AI answers tend to cite sources that carry something specific: a number, a method, a date, a named place. Opinion gets rewritten and dropped. A dental practice already owns original data, from no-show rates to treatment acceptance, and publishing it properly is the cheapest citation strategy available to you.

What The Research Says About Citations

The clearest evidence comes from the Princeton GEO study by Aggarwal and colleagues, presented at KDD 2024 (the Princeton GEO study), which tested a range of tactics against how generative search engines choose sources. Two of its findings matter for a dental practice. Adding citations and adding statistics were the two most effective methods, while keyword stuffing actively reduced performance.

BEFORE
Opinion piece: a data page
–Opinion with no data
–Method not stated
–No date
–Nothing to cite
AFTER
Data piece: a data page
✓Your own numbers
✓Method stated
✓Dated
✓A page worth citing
Illustrative example.

That is a description of a study, not a promise about your answers. Models change, questions get personalised, and the same wording can return different sources on one afternoon and the next. What the work does establish is a direction of travel: engines prefer material they can quote, and an original number beats a borrowed opinion.

There is a duller reason numbers travel well. A figure is short, self-contained and easy to attribute. One statistic fits inside the sentence an engine writes; three paragraphs of context do not, which is why vague articles get paraphrased and dropped.

The Data a Dental Practice Already Has

You do not need a research budget. You need a spreadsheet export and a question worth answering. Most practices already hold six datasets that nobody has ever published.

Six Datasets You Can Export Today

  • No-show rate, counted separately for recall appointments and for emergency appointments
  • Recall reappointment rate, meaning booked hygiene visits that turned into treatment
  • Treatment acceptance by channel, so you can compare what arrived through your website, a referral and a phone call
  • Wait time for a new patient call, measured from the moment someone dials to the moment a person answers
  • Review themes, counted across your last hundred reviews, including the complaints
  • Emergency mix, the share of that week's appointments booked for pain

Price sensitivity is the seventh and the most revealing: how many people ask for a written quote and then do not book. Every one of these numbers describes your own practice, which is exactly why a competitor cannot publish it and an engine can quote it.

How To Run a Small Study

A study, even a small one, has five parts. Write them down before you look at the numbers, because a measure chosen after seeing the data is how small studies go wrong.

  1. The question, phrased so that a number answers it: "How many booked appointments were missed without notice?"
  2. The sample, named exactly: which appointments, which providers, which exclusions.
  3. The period, with a start and an end date, so somebody else can repeat the pull.
  4. The measure, defined in one sentence, including how the awkward cases are handled.
  5. The limitation, stated before publication rather than discovered after it.

A Worked Example

The table below illustrates the method with invented figures rather than data from our clients. It shows how a finding, its method and its limits belong in the same block.

Swipe the table sideways to see every column.

ElementWhat It SaysWhy It Is There
QuestionHow many booked appointments were missed without notice?One sentence, answerable with a single number
Sample150 appointments booked with two clinicians at one siteRemoves the reader's need to guess who was counted
Period1 to 28 June 2026Allows a repeat pull from the same practice management system
MeasureAppointments recorded as unattended, divided by appointments booked, excluding cancellations made more than 24 hours aheadShows the arithmetic and the exclusions in one place
Result18 of 150, which is 12 in every 100 booked appointmentsThe quotable sentence, with the denominator attached
LimitationOne practice, two clinicians, one urban site, one month, no general claim beyond itStops the finding travelling further than the data supports

Twelve in every 100 booked appointments is not a national figure and does not claim to be one. It describes a single practice over four weeks, and the limitation row says so out loud.

How To Present It so It Can Be Lifted

Engines lift sentences that stand on their own. The format that works is a headline finding, one number, one method line, one date and one table.

The Shape Of a Quotable Finding

Headline: "Twelve in every 100 booked appointments at our practice went unattended in June 2026." Number: 18 of 150. Method line: "Practice management system export, 1 to 28 June 2026, excluding cancellations made more than 24 hours ahead." Date: "Published 12 July 2026." Table: the six rows above, unchanged.

Then add the machine-readable layer, so a system can read the same facts you published in prose. The FAQPage schema type carries question and answer pairs, and the documentation for FAQ structured data explains what those pages expect to contain.

No format can promise a citation. This structure makes your material easy to quote; whether a particular answer uses it still depends on the wording of the question, the model and the moment the answer was generated.

Honesty Rules That Protect You

Four rules keep a small study defensible. They are the difference between a measurement and a marketing claim that happens to contain a number.

  • Do not drop the inconvenient months or the inconvenient patients to move the number
  • Do not present your own result as an industry average or a national picture
  • Do not invent a benchmark to compare against when no published benchmark exists
  • Label anything estimated as an estimate and say what it was estimated from
  • Keep patient names, case details and anything identifying out of the published version
  • Do not change the measure after seeing the result unless you rerun it from scratch

Two practical guards go with them. Keep the export that produced the number, so you can answer a question about the method without guessing. And write the limitation in the same paragraph as the finding, because a limit filed at the bottom of a page is a limit nobody reads.

Where To Publish It

Publish the original on your own site first. Your page is the source of truth, the version engines can quote, and the one your other channels point back to. Give it a URL that describes the subject, a publication date, the full table and a named author.

Then put the finding where engines already collect sources: trade press, practitioner forums, local press and your own profiles. A local paper running your number because you sent it the method is worth more than the same number sitting on your own site, because it becomes a citable source with a byline and a date attached. Google's guidance on Business Profile information covers keeping the practice details on those listings aligned with the study.

Give the study a permanent home inside a cluster instead of leaving it orphaned. how a topical cluster is built explains where a study belongs, and the refresh protocol for ageing pages is what you run when the numbers themselves go stale. For context on reading the visibility figures you already track, see what a good AI visibility score looks like.

The audit behind drsmiles.ai asks forty questions across four engines, so you can see which questions your practice is missing before deciding what to measure next. The pricing and terms FAQ sets out what that audit includes, and a sample report shows what comes back. If the work matters enough to hand over, the AEO service is where measuring, fixing and proving sit in one process.