OpenAI put health records into ordinary ChatGPT conversations — including, by its own example, choosing where to eat

What OpenAI actually said

Quoting the announcement rather than the write-ups: "Health in ChatGPT is launching to U.S. users. You can choose to securely connect Apple Health and supported medical records so ChatGPT can help you understand your information in context." Rollout is "to logged-in ChatGPT users 18 and older in the U.S. on web and iOS across Free, Go, Plus, and Pro plans today." OpenAI puts the demand at "more than 300 million people" per week asking ChatGPT health-related questions. Connected records and Apple Health data, and conversations using them, are "not used to train our foundation models or target ads." Health is not yet available in Codex.

The structural change is the second half. An earlier limited release put health behind a dedicated space; OpenAI reports that "more than 70% of health-related conversations took place outside the dedicated health experience", so the connected context now travels across ordinary conversations rather than living in a separate tab. That is the design decision that makes this a local-search story rather than only a health-product story.

Holding it to its stated width

OpenAI gave exactly two examples of connected health context shaping a non-medical answer: the restaurant/dietary-restriction one, and planning lower-impact weekend activities around a recent injury. It said nothing about using health data to select a provider — no claim about which dentist, med spa, clinic or physiotherapist ChatGPT will name. We are not going to extend it there for them, and neither should a pitch deck. What is established today is narrower and still significant: local recommendations are now, by the vendor's own example, personalised on private data. Whether that reaches provider shortlists is an open empirical question, and the honest answer for the next few months is that nobody outside OpenAI knows.

Our baseline, so the change is measurable later

The only useful thing to do with an announcement like this is write down what the world looks like before it, so the after is checkable. From a production export taken today: our corpus holds 112 health-adjacent local businesses (med spas, dentists and dental practices, plus a menopause clinic and independent pharmacies) out of 247 audited, 663 answers to category buyer questions asked 2026-06-12 to 2026-07-14 — never the business's name.

The whole medical-authority layer, in these answers, is Zocdoc (25), WebMD's doctor pages (14), Healthgrades (12), Opencare (12), RealSelf (6) and RateMDs (5) — provider directories with reviews, not clinical references. Ask ChatGPT today which med spa or dentist to use and its evidence is the same generic local-directory pile it uses for a roofer. That is the baseline a personalised health context would be arriving on top of.

What this means if you sell or buy AI visibility

• Personalisation is the thing that breaks third-party measurement, not the thing that breaks ranking. An answer conditioned on the asker's connected records is not reproducible by anyone else — not by the business, not by its agency, not by us. Any tool that reports a single "your AI visibility score" without a sampling method is already reporting one draw from a distribution; connected personal context widens that distribution without announcing itself.
• Repeat-sampling stops being a nicety. We published on Tuesday that asking ChatGPT the same buyer question eight times produced 27 different businesses. That churn is measured on anonymous, context-free asks. Personal context is a second source of variance stacked on the first, and the only defence against both is repeated sampling with the spread reported.
• Do not sell "optimise for health personalisation." There is no surface to optimise: the data is the user's, it is explicitly excluded from ad targeting and model training, and OpenAI has published no signal a business can supply or influence. The defensible work is unchanged — be present, accurately and consistently, in the third-party sources the engine actually reads for your vertical.
• Health-adjacent verticals should be watched, not re-priced. Med spas and dental practices are 112 of our 247 audited businesses, and their answers today lean on provider directories rather than clinical sources. If that composition shifts after this rollout, it will show up as a change in the cited-source mix — which is measurable, unlike the personalisation itself.

Limits

• This is an announcement, not an observed change. Everything above about personalisation is what OpenAI says the product does. We have measured no effect on local recommendations and are not claiming one.
• Our baseline is a convenience sample — 112 health-adjacent businesses we chose to audit, single-sampled per question, US and UK metros, ChatGPT and Perplexity only. The citation figures are ChatGPT-only because the Perplexity API does not return citation URLs in the answer body.
• "Medical or provider-credential source" is a hand-curated domain list (WebMD, Healthgrades, Zocdoc, Opencare, RealSelf, RateMDs, Vitals, Mayo Clinic, NIH, Cleveland Clinic, ADA, AAD). Unrecognised sites count as not-medical, so 15.6% is a floor.
• We sell this measurement. AskedAbout sells AI-visibility audits, so check the method rather than the conclusion.

You can run the free 60-second check on a med spa, dental practice or clinic and see which businesses the engines name instead, and what they cite. Agencies baselining a book of health-adjacent clients before this rollout lands can run five at once with the $249 Agency 5-pack — 25 buyer questions × 4 engines × 3 samples per business, white-labeled.