Short answer
Write 20 questions a patient would really type, half about your category and half about you by name. Ask all of them in ChatGPT, Perplexity, Gemini and Claude while logged out. Record five things per answer: were you named, was a page of yours cited, was anything wrong, who was named instead, and which sources the engine preferred.
Key takeaways
- Logged out matters. Your own chat history will name you and tell you nothing.
- Fix the wording of the 20 questions once and never edit it, or month two is not comparable.
- Record what was wrong about you, not only whether you appeared. Wrong details are the finding that costs bookings.
- One run is a snapshot. The trend over three to six months is the measurement.
We ran this test on our own company before we ever sold it to anyone. It came back named in 6 of 25 questions, and all six already contained the company name. In the 19 questions where somebody was choosing who to hire, we were named zero times. That result is what the rest of this article is built on.
Step 1: write the 20 questions, and do not be clever
Ten should be buying questions with no brand name in them, phrased the way a patient speaks: who to see, what it costs, whether something is worth it, best option near a city. Ten should be about you by name: what the practice is, whether it is any good, who the clinician is, what it charges. The split matters, because the two halves fail for different reasons.
Take the wording from real sources: what patients ask at the front desk, what comes in through the contact form, what your Search Console shows people typing. Invented phrasing tests a question nobody asks.
Step 2: ask all four assistants, logged out
Use a private window with no account signed in, in ChatGPT, Perplexity, Gemini and Claude. A logged-in session carries your history and your location, and it will flatter you. Ask each question once, let the answer finish, and paste the whole answer plus its sources into a sheet rather than summarising it as you go.
Step 3: score five things per answer
| Column | What goes in it | Why it matters |
|---|---|---|
| Named | Yes or no | The headline number, and the one that moves slowest |
| Cited | Which URL of yours appeared, if any | Tells you which page is doing the work |
| Wrong | Any incorrect detail about you | Wrong hours or phone cost bookings this week |
| Named instead | Every competitor mentioned | Shows who the engine trusts and why |
| Preferred sources | Directories, review sites, press, forums | Tells you where to spend the next month |
The "wrong" column is the one people skip and the one that pays first. In rater8’s April 2026 survey, 66% of patients who used AI to research a provider met incorrect provider information, and 60% trusted it without checking. Anything wrong in that column is a live problem, not a ranking problem.
Step 4: read the result honestly
Split the score. Being named in brand questions but not in buying questions is the normal pattern, and it means the model knows you exist but does not consider you an option. Being absent from both usually means an access or indexation problem, not a content problem, and it is fixed at the server rather than by writing.
- Named in buying questions, cited page on your site: keep going, refresh that page every 90 days.
- Named in buying questions, no page of yours cited: the engines are quoting directories and press about you. Get your own page to answer that exact question.
- Named only in brand questions: the entity is known, the recommendation is not earned yet. Directories, reviews and third-party mentions move this.
- Absent everywhere: check crawler access and indexation before writing a word.
Step 5: repeat monthly, judge quarterly
Run the same 20 questions on the same day each month and keep every raw answer. Never edit a question once it is in the list: retire it and add a new one at the end instead. Vendor research reports that 40 to 60% of cited sources change month to month, so a single run tells you almost nothing and the trend tells you everything.
If you want this done for you rather than run by hand, that is what our audit does: the same fixed prompts across the four assistants, plus the entity and crawler checks, with a 90-day plan ordered by what moves the answer soonest. Either way the method is the same, and you can run it yourself this afternoon.
Questions people ask
Why does the test have to be run logged out?
A signed-in session carries your chat history, saved memories and location, so the assistant already knows your practice and repeats it back. That tells you nothing about what a stranger sees. Use a private window with no account for every question.
How many questions do I really need?
Twenty is enough to see a pattern and short enough to actually repeat monthly. Split them evenly between buying questions and questions about you by name. Larger lists are better only if you will still run them in month six.
Can I just ask the assistant whether it recommends me?
No. Asking "do you recommend my practice" contains your name, which pulls you into the answer and proves nothing. The whole finding sits in the questions where nobody has mentioned you and the engine has to choose.
What counts as a good score?
There is no absolute threshold. What matters is the trend across three to six months on a fixed question list, and the ratio between buying questions and brand questions. Named in brand questions only is the usual starting point for a practice.
How long should I wait before re-testing after making changes?
Two to four weeks for crawler access and indexation fixes, which move fastest. Six weeks or more for directory and entity corrections. Content changes take months, so re-test on the monthly cadence rather than the day after publishing.
Sources
- rater8, 2026 Patient Choice Report (April 2026, n=992 US adults, vendor survey): 66% of AI users met incorrect provider information; 60% trusted it without verifying.
- Rock Health, 2025 Consumer Adoption Survey (December 2025, 8,000 US adults, vendor survey): 32% of US adults have used an AI chatbot for health information, double the year before.
- Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024 (peer reviewed): What raises citation visibility, and that keyword stuffing does not.
Published . Last reviewed .
Want this in your business?
Book a call and we will scope exactly which part of your operation an agent should take first.
Book a call with the team