Does ChatGPT Mention Your Practice? The Monthly 30-Prompt Panel Rate

Why there is no position to check

There is no position because there is no list. An AI assistant writes a fresh answer each time it is asked, and the practices it names can change between two sessions that typed the same words. No engine publishes where a practice stands for a question.

Google's guide to its own generative features says that no third-party tool has access to its internal ranking or AI systems 1Ref 1Google Search Central (2026).Google's Guide to Optimizing for Generative AI Features on Google Search.Google's statement that no third-party tool has access to its internal ranking or AI systems, that visibility in generative features is measured with the Search Console generative AI report, and that Business Profile details can surface in AI responses., and the help page for Bing's report on AI citations states that the report does not measure rankings 2Ref 2Microsoft Bing Webmaster Tools (2026).AI Performance - Bing Webmaster Tools.What Bing's AI Performance report counts (total citations, cited pages, grouped grounding queries rather than full prompts) and does not (rankings), and that a citation-trend change cannot be attributed to any single content change or model update.. That leaves one reading that means anything: a rate. Ask the same question enough times, count how often the answer names the practice, and the session-to-session noise averages down into a number that can be set against last month's. Everything else about the prompt panel exists to keep that number comparable from one month to the next.

Wording is the first thing to hold still. A comparison in npj Digital Medicine put 150 health questions to seven language models. The models were sensitive to how each question was framed, and by the paper's own account a model's accuracy can drop from 90 percent to 75 percent or lower when the prompt changes 3Ref 3Marcos Fernández-Pichel, Juan C. Pichel and David E. Losada (2025).Evaluating search engines and large language models for answering health questions.That language-model answers are sensitive to prompt wording, with the paper's own caution that a model's accuracy can drop from 90 to 75 percent or lower on a change of prompt, which is why the panel's wording is frozen.. A panel that rewrites its prompts each month is measuring its own edits.

The second thing to hold still is the asker. A signed-in assistant may carry what it learned in earlier chats. A logged-out window at the practice's own address is the closest a practice can get to the person it is trying to measure: someone in town who has never heard of it.

The 30 prompts, built from what patients ask before they book

The panel is thirty prompts in five groups, six each: fit, fee, plans, location and booking. They are written the way a patient types before booking, filled with the practice's own town, plans and service word, written once from the questions the front desk hears most, and then left alone. Real prompts are short. When 300 laypeople wrote their own health questions to three chatbots for a 2025 study, the median prompt ran seven words 4Ref 4Felix G. Rebitschek, Alessandra Carella, Silja Kohlrausch-Pazin, Michael Zitzmann, Anke Steckelberg and Christoph Wilhelm (2025).Evaluating evidence-based health information from generative AI using a cross-sectional study with laypeople seeking screening information.That real people's prompts to chatbots are short, a median of seven words across 300 laypeople writing their own, so panel prompts are written short..

Write them as questions. Pew followed 900 adults through 68,879 Google searches. An AI summary appeared on 8 percent of one- or two-word searches and on 53 percent of searches of ten words or more 5Ref 5Athena Chapekis and Anna Lieb, Pew Research Center (2025).Google users are less likely to click on links when an AI summary appears in the results.That question-style searches of ten or more words produced an AI summary far more often than one- or two-word searches, and that a link inside a summary was clicked on 1 percent of visits, so being named is not being visited.. The question-shaped prompt is the one that gets an AI answer at all, so it is the one worth measuring.

The thirty below are templates. TOWN is the town a patient would type, PLAN is one of the three plans the front desk is asked about most, and where a prompt says practice, write the practice's own service word.

None of the thirty names the practice. A branded prompt (what are the practice's hours, does it take PLAN) tests whether an engine has the facts right, which is a different question with a different home: the quarterly audit, where every fact an engine states about the practice is checked against the profile.

When a fact behind a prompt changes, a fee, a plan, an opening, the prompt stays and the change is logged beside it. Gale keeps a provider's gale.care page and its openings current, and drafts its Q&A capsules for review, so the fact a prompt asks about exists in one readable place.

How a panel month is run

Run the panel on the same day each month, from a logged-out browser window at the practice's address, on each engine the practice cares about: ChatGPT, Copilot, Perplexity and Google AI Mode. Each prompt goes in three times per engine, each time in a fresh window, and each answer is logged as mentioned or not mentioned before the next prompt is typed. Across four engines that is 360 answers a month.

1. Open a private or logged-out window. Sign out of the assistant and close any window that is signed in. 2. Type the prompt exactly as the sheet has it, TOWN and PLAN filled in. No follow-up questions; the first answer is the one that counts. 3. Read the answer for the practice's name, its street address or its website address. Any of the three is a mention; a near miss is logged as not mentioned, with a note. 4. Save the answer: a screenshot, or the copied text pasted into the log, with the engine, the date and the prompt number. 5. Close the window and open a fresh one before the same prompt's next run. 6. Move to the next engine and start again at prompt 1.

Where an answer names the practice, save the sentence it attached to the name. The quarterly audit reads those sentences against the profile, where a wrong fee quoted back by an engine is caught before a patient repeats it.

Three runs is a floor. Five runs on one engine and three on another make the engines incomparable. Pick three and keep it.

Report the month as a rate, and change nothing on one swing

The month's number is mentions divided by answers, per engine, written with the number of runs beside it. "ChatGPT 14 of 90, three runs" says more than any percentage on its own, because the reader can see how many draws it rests on. It is a rate, not a position, and it is the only form in which two months of this can be compared.

Then leave the site alone for a month. Google's starter guide says some changes take effect in hours and others take months, and suggests waiting a few weeks before judging whether a change did anything in its results 6Ref 6Google (2026).Search Engine Optimization (SEO) Starter Guide.Google's own timeline: some changes show in hours, others take months, and its suggestion to wait a few weeks before judging a change.. The help page for Bing's AI citation report goes further for the AI surfaces: a change in the citation trend cannot be pinned on any single content change or model update 2Ref 2Microsoft Bing Webmaster Tools (2026).AI Performance - Bing Webmaster Tools.What Bing's AI Performance report counts (total citations, cited pages, grouped grounding queries rather than full prompts) and does not (rankings), and that a citation-trend change cannot be attributed to any single content change or model update.. One month up or down, on one engine, is a reading to write down and not a reason to rewrite the hours page.

What a low rate does say is narrower than it feels. Before spending on content marketing, check the plainer cause. Does the fact the prompt asks about exist on a page an engine can read, in the words a patient would use? A prompt about Saturday hours can only be answered from a page that states Saturday hours. Most of the movement a small practice can cause is in that kind of fact, on its own site and on its Business Profile, which Google's guide says can surface in AI responses 1Ref 1Google Search Central (2026).Google's Guide to Optimizing for Generative AI Features on Google Search.Google's statement that no third-party tool has access to its internal ranking or AI systems, that visibility in generative features is measured with the Search Console generative AI report, and that Business Profile details can surface in AI responses..

Rows in the log are never edited after the run; a correction is a new row with a note. And the prompt sheet is saved read-only, so a year from now the series can be trusted by someone who was not there when it started.

The three instruments, and what each one cannot tell you

The panel rate is one of three instruments, and each one counts something the other two cannot. The panel counts whether an engine names the practice for a fixed set of questions. Search Console counts how often the practice's pages were shown inside Google's AI features, whatever the question. The intake tally counts people, the patients who say an assistant sent them. Read together, they tell a practice whether a change in one of them is real.

The Search Console reading is the second instrument. Google announced its generative AI performance reports on June 3, 2026 and had rolled them out to every site worldwide by August 31, 2026. They show impressions, pages, countries, devices and dates, at hourly to monthly granularity, for AI Overviews and AI Mode, and the same impressions stay inside the overall Performance report 7Ref 7Hillel Maoz and Moshe Samet, Google Search Console (2026).Introducing Search Generative AI performance reports in Search Console.What the Search Console generative AI performance report shows (impressions, pages, countries, devices, dates at hourly to monthly granularity for AI Overviews and AI Mode), its June 3, 2026 announcement and August 31, 2026 worldwide rollout, and that the data stays inside the overall Performance report.. There are no queries and no clicks in the AI view. Microsoft's report for Copilot is the same shape: it counts citations and cited pages, its grounding queries are grouped phrases and not the prompts people typed, and its help page says a trend change cannot be attributed to any single change 2Ref 2Microsoft Bing Webmaster Tools (2026).AI Performance - Bing Webmaster Tools.What Bing's AI Performance report counts (total citations, cited pages, grouped grounding queries rather than full prompts) and does not (rankings), and that a citation-trend change cannot be attributed to any single content change or model update..

The third instrument is the only one that counts patients, and it is the cheapest. One line on the new-patient form, how did you find us, with an AI assistant as one of the choices, and a monthly count of that answer is the AI referral tally. It is pixel-free measurement: the patient says so on a form, and nothing on the site watches them. On a provider site that matters, because the tracking-pixel problem turns a script that watches visitors into a question about the practice's privacy posture, while a question on a form is a question for the front desk.

Being named is a long way from being booked. In Pew's data, a link inside an AI summary was clicked on 1 percent of the visits where a summary appeared 5Ref 5Athena Chapekis and Anna Lieb, Pew Research Center (2025).Google users are less likely to click on links when an AI summary appears in the results.That question-style searches of ten or more words produced an AI summary far more often than one- or two-word searches, and that a link inside a summary was clicked on 1 percent of visits, so being named is not being visited.. The panel can be up and the tally flat, and that is information: the engines know the practice and the patients have not moved. When the three move together, over two or three months, something real changed.

Start this month, in an hour

The first panel can be run this afternoon with a spreadsheet and a browser, and the two instruments that take longer, Search Console verification and a line on the intake form, can be started in the same hour. It needs no vendor, only the same sheet, the same day of the month, and a log that is never edited after the fact.

1. Make the sheet: thirty prompts in five groups, TOWN, PLAN and the service word filled in. Save it read-only and date it. 2. Make the log: one row per engine per prompt per run, with the date, mentioned or not, and a note column. 3. Run ChatGPT today and the other three engines this week, all from the practice, all logged out. 4. Verify the site in Search Console if it is not already, find the generative AI performance report, and write this month's AI impressions on the log's summary line. 5. Add the question of how the patient found the practice to the new-patient form, with an AI assistant as one of the choices. 6. Put the three readings on one line per month: panel rate per engine with the run count, AI impressions, and the tally. 7. Calendar the quarterly audit for the third month: branded prompts, and every fact an engine stated checked against the profile.

A year of that is twelve lines. They will not say why an engine names the practice, and no report from any engine says that either. But they will say whether it does, how often, on which engine, and whether the patients followed.