AI Answers Your Health Questions. It Hasn't Earned Your Trust Yet

Fewer than one in twenty of them fully trust what they're told back. That's from a nationally representative survey by Gallup and West Health, published this year: among people who'd used AI for health advice in the past month, only 4% said they strongly trust its accuracy.

That gap, not the usage number on its own, is the story worth telling.

AI Didn't Make Health Information Less Trustworthy. It Made Trust Invisible.

Before generative AI, trusting a piece of health information meant doing a bit of visible work. You could see who wrote it, when, and where the claims came from. You made a judgment, even a rough one, based on what was in front of you.

Ask an AI system the same question today and you get a single, confident paragraph. The sources behind it, how they were weighed, whether they actually said what the AI claims they said: all of that gets folded into the answer instead of sitting next to it. The trust judgment still has to happen. You just can't see it happening anymore, and neither, really, can the AI.

That's the shift. Not less trust. Hidden trust.

The Doctor Is Still the Most Trusted Source. The Chatbot Is Just the Most Convenient One.

It's tempting to read the usage numbers as AI winning ground on trust. It isn't, and the data on this is unusually clear.

A Gallup study found that 73% of US adults still turn to their doctor for medical information, against 16% who reach for an AI chatbot. When Salesforce asked patients whether they trusted an AI assistant more inside their own doctor's system or as a public chatbot answering the same question, patients trusted the doctor's version three times as much. Same underlying technology. Different owner. A completely different level of trust.

People aren't confused about this hierarchy. Ask them directly and they'll tell you exactly which source they'd believe if the two disagreed. What's changed is how often they ask the convenient one anyway, as a first pass before or after seeing the person they actually trust.

That's not a trust economy in the way the phrase usually gets used, where a scarce resource gets fought over and handed to the most deserving source. It's closer to an arbitrage: people borrowing convenience from a system they've already decided not to fully believe.

Why Only 17% of What AI Tells You About Your Health Comes From a Source Worth Trusting

Here's the part that should give pause to anyone treating “get cited by AI” as the new SEO.

An independent analysis of more than 800,000 AI health citations, run by LLM Pulse, found that only 17% came from a source most people would call authoritative: government health bodies, major medical institutions, peer-reviewed research. The rest came from a long tail of commercial health sites, forums, apps and video platforms that happened to rank well for the words in the question. Separately, an audit published in BMJ Open this year found that roughly half the responses from five popular chatbots, across topics like cancer, vaccines and nutrition, were rated somewhat or highly problematic.

None of this means the AI is lying on purpose. It's built to sound confident regardless of where the answer came from, and confidence isn't the same thing as being right.

It's also not consistent. Researchers demonstrated in 2026 that rewording the same health question slightly, without changing its meaning, can change which sources an AI system cites entirely; for some systems, every single reworded version in the test pulled from a different set of sources. There's no fixed, stable answer sitting underneath, waiting to be optimized for. Which is exactly why chasing it as the goal is the wrong instinct.

What We Think Brands Should Actually Do About This

If the target keeps moving and the scoreboard is mostly invisible, the answer isn't to try harder at winning the AI's trust. It's to stop treating that as the goal.

Winning an AI's trust is a visibility contest, decided by a system nobody outside a handful of companies fully understands, running on citation patterns that can shift week to week. Earning a person's trust is a different job entirely: your claims need to hold up when a doctor, a journalist, or a genuinely skeptical reader goes looking for the source themselves. One of those is a game played against a black box. The other is a standard you set for yourself, and it happens to be the same one Google has graded health content against for over a decade. Google's own Search Quality Rater Guidelines classify health as “Your Money or Your Life” content, held to the highest quality bar it applies to any category, precisely because getting it wrong has real consequences for real people.

That standard is worth building toward regardless of what any AI model does next. Three moves, specifically:

• Get corroborated somewhere you don't control. If a claim only exists on your own domain, nobody, human or algorithm, has much reason to believe it over the hundred other claims competing for the same answer. Get your evidence repeated in a named study, an industry report, a journalist's coverage, or a client saying it in their own words.

• Write the sentence that survives a fact check. Every stat and every claim needs to name where it came from, in the sentence itself, not tucked into a footnote nobody clicks. If a claim can't be sourced properly, say plainly that it's an opinion. That single habit is most of the difference between sounding confident and being credible.

• Stop treating an AI citation as the finish line. Build for the moment after the AI answer, when a real person decides whether to believe it, act on it, or go check with someone else. That decision is still human, and it still rewards brands that earned the right to be believed long before an algorithm noticed them.

The Bottom Line

Trust in health information didn't disappear when AI arrived. It moved somewhere you can't see it happening, into a black box that borrows credibility from sources it won't always name properly, and hands it to whoever asked the most convenient question.

Brands chasing a place inside that box are optimizing for a target that changes shape depending on how a question is phrased. Brands building the kind of authority that survives someone actually checking are building something that outlasts whichever model is popular this year.

That's the harder version of GEO. It's also the only version worth doing.

If you're working out what that looks like for your own brand, that's the conversation NMQ has with clients building their GEO strategy from the ground up. Get in touch to talk it through.

References

• OpenAI, “Introducing ChatGPT Health” (7 January 2026): official confirmation that health questions are among the most common uses of ChatGPT.
• Forbes, “Everything About ChatGPT Health You Need To Know” (8 January 2026): reports OpenAI's disclosed figure of 230 million health-related questions per week, roughly 29% of ChatGPT's weekly users.
• West Health and Gallup, survey on AI use in healthcare (15 April 2026): nationally representative study; source of the 4% “strongly trust the accuracy” figure.
• TechTarget, reporting on Gallup data: source of the 73% doctor vs. 16% AI usage comparison.
• Fierce Healthcare, on the Salesforce patient trust survey (24 June 2026): source of the 3x trust gap between provider-embedded and public AI assistants.
• LLM Pulse, “What Sources Does AI Trust for Health Questions?”: analysis of 825,000 AI health citations; source of the 17% authoritative-source figure.
• BMJ Open, Tiller et al., chatbot medical misinformation audit (2026): source of the misinformation-rate finding across five popular chatbots.
• Wen et al., “Position: Generative Engine Optimization Creates Underexamined Risks” (2026): source of the citation-instability finding under query rewording.
• Google, Search Quality Rater Guidelines (official document): primary source for the YMYL classification of health content.