AI visibility for
health tech.

Health is the most heavily filtered vertical there is. Two stages run before source selection: a safety layer that hedges or refuses specific medical guidance, and a source-preference layer favouring government health agencies, academic centres and peer-reviewed literature. Commercial health sites are not blocked outright, but they enter the candidate set only when they look like clinical publications rather than marketing.

The composed answer to a general health question is therefore general information, a recommendation to consult a clinician, and citations to institutional sources. Product names appear only when the question is explicitly commercial — 'HSA-eligible fitness apps', 'telehealth that accepts <insurer>'. The winnable surface in health tech is narrower than in any other vertical, and knowing where its boundary sits is most of the strategy.

What makes a commercial health page eligible is mechanical: a named reviewer with a credential, a visible review date, links from clinical claims to primary literature such as PubMed Central–indexed studies, and unambiguous language about what the product is. MedicalWebPage and Physician schema make the reviewer machine-readable rather than merely present on the page.

FAQPage markup deserves a specific note here. Google's August 2023 restriction of FAQ rich results kept well-known authoritative government and health sites as the exception — which does not extend to a commercial health-tech brand. The search feature is gone for you; the four-point extraction value in the rubric is not. Ship the markup for the model, not for the search result.

Coverage and eligibility queries convert harder than anything else in this vertical and are almost unclaimed. 'Telehealth providers that accept <insurer> in <state>', 'HSA-eligible <category>' — factual, high intent, and the answer is a table nobody publishes because maintaining it is tedious rather than difficult.

Comparison queries with a measurable attribute — 'sleep tracker accuracy Oura vs Whoop' — are winnable if you have published validation data and unwinnable if you have not. The attribute has to be measured by somebody, and the citation follows the measurement rather than the marketing.

Price-without-insurance queries are the third unclaimed shape. 'How much is an online therapy session without insurance' is asked constantly and answered vaguely, because publishing a number feels like a commercial risk. It is the same trade a law firm makes with fee ranges, and it pays the same way.

Symptom, treatment and dosage queries are the vanity trap here, and they are worse than vanity: they carry enormous volume and are structurally closed to commercial sources. The safety layer routes them to institutional sources before extractability is ever assessed, so the quality of your content is not the operative variable. Dosage and diagnosis phrasings are frequently refused outright.

'Is <brand> legit' behaves as it does in ecommerce — decided by third parties, regulatory status and discussion you do not control. Publish your regulatory facts and let the query resolve itself.

Oura's position rests on something most health-tech companies never attempt: peer-reviewed publication of its own sleep data. The citation the model wants already exists inside a corpus it trusts, and the brand name travels with it. This is the single highest-leverage move in the vertical and also the slowest to execute.

Headspace maintains a clinician-reviewed library with visible reviewer credentials, which passes the attribution gate at scale. Ro states its regulatory status as a plain fact rather than a trust badge. Hims & Hers runs a page per condition per treatment, so there is always a page matching the shape of the query. Calm covers the phrasing variants people actually use for mental-health questions, in FAQ blocks that lift cleanly.

Four of those five are reproducible without a research budget: the credentialed reviewer byline with a date, the condition-per-page structure, the plain regulatory statement, and citing peer-reviewed literature you did not produce. Citing PubMed Central–indexed studies costs nothing and is what separates a health page that gets extracted from one that gets filtered. Only the original research requires real investment.

Wellness-blog volume. Publishing two hundred posts on sleep hygiene or gut health does in this vertical what it does nowhere else — nothing at all. The safety and source-preference layers remove commercial sources from general health answers before extractability is assessed, so the long-tail coverage that volume buys in software or ecommerce simply does not exist here.

Testimonials and before-and-after content. In a category judged on evidence, an unverifiable personal outcome is the weakest possible sentence, and it occupies the exact position where a cited study should be sitting.

'Clinically proven' with no linked study. The link is the entire signal. An unlinked efficacy claim on a commercial health page is the precise pattern the source-preference layer exists to filter out, so the phrase actively works against you.

'Medically reviewed by our clinical team'. No person, no credential, no date. It fails the verifiable-byline check worth two points and, more importantly, it fails the vertical's attribution gate — which is looking for somebody whose credential can be checked, not for a reassuring phrase.

Regulatory ambiguity. Leaving 'is this a medical device, a wellness product or a telehealth service' unanswered is a trust cost in a category where scope is the first thing both a reader and a model try to establish. HowTo markup on treatment steps fails similarly: it is deprecated as a rich result and it frames clinical guidance in a way you do not want attributed to you.

Health-tech benchmarks measure attribution and scope clarity. Unlike every other vertical, no amount of publishing volume compensates for a failure in the first two rows — they behave as gates rather than as weights, which is why large content investments here so often return nothing measurable.

Fix reviewer attribution and literature linking, then judge everything else. A site failing either of the top two rows will not move on the remaining four, and teams routinely misdiagnose that as a content-quality problem when it is an eligibility problem.