The bottom line: Generative AI systems structure clinic recommendations in consistent ways that can influence how trustworthy and complete options appear to patients. The way a question is framed shapes the length, organisation and tone of responses more than the clinic type or city, affecting first impressions before any quality judgement is made. As more people rely on AI for local healthcare searches, policymakers and practitioners should recognise that presentation alone can shape health decisions and public confidence.
When people search for a dentist, physiotherapist, or cosmetic clinic, many now rely on generative AI rather than traditional search engines. A new independent study suggests that the way these systems recommend clinics follows consistent patterns that may subtly shape how people interpret healthcare options.
The research was conducted by José Francisco Ouviña and focuses on how generative AI systems structure their answers when asked to recommend clinics in Spanish cities. The study does not assess whether the clinics mentioned are good or bad, nor does it verify the accuracy of the information provided. Instead, it examines the form and organisation of the responses users receive.
Explaining the rationale behind the project, Ouviña said, “This study stems from a fairly straightforward observation: more and more people are using generative AI systems as a first point of reference to find local healthcare services. However, there is very little public information about how these systems actually construct their recommendations.”
The analysis covered 1,530 responses generated in the first half of January 2026. These responses were produced by three widely used generative AI systems and were based on identical questions asked across 51 Spanish cities with populations above 100,000. The study included dental clinics, physiotherapy clinics, and aesthetic medicine clinics, reflecting common types of local healthcare searches.
Each system was asked the same fixed set of ten questions, ranging from open requests for clinic recommendations to short summary prompts. This design made it possible to isolate how question framing affected the structure of the replies. As Ouviña clarified, “The objective was not to evaluate clinical quality or care outcomes, but to observe patterns in how information is presented, organised and qualified, since this intermediate layer shapes user perception from the very first moment.”
One of the clearest findings was that the wording of a question had the strongest influence on how an answer was structured. Open or comparative questions tended to generate longer responses, often organised into lists with descriptive detail, while short prompts asking for brief explanations produced more concise replies.
There were also consistent differences between the AI systems themselves. Some generated longer and more detailed text, while others favoured shorter and more compact responses. Recurring contrasts appeared in the use of lists, warnings or limitations, references to external sources, and first person language. According to Ouviña, “The analysis also shows that different AI systems tend to maintain stable and recognisable response styles, particularly in their use of lists, disclaimers and summarising conclusions. These patterns repeat consistently regardless of location or clinic type.”
The type of clinic being recommended played a comparatively minor role. Although there were small variations in how often practical details or cautionary notes appeared, these differences were limited when set against the stronger effects of question design and system choice.
Crucially, the study makes clear what it does not attempt to do. It does not analyse clinical quality, reputation, patient outcomes, or commercial visibility, and it does not explain how AI systems decide which clinics to mention. The findings describe observable patterns in generated text rather than the underlying mechanisms that produce them.
Ouviña described the work as the first stage of a broader research programme examining AI visibility in local healthcare. Future phases will analyse which public signals appear among recommended clinics and then explore how those signals relate to visibility in AI generated responses.
