What Is Prompt Tracking?

Prompt tracking is the routine of running a fixed set of patient questions through AI tools on a schedule and recording what each answer says. For a health practice, it is the way to find out whether AI tools name you, cite your website, or send patients to competitors. It replaces guesses and screenshots with a record that builds over time.

How prompt tracking works

A prompt is the question or request a person types into an AI tool. Prompt tracking treats a list of those questions like a ranking report. You choose the questions once, keep the wording fixed, run them on a schedule, and write down what comes back.

There are four moving parts.

• The question list. Write them from real patient wording, such as intake forms, front desk calls and your own reviews. A tool's suggested prompts are often guesses at what patients ask, not data. Sort them into three families and keep the families fixed: questions that name your practice, questions that name only a need and a city, and questions that describe a problem in a patient's own words. Most should not name you, since that is how new patients ask, and the branded and unbranded questions behave differently.
• The tools. ChatGPT, Google AI Mode, Google AI Overviews, Perplexity and Claude are the common ones. Each can answer differently.
• The repeats. Run each question several times per tool, because answers can differ from run to run. One run is an anecdote.
• The log. For each run, record the date, tool, mode, exact question, whether you were named, whether your site was cited, and which competitors appeared. Add the cited web addresses in their own column, because being cited and being recommended in the written answer are different results.

Mode matters. Some tools answer from what the model learned in training, and some search the web live. An answer from one is not a prediction of the other. For Google's AI Overviews and AI Mode, Google's documentation says a page must be indexed and eligible to show with a snippet. That is a different path from a model answering from memory. Write the mode down every time.

Personalization is the other trap. A logged-in account can shape its answers around what the tool already knows about you. For a cleaner read, use a ChatGPT temporary chat, a new one for each run, and log that you did.

The log feeds three monthly numbers. AI visibility is how often you appear at all.

AI share of voice compares your mentions with competitors.

Citation share, covered under citation share, tracks how often your pages are the sources cited. Prompt tracking is the raw data under all three.

A practical size for a first list is one service at a time. Start with your top three services in your top city, which gives you around ten questions and a log you can finish in half an hour. Grow toward 20 to 30 questions once the routine feels easy.

Three habits make the numbers believable. First, label a few questions as direct (about a page you just changed), a few as related (a nearby topic) and a few as control (a topic you did not touch). If all three groups move together, suspect a change in the tool and not your edit, a point covered under AI answer variability. Second, keep a dated list of your own site changes next to the log. Third, test with invented patients only. Do not paste real patient information into an AI tool.

Give a fix several weeks before judging it. That timing is my practical judgment and not a published rule.

A related concept is query fan-out, where an AI tool splits one patient question into several smaller searches. See query fan-out for the detail. It is one reason a small change in wording can change the answer, and one reason to keep your wording fixed.

Example: prompt tracking at a health practice

Take a dental practice that wants to know whether it appears for "best cosmetic dentist in Plano." It writes 25 questions like that, covering veneers, implants, whitening and emergency visits. This example is illustrative, not a transcript of a real practice.

Every week, someone on the front desk runs each question three times in a ChatGPT temporary chat. For each run, she records whether the practice was named, whether its website was linked and which competitors appeared. After two months, the sheet shows the practice appears for implants in most runs, appears for whitening in some, and is missing for veneers. That tells the practice exactly where to work. It writes a stronger veneers page with real photos, pricing and provider credentials, then watches that row.

Now a piece of real data that shows why the "mode" note matters. Health Biz Scale ran a study of 764 established US health practices between May 11 and August 20, 2026. We asked an AI model, GPT-4o-mini through OpenAI's API, 2,292 patient-style questions. The model answered from memory with no web search, one check per practice, test runs excluded. Across those answers, the model named 3,554 practice websites, and 1,153 of them, 32.4%, did not load. Only 6 of the 764 practices, 0.8%, were named. The ChatGPT app searches the web live and can answer differently.

Read that carefully. The finding describes one model, answering from memory, on one set of questions. It is not a prediction of what the ChatGPT app shows a patient today. That is the whole reason to log the tool and mode with every run. A practice that tested only one setup, and treated the result as universal, would draw the wrong conclusion in either direction.

The practical lesson is simple. Track the setups a real patient is likely to use, keep the wording fixed, and report the result as a rate across runs, such as "named in two of three runs," and not as a yes or no.

How to check yours in 60 seconds

1. Open a ChatGPT temporary chat.
2. Paste one real patient question, such as "best functional medicine doctor in [your city]".
3. Write down whether you were named, whether your site was linked and who else appeared.
4. Repeat once more in a new temporary chat. If the answers differ, you have seen the reason to track.
5. Save both answers with the date and the exact wording you used.
6. Do the same for one more question, so you have two data points to start your sheet.

What to do about it

1. Start with about 10 questions in patient language and grow toward 30, covering each service and each city you serve.
2. Include a few that name your practice directly and many that do not.
3. Keep the wording fixed. Changing it breaks comparisons.
4. Run each question three times per tool, every week or every other week, in a temporary chat. If that is too heavy, ten to fifteen questions run monthly is a realistic start.
5. Log date, tool, mode, question, whether you were named, whether you were cited and which competitors appeared.
6. Include the hard questions, since those show where the work is.
7. Give one person ownership of the log, and put the run on their calendar at a fixed time.
8. Summarize monthly as a rate, such as "named in 14 of 25 questions across three runs," and pick one page to improve based on the weakest row.
9. Rerun the same list after the improvement ships, so you can see whether the row moved. Keep the old results and your list of site edits for comparison.