Health systems measure market share by service line. They measure referral patterns, payer mix, campaign attribution, cost per acquisition, and downstream revenue per encounter. The instrumentation is sophisticated and the reporting is monthly.
Almost none of them can answer this question: when someone in our market asks an AI assistant where to go for a hip replacement, how often do we come up?
That was a reasonable gap two years ago. It is no longer.
The shift, in numbers
In a survey of nearly 1,000 US adults published in June 2026, rater8 found that among patients who had actively searched for a doctor in the previous year, 36% cited AI tools such as ChatGPT as an influence on their choice. The prior year's figure was 17%. It roughly doubled in twelve months.
The comparison that should stop a marketing executive is what sits beside it. In the same survey, Google search was cited by 34% and personal recommendations from a doctor by 32%.
of patients who searched for a doctor in the past year cited AI tools as an influence — ahead of Google search (34%) and physician referral (32%). The figure was 17% a year earlier.
Source: rater8 Patient Choice Report, ~1,000 US adults, June 2026
Read that carefully. AI assistants did not merely become significant. On this measure they moved past general search, and past word of mouth from a physician — the channel healthcare marketing has treated as unassailable for decades.
The second finding is quieter and arguably more consequential. Asked which part of a Google results page they trust most for healthcare research, patients named AI Overviews at 37% — against 20% for the organic blue links, 13% for the local map pack, and 7% for sponsored results.
So the synthesized answer at the top of the page is now not just the most seen element. It is the most trusted one. A health system that ranks third organically and is absent from the AI Overview has lost the part of the page that carries the most weight, while its analytics still report a respectable position.
Why this is not an SEO problem wearing a new hat
The instinct is to treat this as search engine optimization with different plumbing. It is a different problem, and the difference is structural.
Traditional search returns a ranked list. Ten results, and a patient scans several. Being fourth is worse than being first but it is not being nowhere — you are on the page, you can be clicked, the competition for attention is visible and gradual.
A generative system returns an answer. It names two, three, perhaps five organizations. There is no fourth position. You are in the answer or you are not, and the patient generally never learns that alternatives existed. The gradient collapses into a binary.
There is no page two of a ChatGPT answer.
The second structural difference matters more for healthcare specifically: the sources are different. Generic advice about generative search reduces to content, schema markup and links — which is roughly right for a software company, whose own website is the primary text about it in the world.
Healthcare is not like that. A hospital's own website is a modest fraction of what is written about it. There is a dense substrate of registries, directories, ratings bodies and credentialing data that describes healthcare organizations in structured, machine-readable, heavily-crawled form: NPPES and NPI records, Google Business Profile, Healthgrades, Vitals, Zocdoc, WebMD Care, US News, Leapfrog, CMS Care Compare, payer provider directories. When a model constructs an answer about where to get care, it draws disproportionately on that substrate.
Which produces a counterintuitive result that we see repeatedly. An organization can have an excellent website, a large content operation and strong conventional SEO, and still be poorly represented in generative answers — because its NPI records are stale, its specialty taxonomy in Google Business Profile is wrong, and it is listed at three slightly different addresses across four directories. The model resolving that ambiguity resolves it toward a competitor with a cleaner record.
Brand visibility and service line visibility are two different problems
This is the finding that most surprises health system executives, and it is worth stating plainly.
"Tell me about Memorial Health System" and "best hospital for joint replacement in South Florida" are not variations on one question. They are two different retrieval problems with two different source sets, and an organization's position in one predicts very little about its position in the other.
Institutional authority — the thing built over decades through brand, size, reputation and press — helps considerably with the first. It helps much less with the second, where the model is looking for evidence about a specific clinical program in a specific place: procedure volumes, sub-specialty depth, named surgeons, outcomes, program-level pages, condition-specific content.
The practical consequence is that a well-known system routinely discovers it is strong in three service lines and effectively invisible in two others — including, often enough, a line it has recently invested heavily in. Averaged across the organization, the picture looks fine. Broken out by service line, it does not.
The same is true of geography. "In South Florida" is not a modifier on the question; it defines the competitive set entirely. A system operating in four markets has four different pictures, and averaging them destroys the finding.
Accuracy is the other half, and it is the half that reaches the CEO
Everything above is about presence. There is a parallel problem about correctness that gets far less attention and moves faster inside an organization.
Generative systems describe organizations as well as naming them, and they are confidently wrong at a non-trivial rate. We routinely observe AI-generated descriptions of health systems that reference service lines that have closed, attribute physicians who left years ago, misstate locations and affiliations, or surface a lawsuit or regulatory action with no context and no resolution.
No one is monitoring this. There is no alerting for it, no owner, and no established remediation path. It propagates silently, and the organization typically learns about it when a patient, a referring physician or a board member mentions it — which is the worst possible discovery mechanism.
What to do about it
The honest first step is not a project. It is a measurement, because almost every organization's assumption about its own position turns out to be wrong in one direction or the other.
A useful measurement has four properties. It runs the same questions repeatedly, because generative systems are non-deterministic and a single run tells you nothing. It covers the platforms patients actually use rather than just the convenient one. It scores presence, recommendation and citation separately, because appearing in a list and being recommended are very different commercial events. And it reports against a named competitive set, because a visibility figure with no comparison is a number without meaning.
Once you have that, the work becomes ordinary — and ordinary is good news. Fix the entity records. Build genuine service-line depth where the model has nothing to draw on. Correct the third-party sources doing the most citation work. Monitor the accuracy problem so it stops being discovered by accident.
None of it is exotic. It is simply unassigned, which is why the practices MGMA surveyed named ownership as a barrier alongside time and money. The organizations that move first here will mostly be the ones that gave it to somebody.
Related
Where does your organization appear?
A Share of Discovery assessment measures presence, recommendation and citation by service line, market and competitor — across all four major platforms.