To track AI visibility for a clinic client, run a fixed panel of patient-style prompts (specialty, city and intent) in the same AI engines on a set schedule, and log whether the clinic is named, where, and which sources are cited. Report the mention rate over time, not single answers, because responses vary between runs.
Last verified: November 9, 2026.
Clinic owners have started asking their agencies a new question: "Does ChatGPT recommend us?" The tempting answer is a screenshot. It is also the least reliable answer you can give, because the same prompt can produce a different set of clinics an hour later, from another city, or in another engine.
Most guides to AI visibility tracking are written for national brands with one product and one audience. Healthcare clients are local, multi-provider and sometimes multi-location, and their patients ask location-bound questions. This guide covers how to design a prompt panel for a clinic, which engines to include, how to deal with answer variance, and how to present the result so a practice owner can act on it.
What to track in each AI answer
For every prompt and engine, record the same five fields:
Mention. Is the clinic named anywhere in the answer? Count the practice name, common short forms, and the names of its providers if the client wants provider-level tracking.
Position in the answer. First recommendation, in a list, or only in a footnote. Being one of three named clinics is different from being the eighth link.
Citation. Is the clinic's own website cited as a source?
Cited sources. Which domains the engine cited: directories, review sites, news, hospital systems, competitor sites. This is where the fixes come from.
Accuracy. Did the answer state anything wrong about the clinic, such as hours, insurance, a provider who left, or an old address?
The accuracy field is the one healthcare clients care about most once they see it. A wrong statement about which insurance a clinic accepts is a patient-experience problem, not just a marketing one.
Building a prompt panel for a clinic
A prompt panel is the fixed list of questions you ask every cycle. Build it from three parts.
Specialty
Use the words patients use, not billing terms. "Kids' dentist" and "pediatric dentist" can produce different answers; include both if both are common in the client's area.
Place
Use the city and state, and add neighborhoods or nearby towns the clinic draws from. A multi-location group needs a separate set per location.
Intent
Cover the different reasons someone asks:
| Intent | Example prompt |
|---|---|
| Recommendation | Which pediatric dentists in Boise, ID would you recommend? |
| Best/top | Who is the best pediatric dentist in Boise, ID? |
| New patient | Pediatric dentist in Boise, ID accepting new patients |
| Insurance | Which pediatric dentist in Boise, ID takes Delta Dental? |
| Access | Pediatric dentist in Boise, ID open on Saturdays |
| Reputation | Most trusted pediatric dental practice in Boise, ID |
| Booking | Where can I book a pediatric dentist appointment in Boise, ID? |
Fifteen to twenty-five prompts per location is a workable size for monthly or weekly tracking. Fewer than ten makes the mention rate jumpy; many more makes the review work expensive across a client list.
Keep the panel fixed. If you change wording every month, you cannot tell whether a change in mentions came from the engine or from your edits. Add new prompts as a separate group and keep the core set stable for at least a quarter.
Avoid health details in prompts. Use access and specialty questions rather than symptoms tied to a person. You are measuring visibility, not testing medical advice.
Which engines to include
Pick the engines your client's patients are likely to use and that you can query consistently. For most clinic clients that means three:
ChatGPT with web search. OpenAI's help center says ChatGPT may rewrite a question into one or more search queries sent to search partners, may use an approximate location from the user's IP address, and may use saved memories when rewriting queries. It also says search results and citations "can be incomplete, outdated, or incorrect." For a website to be eligible, OpenAI says to allow OAI-SearchBot to crawl it.
Perplexity. An AI answer engine that patients can use for local questions; log its cited sources the same way.
Google AI Mode and AI Overviews. Google says the two may use different models and techniques, so their responses and links will vary, and both may use "query fan-out," running several related searches to build an answer. AI Overviews appear only when Google judges them additive and often don't trigger.
Log each engine separately. Averaging them into one score hides the most useful finding, which is usually that a clinic is named by one engine and missing from another.
Controlling for location
Because local answers depend on location, put the city and state in the prompt text rather than relying on the engine to infer it. OpenAI's own guidance for better local results is to include your city, neighborhood or postal code in the question. If you run prompts through an API or a tracking tool, record the location setting it uses.
Variance: why one-off checks mislead
AI answers are generated each time. The same prompt in the same engine can name a different set of clinics on two runs, and engines update their models and search behavior without notice. Several documented behaviors add to this:
ChatGPT may rewrite the question into different search queries, and memory or IP-based location can change them.
Google's AI features may run multiple related searches, and AI Overviews do not trigger on every query.
In Google Search Console, a follow-up question in AI Mode counts as a new query, which shows how quickly the context shifts.
We are not quoting a variance rate here; we have not published one, and you should be wary of anyone who gives a single universal figure. What matters is the reporting discipline that follows:
Report rates, not anecdotes. "Named in 9 of 25 prompts in ChatGPT this month" is more reliable than "ChatGPT recommended us."
Use a consistent cadence. Weekly runs smooth out noise better than one monthly check. Report the monthly total of weekly runs.
Show a trend over at least three periods before you call anything a change.
Treat small moves as noise. If the clinic went from 8 to 10 mentions out of 25, do not claim a win. If it went from 2 to 12 over three months, that is worth discussing.
Keep raw answers. Store the answer text and cited links for each run so you can explain any number in the report.
Reporting AI visibility to the client: one chart, one table
Clinic owners do not need a dashboard. Two elements are enough.
The chart: mention rate per engine per month (prompts where the clinic was named divided by prompts run), as three lines. Add a marker for each month the agency made a relevant change.
The table: the sources engines cited most often for the client's prompts, with a status column.
| Cited source | Times cited this month | Clinic listed there? | Details correct? | Action |
|---|---|---|---|---|
| A local directory | 14 | Yes | Old phone number | Update listing |
| A review platform | 11 | No | n/a | Claim profile |
| A competitor's site | 9 | n/a | n/a | None |
| Client's own site | 4 | Yes | Hours outdated | Fix hours page |
Then add one paragraph: what changed, why you think it changed, and what you will do next. For how this section fits into the client's wider monthly report, see monthly SEO reports for healthcare clients.
Linking AI mentions back to fixes
AI visibility is only worth reporting if it leads to work you can do. The fixes usually fall into four groups:
Access. Make sure AI crawlers and search bots can reach the site. OpenAI names OAI-SearchBot for ChatGPT search eligibility, and Google says AI features use the same Search technical requirements and robots.txt controls as regular Search. Our AI search readiness checklist covers robots.txt, server-rendered pages and llms.txt, and there is a free llms.txt generator.
Quotable facts. Put specialty, providers, locations, hours, insurance accepted and new-patient status in plain text on the site, in a form an engine can quote. Google notes that structured data should match visible text.
Consistency across sources. If the cited-sources table shows directories and review sites, align the clinic's name, address, phone and provider list across them and the Google Business Profile.
Reputation. Recent reviews and owner replies on the platforms engines cite.
The AI visibility guide explains how assistants find and name clinics in more detail.
Doing this across many clients
For agencies, the cost is review time, not running prompts. Standardize the panel template per specialty, fill in each client's city and specialty automatically, and review only exceptions: clients whose mention rate fell across two periods, or whose answers contained a factual error. If you would rather not build the pipeline, ClinicStats Live Ops runs 25 prompts across ChatGPT, Perplexity and Google AI Mode weekly for each monitored clinic site; the details are on the agencies page.