AI visibility for a dental practice is one number: the share of the questions your patients ask where a model names your practice and recommends it. You can measure that yourself in an afternoon: twenty fixed questions, four engines, three runs each, a spreadsheet and a scoring rule. No vendor, no dashboard, no argument about what counts as a win.
What AI Visibility Actually Means
Start with the metric and the metric is unremarkable: mention rate. Take the runs in which your practice is named, divide by the runs you made. A run is one answer, from one engine, to one question, on one occasion. Nothing else enters the calculation, which is exactly why it survives a busy week.
answers name your practice
answers name your practice
Position, the thing most dashboards sell, does not survive contact with a paragraph answer. A model that writes six sentences about dentists in your city has not ranked you sixth. It has mentioned you, or it has not. There is no first place in a paragraph, and any tool promising you a position is selling a metric your patients never see.
The Three States Of a Mention
Not every mention is the same, and one number hides the difference that matters:
- Named and recommended. The practice appears with an endorsement attached, described as the best, the top rated, or the one to choose for a particular need.
- Named but not recommended. The practice turns up as one option in a list or a comparison, with no endorsement attached to it.
- Absent. The practice is not in the answer at all, or appears only in passing, such as a hospital dental clinic or a practice in another city.
Only the first state is a result. The second is an opening: you are already in the set, and one missing fact may be the thing that would tip the recommendation. The third is a gap that needs a diagnosis. Counted as a single number, you cannot tell which of the three you are in.
What a Run Is, and What a Run Is Not
One answer is an observation, not a verdict. The same question asked twice can return different names, because models vary by wording, version, location, time of day and whether the account behind the session carries any history. A single miss tells you almost nothing. Three runs of the same fixed question tell you something real.
drsmiles.ai scores every run individually and keeps the ones that did not help. A scorecard built from the flattering runs is not a measurement, it is a marketing artefact, and it collapses the first time a patient notices the difference.
The Question Set: 20 Questions That Find Money
Twenty questions, six buckets, and never change them. Each bucket maps to a patient who either needs care now or is deciding whether to trust you with care later. The allocation below is a starting point rather than a law, but set it once and then freeze the list for at least a quarter.
Swipe the table sideways to see every column.
| Bucket | What The Patient Is Really Asking | Share Of The 20 | Example Prompt |
|---|---|---|---|
| Emergency | I am in pain now and need someone who will see me | 4 | Urgent tooth pain tonight in [CITY] |
| Treatment | Who actually does this procedure well | 3 | Who does tooth replacement with implants in [CITY] |
| Price | What will this cost me before I commit | 3 | What a first visit and exam costs in [CITY] |
| Availability | Can I actually get an appointment this week | 3 | Saturday and evening appointments in [CITY] |
| Insurance | Will you take my plan and my budget | 4 | Dentists in [CITY] taking my insurance plan |
| Neighbourhood | Who is trusted by people like me, close to me | 3 | Best dental practice near [NEIGHBOURHOOD] in [CITY] |
Freeze the set because a moving question set produces a moving number, and then you cannot tell whether the practice moved or the questionnaire did. If you must add a question, add it as number 21, keep the original twenty in the report and report the two series separately for a quarter. Our full list of patient questions for ChatGPT and the other engines shows how the wording is written so that you can copy it exactly rather than approximate it.
How To Run and Capture The Answers
Write the prompt once, as a template, and change nothing between runs.
I'm in [CITY] and I need [NEED].
Which dental practices should I consider? Name them and say why.
Fill [CITY] with one city, not a metro area and not a neighbourhood. Fill [NEED] with the need in the patient's own words. Leave the practice name out of this prompt entirely and add it only in the reverse test below, because a prompt containing your name measures your own recognition rather than the set.
Run each of the twenty questions three times against each of ChatGPT, Gemini, Perplexity and Claude. Three is the smallest number of runs that tells you anything; one run can be an artefact of a single sample. Where the tool allows it, use a signed-out session, because a personalised history changes what the model assumes about you.
Capture every answer verbatim before you interpret any of it. Record six fields per run: question number, exact question text, engine, run number, timestamp and the full answer. If you summarise instead of storing the text, you cannot audit a disagreement later, and the whole point of the record is the afternoon you need to check what was actually said.
A run is noise when the model refuses the question, asks for your address before answering, answers about a different city, or returns a generic paragraph that names no practice. Log it as noise and keep it. Never delete it, and never re-run a question until you like the answer, because a re-run replaces a data point with a sample you chose.
How To Score The Answers
Fix the rule in writing before you read the answers, because the temptation to redefine "recommended" is strongest once you know what you would like to claim.
Swipe the table sideways to see every column.
| State | What It Looks Like In The Answer | Score | What It Tells You |
|---|---|---|---|
| Named and recommended | Your practice leads the answer, or is singled out with the reason you would want to be singled out | 2 | You own this answer for this run |
| Named, not recommended | Your practice appears among the options with no endorsement attached | 1 | You are in the set and a fact is missing |
| Absent | Your practice does not appear, or is named only as an aside | 0 | You are outside the set and need a diagnosis |
When you are named, record where you appeared: first, middle or last in the list. That is not a ranking, but it tells you how much of the answer you own. A first-listed mention with a reason attached is worth more than a name buried at the end of a long list, because readers read the first line and stop.
Total the scores and divide by the maximum possible score for that engine and question set. Report one figure per engine as well as the combined one, because a practice can be strong in Perplexity and invisible in Gemini, and the average hides exactly that.
A Worked Example
Five runs from one practice in one month, with the next move written against each one:
Swipe the table sideways to see every column.
| Question | Engine | What The Answer Said | State | Position | Next Move |
|---|---|---|---|---|---|
| Urgent tooth pain tonight in [CITY] | Perplexity | Named two practices, one of them yours, with no reasons given | Named, not recommended | Second of two | Your page shows no hours or phone details an engine can quote |
| Who does tooth replacement with implants in [CITY] | Gemini | Named you first and explained why, pointing at a treatment page | Named and recommended | First of three | Hold, then re-check next month |
| What a first visit and exam costs in [CITY] | ChatGPT | Gave a range and named nobody | Absent | Not applicable | Publish a page stating what a first visit includes |
| Dentists in [CITY] taking my insurance plan | Claude | Named you third, with no reason attached | Named, not recommended | Third of four | Make the insurance answer explicit on your own page |
| Best dental practice near [NEIGHBOURHOOD] in [CITY] | Gemini | Named two competitors, both of them with a page for that area | Absent | Not applicable | Build a page for that neighbourhood |
This score is a sample, not a census. Three runs per engine is enough to see a gap and not enough to claim a stable average. If a number is going to decide whether you spend money, run it again before you act on it.
What To Fix First, In Order
Work the list in this order, because each fix makes the next one more effective. The sequence is ordered by how fast a fix moves the number, not by how much effort it costs.
Why This Order
- Entity facts. Name, address, phone, services, hours and practitioner names, stated plainly on one page and marked up as a Dentist entity and an organisation so a machine can read who you are.
- Directory consistency. The same facts, spelled the same way, on every profile that lists you. Google's Business Profile help sets out what belongs on the profile most engines read first.
- Missing pages. One page per service you want to be found for, one per neighbourhood you serve, one per question in your price bucket. A question with no page behind it is a question you cannot win.
- Content depth. Detail, specific answers, review proof and markup that makes a page quotable. The growth work belongs here, and it comes last because a longer page on a broken foundation still reads as a longer page on a broken foundation.
Entity facts come first because everything else depends on them. A page an engine cannot identify as a dental practice cannot be quoted as one. Directory consistency comes second because directory profiles are what engines most often name, and they are the cheapest item on this list to change. Missing pages come third because they are the only fixes that create something to be found. Depth comes last because a longer page on a wrong foundation still reads as a longer page on a wrong foundation.
When you would rather start with a second pair of eyes, the visibility snapshot shows where you stand in a couple of minutes, and the pricing and terms FAQ explains what a full engagement includes.
How Often To Measure
Three cadences, three purposes.
- Weekly, for tracking. Five questions on one engine, one run each. It catches sudden losses, which are usually a directory edit, a review surge on a competitor or a changed opening time in a listing.
- Monthly, for decisions. The full twenty questions across all four engines. This is the run that tells you what to fix next.
- Quarterly, for strategy. The full set plus a comparison against the competitors who take your answers. Our guide to benchmark scores covers how to compare against your own city rather than an invented national average.
Cadence matters more than a longer list. Twenty questions run consistently for a year tell you more than two hundred questions run once. The second number is a snapshot; the first is a trend, and trends are what you plan against.
For the background on why mentions matter more than position, start with what AI visibility is and where these answers come from. For the monthly run, use the twenty above. For the full catalogue with the exact wording, see the forty patient questions we ask every week.