Medical and Healthcare: How Patients Discover and Choose Providers Now

The Old Baseline: Search Was Already the Front Door

Long before anyone was asking a chatbot about a symptom, search was already the default entry point for health questions. Pew's 2013 research found that 59% of US adults had looked online for health information in the prior year, and among them, 77% started at a general search engine like Google, Bing, or Yahoo rather than going directly to a dedicated health site like WebMD. A third of adults, 35%, had gone online at some point specifically to try to self-diagnose a condition. That baseline matters because it establishes the trend line the current AI-chatbot conversation is actually competing against: general search has been the dominant health-research habit for well over a decade, and any new surface has to take share from that habit, not invent a new one from nothing.

Search has been the dominant health-research habit for well over a decade; anything new has to take share from that habit, not invent one from nothing.

A Third of Adults Now Ask a Chatbot First

KFF's tracking poll, fielded in late February and early March 2026 among 1,343 US adults, found that 32% had used an AI chatbot for health information in the past year, split between 29% who used one for physical health questions and 16% for mental health questions. That is no longer a fringe behavior; it is a share of the population comparable to the 36% who use social media for the same purpose (Pew, April 2026). What KFF's data also surfaces is the risk sitting inside that convenience: 41% of AI-health-info users, or 13% of all adults, have uploaded personal medical information into a chatbot, and 77% say they are concerned about the privacy of medical data shared with an AI tool. People are using these tools, and worrying about them, at the same time.

Convenient Is Not the Same as Trusted

Pew's April 2026 survey, fielded the previous October among 5,111 US adults on its American Trends Panel, put a hard number on the trust gap. Across the population, 85% get health information from healthcare providers, 60% from major health websites, 36% from social media at least sometimes, and 22% from AI chatbots at least sometimes. But when asked about accuracy specifically, 65% rate their provider's information as highly accurate, compared with just 18% for AI chatbots and 7% for social media. Pew's own framing captures it precisely: users of AI and social media for health information are more likely to call them convenient than to call them accurate. That distinction, convenience without trust, describes the AI-answer layer in healthcare right now. People are willing to ask; they are not yet willing to rely.

Users of AI and social media for health information are more likely to call them convenient than to call them accurate.

Google's Two Different Answers to the Same Category

The most striking pattern in this research is how differently Google treats the same subject depending on what the searcher seems to want. BrightEdge's daily tracking of health-related keywords, a single-vendor, proprietary methodology that has not been independently audited and should be read as directional rather than definitive, recorded AI Overview presence on general healthcare queries rising from 59% in December 2023 to 84% a year later and 89% by December 2025. Over that same window, AI Overview presence on local, provider-intent queries, the kind of search someone runs when they already know they need a dermatologist or a family doctor nearby, fell from 100% in December 2023 to 14% a year later and effectively 0% by December 2025. Underneath that behavior sits Google's own quality standard: health and medical topics are almost always classified as Your Money or Your Life content, a category Google introduced into its Search Quality Rater Guidelines in 2013, and Google's guidance is explicit that trust is the most important of the Experience, Expertise, Authoritativeness, and Trust signals it evaluates. Read together, the pattern suggests Google is comfortable summarizing general health information but is deliberately routing the moment someone is choosing a specific provider away from a generated answer and toward the local pack, Maps, and organic results instead.

The Review Is the New Referral

If the AI-answer layer is not deciding who gets chosen, something is, and the data points squarely at reviews. rater8's 2025 report on how patients choose their doctors, based on a survey of 1,008 US adults fielded in December 2024, found that 84% of patients check online reviews before selecting a new provider, 61% say they prioritize online reviews over a referral from friends or family, and 51% read at least six reviews before booking. Negative reviews are not passive background noise either: 40% of patients have canceled an appointment or changed a care plan because of one. Yet the supply of that signal is thin. Only a minority of patients generate it: 57% say they rarely or never leave a review themselves, which means the reviews that exist carry outsized weight over the ones that were never written.

A Legal Floor Under the Review Signal

That review economy now operates inside a real legal boundary. The Federal Trade Commission's Trade Regulation Rule on the Use of Consumer Reviews and Testimonials passed by a 5-0 vote and took effect October 21, 2024. It prohibits fake or fabricated reviews, undisclosed insider reviews, the suppression of negative reviews, and outright review-buying. For a healthcare practice, that closes off the shortcuts, buying reviews, hiding bad ones, having staff post as patients, that some competitors have relied on, and it puts the honest weight back on actually earning reviews at the volume the data above shows patients expect to see.

What Patients Are Doing With AI Tools Today, Specifically

A second rater8 report, fielded in June 2025 among 1,024 US adults, gives a more granular picture of AI's early role inside the provider-selection process itself. It found 73% of patients had adopted a new provider-research behavior or tool during 2025, and within that, 31% had used a generative AI tool such as ChatGPT or Google's AI Overviews to research or compare providers, with 26% saying an AI tool directly influenced their eventual choice. Alongside that, 25% had begun using voice assistants to research providers, and 35% said they chose a provider partly based on that provider's social media presence. None of these figures suggest AI has replaced search, reviews, or word of mouth. They suggest it has been added as one more input that a meaningful minority of patients now consult on the way to a decision, most often early in the process rather than at the moment of booking.

What the Medical Literature Actually Says About AI Answers

Underneath all of this sits a harder question: when a patient does get an answer from an AI tool, is it any good? The peer-reviewed evidence is real but narrow, and the researchers themselves are careful not to overclaim it. A 2026 study in Nature Scientific Reports tested ChatGPT and DeepSeek against 123 oral and maxillofacial radiology exam questions across three phases between May 2025 and February 2026, producing 1,230 total responses. Overall accuracy was 83.7%, and the strongest predictor of a correct answer was the cognitive complexity of the question itself, not which model answered or when, a relationship the researchers measured at an odds ratio of 6.15 (p=0.003). A separate 2026 study published in Medicine tested DeepSeek V3 and ChatGPT-4o against 52 patient questions about esophageal cancer, with both models scoring a median 4 out of the rated scale (interquartile range 3 to 4) and no significant difference between them. But the study's own conclusion is explicit: those results do not confirm the models' general reliability for esophageal cancer health information in routine clinical or public use. That caveat, from the researchers who ran the study, describes the state of this evidence: accuracy that is real and often high on the specific questions tested, paired with an explicit refusal by the people who tested it to generalize beyond that.

The researchers' own conclusion is explicit: strong scores on the questions tested do not confirm general reliability for routine clinical or public use.

Local Intent Still Rules the Moment of Choice

The broader local-search context reinforces why Google keeps the local pack, not a generated answer, in front of provider-intent queries. A compilation of BrightLocal, Whitespark, Google, and Semrush data published by Digital Applied, itself a third-party synthesis of other firms' primary research rather than a primary or independently audited source, so its figures are contested rather than established, estimates that 46% of all Google searches now carry local intent, up from roughly 30% in 2019, that the top local-pack position earns roughly 1.8 times the click-through rate of the top organic result, and that 76% of people who run a local search take an action, typically visiting the business, within 24 hours. That pattern lines up with a broader shift toward digital care touchpoints generally: Deloitte's 2024 consumer survey found 64% of consumers say virtual visits are more convenient than in-person care, their top reason for choosing one, and that 43% used a connected health monitoring device or digital health tool in 2024, up from 34% in 2022. Put simply, patients are more digital across every part of the care process, but the specific act of choosing who to call still runs through the local pack and the reviews sitting inside it, not through a chat window.