one percent is a surface where position is still being decided. Compare it with classical organic search in health, where a small group of domains has held the head terms for a decade and the cost of entry is measured in years. The citation surface is not yet consolidated, and the window for that to remain true is closing rather than closed.

How do I get my medical content cited in AI answers?

Only one of the three studies examined what cited pages had in common, and its findings contradict most published advice on the subject. Among 615 sources ChatGPT cited across 100 consumer health questions, 64.7% carried no author attribution at all and only 7.0% carried an author name together with credentials. Only 29.6% carried a medical review statement. 40.7% was published before 2020 or carried no date. The signals that were almost universally present were domain level rather than page level: a median third party domain authority score of 89 and schema markup on 74.3% of cited pages. The defensible reading is that domain strength and machine readable structure correlate strongly with being cited, while the page level trust markers the content marketing industry sells correlate weakly or not at all. None of this is causal, it comes from a single preprint, and it is still the best evidence currently available. The practical answer is to concentrate authority on fewer, deeper, better structured destinations rather than to spread bylines and review stamps across a wide library.
 

What cited health pages actually had in common

This is the authority signal audit from Study C, reported in full because the selective version of it is already circulating.

The gap between the domain authority median of 89 and the page authority median of 55 is the single most strategically useful line in this audit. It says that engines appear to be selecting at the level of the site, and then picking whichever page on that site matches the question, rather than evaluating pages on their individual merits. If that holds, a new page on a strong health domain starts with an advantage a strong page on a weak domain cannot buy. It is a correlation in one study of 615 sources and it is consistent with everything else in the dataset.

The medical review finding, in full, because it is being misquoted

Study C broke the medical review statement down by the type of organisation publishing the page. The pattern is the opposite of what a trust model predicts.

So medical review statements correlate with being a commercial publisher, not with being cited. That is a categorically different claim from the one usually made. It does not mean medical review is worthless. Reviewing health content is the right thing to do for readers, it is defensible under advertising and professional conduct rules in most markets, and in regulated pharma communications it is not optional. What it does mean is that nobody should fund a medical review programme on a business case built on AI citation lift, because the only dataset that has looked does not show one.

This complicates a page we have not written yet, and it is better to say so here than to quietly drop the thesis later. The citability page planned for this cluster argues that medical review is an underexploited asset in moving from classical E-E-A-T to AI citability. That argument now has to be made on reader trust, regulatory defensibility and conversion grounds rather than on citation frequency, and whoever writes it should start from this table rather than around it.

What to actually do, ranked by how strong the evidence is

The ranking matters more than the list. Most AI visibility checklists present twenty tactics as though they carried equal weight. These do not.

Two of those eight rows advise against work the industry currently sells, and one of them is labelled as our own inference rather than evidence. That is uncomfortable to publish and it is the only honest shape for this list. A page claiming eight evidence backed tactics from a body of research this thin would be telling you something about the author rather than about the engines.

A correction to two of our own pages

This section exists because a programme whose stated differentiator is provenance discipline cannot leave an unsupported claim standing on two live pages. Both of the pages above this one in the cluster published a citation tier hierarchy for health content. Both labelled it a strategic position rather than measured data, and both said explicitly that this page would measure it. The measurement does not support the hierarchy as published. Here is the correction, in public, rather than a silent edit.

The substantive point under the correction is worth stating on its own terms. The reason a trust hierarchy predicts prominence but not frequency is that the two are produced by different mechanisms. Frequency is a retrieval outcome and retrieval follows the question. A commercial question retrieves commercial sources because that is where the answer lives. Prominence looks more like a ranking decision taken after retrieval, and that is the stage where something resembling source quality appears to be applied. A hierarchy of authority was always a model of the second stage. It was published as though it described the first.

How to run this analysis on your own content

Short version, because the method belongs to another page. The measurement page in this cluster covers sample construction, how many questions you need, how many runs per question, how to deduplicate, and how to handle the fact that the same prompt returns different sources on consecutive days. What follows is only the shape of it, so you can judge whether the three studies above were competently built.

1. Define the question set before you look at any results. Pull from real demand data, your own site search, and the questions your commercial teams are asked. Segment by the intent types in the table above. A question set assembled after seeing which questions you already win is not a measurement, it is a trophy cabinet.
2. Run each question in a clean session. Study C used a fresh account and a separate session per question, which is the right standard. Personalisation and conversation memory will otherwise contaminate every result after the first.
3. Run across engines, not one. Minimum of one OpenAI surface and one Google surface, given the nine point authoritative share gap between them.
4. Record the position of each citation, not only its presence. This is the step almost every commercial tool skips and the one that makes the data actionable.
5. Classify domains with a written taxonomy and publish it. Study A and Study C disagree by a factor of thirty on hospitals partly because they drew the boundary differently. Your own taxonomy will shape your own answer the same way, so write it down before you start and show it to anyone who reads the output.
6. Repeat on a fixed cadence and treat single run movements as noise. These systems are non deterministic. A citation that appears one week and not the next has told you very little.

If you want a worked version of that method with sample sizes and a reporting template, the measurement page is the one to read. The prompt map page covers how to build the question set itself, which is the step that decides whether any of the rest is worth doing

Limitations, and what would change our mind

Three studies is a thin evidence base for a page this confident, so here is the full accounting.

What would change our mind. A replication of the position finding on a second engine panel would move the frequency versus prominence argument from one study to a pattern, and it is the single most useful piece of work anyone could publish next. A peer reviewed version of the authority signal audit with a larger sample would settle whether the author attribution finding is real or an artefact of 100 questions. A non United States dataset of any size would tell us whether any of this generalises. If a well constructed study shows a measurable citation lift from medical review statements, the ranked list in the previous section is wrong at row eight and we will say so on this page.

Conclusion: stop asking for the hierarchy, start asking for the question set

The reason three competent studies of AI health citations disagree is that they were never answering the same question. One asked what ChatGPT cites when a member of the public asks what causes a symptom. One asked what Google AI Overviews cite when somebody is choosing a hearing aid. One asked what five engines cite across a broad panel weighted toward the categories brands pay to monitor. Those are three different competitions, and a publisher who treats them as one market will optimise for an average that does not exist.

What does survive across all three is smaller and more useful than a tier list. Engines retrieve according to the question and rank according to something that looks like source quality. That is why the conventional authority hierarchy predicts position well and volume badly, and it is why the correct response is to choose which questions you intend to be the answer to, measure those questions by segment and by engine, and report count and position together. Everything else in this category is currently advice without a number behind it, including some of ours, which is why this page labels which is which.

The strategic framework this sits inside is in the generative engine optimisation pillar, and the commercial argument for publishers specifically is in GEO for medical publishers. Both now carry a correction pointing here. For building the question set itself, see the health publisher's prompt map. For running the measurement at library scale, see measuring AI visibility across a large content library. For the editorial standards argument, with the complication this page introduces, see from E-E-A-T to AI citability. For the markup and architecture that the schema finding points at, see condition hub architecture for AI answers.

Frequently Asked Questions For AI Citations for Health Content