METHOD

How the number is made.

Every report on this site comes from the same loop: a fixed question set, four engines, three runs per question, a union count, and a 30-day window. Here is the whole method, including the parts where it is weak.

THE LOOP, IN FOUR NUMBERS

Engines measured
4
Runs per question
3
Days in the window
30
Reported window
Rolling
Questions × engines × days. The arithmetic is the same for every client.

WHAT WE TRACK

Five kinds of question, in the words a patient uses.

Keywords are a search-engine unit. A patient asks a sentence, and the sentence carries the treatment, the place and sometimes the urgency. The tracked set is built from those sentences, then approved by the practice before anything is measured.

TypeExampleWhy it is tracked
Local recommendationBest dentist in Naples, FLTests whether the practice exists at city level
TreatmentBest All-on-4 dentist in NaplesThe question a patient asks once a treatment is chosen
ClinicianTop rated implant dentist near NaplesTests whether the named dentist is associated with the practice
PriceHow much do implants cost in Naples?High intent, and the answer often cites directories
ProblemWhere can I get emergency dental care today?Tests recency and hours, which AI answers get wrong often

HOW A MONTH RUNS

Same questions, same order, every month.

Day 1

Baseline

The agreed question set runs once across four engines, three runs per question. The union becomes the starting number.

Daily

Collection

The same questions run every day. Answers, citations and competitor mentions are stored raw, with the timestamp and model name.

Day 25

Report

We score mentions, share of voice, citation domains and factual accuracy. Two reviewers score accuracy; disagreements over 20 points are re-scored.

Day 30

Review

Thirty minutes on what moved, what did not, and which source to fix next. The question set is only changed at a review, never mid-month.

HOW THE NUMBERS ARE BUILT

Four numbers, four formulas. Check them on a call.

Mention rate

questions naming the practice ÷ questions tracked

Counted as a union across three runs, so one lucky answer cannot carry the month.

Share of voice

own mentions ÷ (own + three named competitors)

Competitors are fixed in the contract and can change once a quarter.

Citation share

own-domain citations ÷ all citations in the answers

Directory pages do not count as own citations, even when they quote the site.

Accuracy

correct fields ÷ 3 (treatment, clinician, location), scored 0 to 100

A practice can be named and still be described wrongly. This number catches that.

WHAT MOVES, AND WHAT WE WILL NOT PROMISE

Answers are volatile. The method has to survive that.

In a published sample of roughly 80,000 prompts per engine, between 40% and 60% of the domains cited in an answer differed a month later. One evening's screenshot is a photograph, not a trend. So the report shows a union over three runs and a rolling window, and every table carries the date it was produced.

  • We do not promise placement inside an answer. No agency can buy that slot.
  • We do not promise patient volume. We report the questions, the citations and the accuracy of the description.
  • We report four engines. Claude and Grok are observed, not measured, until their output is stable enough to bill against.
  • We score accuracy by hand, with two reviewers. A disagreement over 20 points triggers a re-score, not an average.

QUESTIONS ABOUT THE METHOD

The five that come up first.

Why three runs instead of one?

Because the same question asked twice in an evening can return a different order. Three runs turn a single answer into a small sample, and the union stops one outlier from becoming the headline.

Which engines are measured, and which are only observed?

ChatGPT, Perplexity, Gemini and Google AI Overviews are measured and reported every month. Claude and Grok are observed on the Dominator plan and are not part of the reported number yet.

Are API answers the same as what a patient sees?

Not always. API responses can differ from the consumer apps in wording and in the source list. We collect through APIs for repeatability, and we say so in the report rather than implying we sat in front of the app.

How volatile is this data?

In a published sample of roughly 80,000 prompts per engine, 40% to 60% of the domains cited in an answer differed a month later. That is why every table carries a window and a date, and why a single month is never sold as a trend.

What happens if the number goes down?

It is reported. A month where mentions fall is more informative than one where they rise, because it usually points at a source that changed. The report keeps the same question set so the fall can be traced rather than explained away.

START WITH A MEASUREMENT

See the method applied to your own practice.

The free snapshot runs five questions on two engines and returns the same fields a retainer reports: mention, description, citation. No call required.

Request the free snapshot