How to Check for Yourself How Your Clinic Appears to AI

How is your clinic treated in AI search? Even without any special tools, you can begin observing simply by doing the work yourself. This article is that procedure. It is published ahead of time as one chapter of the AI Search Handbook for medical institutions.

First, prepare your questions. Create three categories. First, questions by name—"What kind of clinic is (clinic name)?" "What is (clinic name)'s setup for painless (epidural) delivery?" These reveal what the AI knows about your clinic and which sources it speaks from. Second, condition-based local questions—"Which dermatology clinics in (city name) see patients on Saturdays?" "Which facilities in (area name) offer 24-hour support for painless (epidural) delivery?" In a form close to a patient's own inquiry, you see whether your clinic is surfaced as a candidate. Third, comparison questions—"What is the difference between (your clinic) and (a nearby clinic)?" For each, create two or more variations with different wording. You will find later that even with the same intent, the wording changes the answer.

Next, the environments. Do not test with a single AI and stop there. Try at least three—ChatGPT, Claude, and Gemini—and, if possible, switch Web search on and off as well. The reason lies in the actual measurement data. In this site's verification (150 responses), even with the "Web search on" setting, the proportion of cases where a search actually triggered ranged from 0% to 83% depending on the service. An AI that does not search answers only from the knowledge it had at training time, so even if you improved your site yesterday, it is not reflected in the answer. An AI that searches looks at today's web. In other words, depending on which environment you ask in, "what it is looking at" is fundamentally different.

Keep records. Date, service name, Web search setting, the question text, whether your clinic was mentioned, the clinics surfaced as candidates, the pages shown as citation sources—record these seven items in a table. The citation sources in particular are important. When your clinic is described, is it from your official site, from a portal site, or from reviews? Where the "material" for the AI's answer lies will tell you where you need to improve.

And the most important caution. Do not judge from a single result. Even with the same question, the answer fluctuates. A single judgment—"it didn't come up as a candidate, so it's no good," or "it came up, so we're safe"—holds no statistical meaning. Repeat the same question set once a week and record the changes—this is the form that holds meaning as observation. Also, repurposing the content of an AI's answer into advertising material can create problems under the Medical Advertising Guidelines, so avoid it. Observation is, ultimately, material for reviewing how your clinic communicates its own information.

Manual observation also has its limits. What you can see with a few questions × a few environments × once a week goes only to the entrance of a trend; comprehensive coverage of question patterns, quantification of fluctuation, and tracking changes in competing candidates require systematic verification on the scale of hundreds of trials. The verification reports this site is preparing are meant to take on that part. We believe the healthiest way to use this is to first begin observing with your own hands, and then bring in the questions that arise from it.