Three-quarters of what AI knows about a practice comes from somewhere else

By: MDG

Healthcare is a Your Money or Your Life category, so search engines and AI platforms hold it to a higher standard of trust than almost any other vertical. That standard is likely not met on your website. When ChatGPT answered patient-style questions about 200 independent practices, only 23% of its sourcing came from those practices’ own sites. We are seeing groups fund owned content heavily while the records that actually describe their clinicians go unmanaged.

Key Facts About Healthcare Authority In AI Search

In one analysis of roughly 4,950 patient-style queries about 200 independent practices, ChatGPT drew 27.5% of its sourcing from hospital staff rosters, 23% from practice websites and 22.1% from specialty association and credentialing pages.
There is no single AI to be trusted by. Google AI Overviews drew a third of its healthcare citations from elite hospital systems. ChatGPT drew 1% from them and 27% from government domains.
E-E-A-T is not a ranking factor, and Google says so in its own documentation. It describes what Google is trying to reward, which is more durable than a factor would be, but it is critical for this category.
Structured data does not earn AI citations. Google states no special markup is needed, and the largest study of schema adoption found no meaningful lift. It remains worth implementing for other reasons.
Half of Americans say it is at least somewhat difficult to judge whether health information is accurate. That is the problem authority actually solves.

What is YMYL, and why does it matter more in healthcare?

YMYL stands for Your Money or Your Life. It is Google’s term for topics where bad information causes real harm, and healthcare sits at the center of it. Google’s Search Quality Rater Guidelines, last revised in September 2025, define it as content that “could significantly impact the health, financial stability, or safety of people,” with a specific health category covering anything that could harm mental, physical, or emotional health.

The consequence is stated plainly in Google’s own documentation: it gives “even more weight to content that aligns with strong E-E-A-T” on exactly those topics. Experience, expertise, authoritativeness, and trust matter everywhere. In healthcare, they matter more, and Trust is the one Google calls the most important of the four.

That is the part most healthcare marketers already know. The next three sections are the part almost nobody has priced in.

Is E-E-A-T a ranking factor?

No, and getting this right separates a credible content program from a superstitious one.

Google’s documentation states directly that “E-E-A-T itself isn’t a specific ranking factor,” and that quality rater data “is not used directly in our ranking algorithms.” The rater guidelines themselves say no single rating “can directly impact how a particular webpage, website, or result appears in Google Search.”

Read that as good news rather than bad. The guidelines are not a scorecard that gets applied to your pages. They are a published specification of what Google is trying to reward, written for the humans who evaluate whether the algorithm is succeeding. A ranking factor decays with the next update. A statement of intent does not. Building against the intent is the more durable investment, and it is the only version of this work that also survives contact with AI platforms, which publish no ranking factors at all.

One caution while we are correcting the record. Several well-ranking pages currently reference a 2026 update to the rater guidelines. We could not confirm any such update from Google or from any established trade publication. September 2025 is the current revision. This category is unusually polluted with confidently written content asserting things that did not happen, which is its own argument for sourcing everything.

There is no single AI to be trusted by

The most useful finding for any multi-location group is that the major platforms do not agree on who is credible in healthcare. They are not close.

BrightEdge tracked healthcare citations across 14 weeks from October 2025 to January 2026, covering more than 113,000 URL and prompt pairs across ChatGPT, Google AI Mode and Google AI Overviews.

Source type	Share of ChatGPT healthcare citations	Share of Google AI Overview healthcare citations
Government (.gov)	27%	10%
Medical specialty organizations	17%	2%
Elite hospital systems	1%	33%

A group that optimizes against one of those trust models is close to invisible in the other. Google AI Overviews is looking for institutional healthcare brands. ChatGPT is looking for government and specialty society sources. Neither is primarily looking at your website.

This is vendor research, and we label it as such, but it is the largest dataset available on the question, and the direction is not subtle. It also lines up with an academic finding: an audit of 15,942 citations across ChatGPT, Perplexity, and Google AI Overviews, published in August 2026, found the top ten domains accounted for 43.6% of English-language citations on mental health queries. Health authority in AI answers is extremely concentrated, and the concentration differs by platform.

The Provider Record: where AI actually reads your clinicians

Here is the finding that should change how a DSO, MSO, or multi-location group allocates its content budget.

An analysis published in Medical Economics in July 2026 took 200 randomly selected independent practices from the national provider registry, across 18 specialties and 20 metro areas, and ran roughly 4,950 patient-style queries against ChatGPT. It then traced where the answers came from.

Where ChatGPT sourced provider information	Share	Who controls it
Hospital and health system staff rosters	27.5%	The affiliated institution, not you
The practice’s own website	23.0%	You
Specialty association and credentialing pages	22.1%	Boards and societies, updated by your clinicians
National directories	8.3%	Third-party aggregators

Roughly three-quarters of what an AI platform said about these practices came from records the practice does not own. Hospital affiliation pages that still list a physician who left two years ago. Board certification records that were never updated after a subspecialty. Directory entries carrying a legacy practice name from before the acquisition. None of that is a content problem. All of it is an entity problem, and it is the layer most marketing teams have never been asked to manage.

We call this the Provider Record: the set of third-party institutional sources an AI platform reads to decide who a clinician is and whether the organization behind them is credible. It has four parts, and only one of them lives on your domain.

Institutional affiliation. Hospital and health system staff pages, academic appointments, residency and fellowship listings. The single largest source in the analysis above, and the one most likely to be stale.
Professional credentialing. Specialty boards, state licensure records, professional society directories. Authoritative by construction, which is exactly why AI platforms lean on them.
Aggregator directories. Insurer provider finders, physician rating sites, health directories. Frequently wrong after an acquisition, and almost never audited.
Your own clinician pages. The 23%. Still the only layer you write yourself, and still the one that has to be right for the other three to reconcile against it.

The methodology deserves its caveats. It is a single platform, a single month, and the author discloses possible survivor bias in the sample. But nothing else published measures this, and the practical implication holds regardless of whether the true figure is 23% or 35%. If most of what AI says about your clinicians comes from records you do not control, auditing those records is not an afterthought. It is the work.

This is the healthcare-specific version of the broader model we set out in our piece on how AI platforms decide who to cite [a][b]. There the point was that roughly 90% of citations sit outside brand-owned domains. Here it is narrower and sharper: in healthcare, the off-domain layer is institutional rather than social, which means it is auditable, correctable, and almost entirely neglected.

Does schema markup or an author bio earn AI citation?

Not on the current evidence, and we would rather tell you that than sell you a tactic that does not do what it claims.

Google’s guidance on generative AI features, published in May 2026, states that “structured data isn’t required for generative AI search, and there’s no special schema.org markup you need to add,” and that no machine-readable files are needed to appear in Google Search. Its December 2025 AI features documentation says the same thing.

The empirical picture matches. Ahrefs tracked 1,885 pages that added structured data between August 2025 and March 2026, matched against 4,000 control pages using a difference-in-differences design, and found AI Overview citations moved by minus 4.6% while AI Mode and ChatGPT effects were statistically indistinguishable from zero. Separately, an agency analysis of 8,500 keywords found that the publishers with the highest rate of author bios held the smallest share of AI Overview citations.

Three things are worth saying about that before anyone strips the schema off their site.

First, the Ahrefs study pooled schema types. Author, Person, and medical review markup were not tested separately, and the sample was pages already earning heavy citation, so it says nothing about what markup does for a page nobody cites yet. Second, structured data remains how you assert entity identity unambiguously, and Google still recommends it for rich result eligibility in classic search, which is where your local and transactional demand still resolves. Third, adoption is early. A signal with no measured effect today is not the same as a signal that will never matter, and the cost of having it in place is close to zero.

So the honest position is this. Schema and credentialed bylines are hygiene, disambiguation, and risk control. They are not a citation lever today, and we expect their weight to rise as platforms mature. Anyone selling them as the mechanism that earns AI citation is asserting something the published evidence does not support.

What does move the needle?

One tactic has real measured support, and it is the least glamorous one.

The GEO benchmark study, presented at KDD in 2024 and built on 10,000 queries across 25 domains, tested what actually changes a page’s visibility in generative engine answers. Adding citations, quotations from relevant sources, and statistics raised visibility by up to 40%. Adopting an authoritative tone produced only modest gains. Keyword stuffing made things worse.

Sounding authoritative does not work. Sourcing does. The study predates the current generation of platforms and has no healthcare-specific breakdown, so treat the 40% as a ceiling rather than a forecast. The direction is the useful part, and it happens to align with everything else in this post: content that cites primary literature, names its sources, and dates its claims is the content that gets picked up, in a category where the reader cannot easily verify anything themselves.

That last point is not a marketing observation. Pew found in April 2026 that half of Americans say it is at least somewhat difficult to judge whether health information is accurate. Healthcare providers remain the most-used source at 85%, well ahead of major health websites at 60% and AI platforms at 22%. Authority is not a ranking tactic in this category. It is the thing the audience is actually short of.

The accuracy problem is a governance problem

There is a reason to take named clinical accountability seriously that has nothing to do with citation share.

A study published in npj Digital Medicine in February 2026 had 16 board-certified physicians grade 888 AI responses to 222 patient-posed medical questions. Between 5% and 13% of answers were rated outright unsafe depending on the model, with problematic response rates ranging from 21.6% to 43.2%. The responses were collected in late 2024, and models have changed since, which is a real caveat worth stating.

More directly relevant to anyone publishing health content: a BMJ Open audit published in April 2026 assessed 250 medical questions and found half the responses problematic, with reference completeness averaging 40%. Its finding on sourcing is the one to sit with: no chatbot tested produced a fully accurate reference list, and fabricated citations were common.

Now put that next to what patients are doing. Pew found in August 2026 that 34% of Americans use AI platforms for at least one health reason, including 25% trying to work out what is causing symptoms. KFF found that among people who use AI for health information, 69% trust it a great deal or a fair amount, against roughly 18% of non-users.

The professional bodies have started to respond. The AAFP now encourages organizations to disclose when AI has been used to generate or meaningfully shape educational content, and holds them accountable for accuracy and scientific integrity regardless of how the content was produced. That norm is about continuing medical education rather than marketing, but the logic transfers cleanly. If a physician’s name is on it, a human being has to have checked it.

None of this earns a citation. All of it is why medical review, named authorship, and dated content belong in a healthcare content program anyway, and why we build them in by default rather than as an upsell. Our healthcare marketing compliance playbook covers the wider regulatory perimeter.

Where to start

Audit the Provider Record before writing anything new. Every clinician, every institutional affiliation page, every board and society listing, every directory entry. Find the ones that are wrong. In an acquisitive group, they will number in the hundreds.
Reconcile to one entity. Organization name, service lines, locations, and clinician records saying the same thing everywhere. This is the highest-yield work available to any group that has acquired in the last three years, and it is the only work in this post with evidence behind it on both classic and AI search.
Test both trust models. Run your priority questions through ChatGPT and Google AI Mode separately and record who gets named. They will not agree, and the gap tells you which gap to close first.
Source everything you publish. Primary literature, named clinical author, review date visible. The one tactic with measured support, and the one that also holds up if a regulator or a journalist looks.
Keep the schema. Not because it earns citations today, but because it is how you state unambiguously who you are, and because the cost of already having it is lower than the cost of adding it later.

Frequently asked questions

What does YMYL mean in healthcare?

YMYL stands for Your Money or Your Life, Google’s term for topics where inaccurate information could cause real harm. Google’s Search Quality Rater Guidelines define it as content that could significantly affect health, financial stability, or safety, with a specific category for anything affecting mental, physical, or emotional health. Google states it gives even more weight to strong E-E-A-T signals on these topics.

Is E-E-A-T a ranking factor?

No. Google’s documentation states that “E-E-A-T itself isn’t a specific ranking factor” and that quality rater data “is not used directly in our ranking algorithms.” The rater guidelines describe what Google is trying to reward rather than a scorecard applied to individual pages, which makes them a more durable thing to build against than a ranking factor would be.

Does schema markup help a healthcare page get cited by AI?

Not measurably today. Google states no special markup is required for generative AI features, and a study of 1,885 pages that added structured data found no meaningful citation lift on any platform. Schema still earns rich results in classic search and remains the clearest way to assert entity identity, so it is worth implementing. It is not an AI visibility strategy on its own, and we expect its weight to rise as adoption matures.

Why does ChatGPT cite different healthcare sources than Google?

Because they apply different trust models. Across 14 weeks of healthcare queries, Google AI Overviews drew 33% of citations from elite hospital systems and 10% from government domains, while ChatGPT drew 1% from hospital systems and 27% from government sources. A program built against one platform’s preferences will underperform badly in the other.

Where does AI get information about a medical practice?

Mostly from somewhere other than the practice. In an analysis of roughly 4,950 patient-style queries about 200 independent practices, ChatGPT drew 27.5% of its sourcing from hospital and health system staff rosters, 23% from the practices’ own websites, 22.1% from specialty association and credentialing pages, and 8.3% from national directories.

Should healthcare content have a named clinical author?

Yes, though not for the reason usually given. There is no evidence that credentialed bylines earn AI citations. They matter because half of Americans find health information hard to evaluate, because AI-generated medical content has a documented accuracy problem, and because professional bodies including the AAFP now expect disclosure of AI involvement and named human accountability for accuracy.

Ask for an authority and E-E-A-T content audit

MDG builds provider-authored, source-cited authority content and the entity work underneath it: clinician and organization records reconciled across institutional affiliations, credentialing bodies and directories, listings accuracy across 60 or more directories, schema and entity engineering, and AI visibility tracking measured separately for each platform. It runs as one in-house discipline alongside paid, local, and organic search rather than as a content bolt-on.

The audit maps your Provider Record clinician by clinician, shows where the records disagree, and benchmarks how often each major platform names you against your competitors. We serve more than 2,700 client locations across 48 states. Our approach is all informed by more than 25 years of doing multi-location, high-growth marketing for PE-backed companies.

Start the conversation here, or read more about how we support multi-location healthcare organizations.