Executives do not need another marketing metric. They need this one translated into a frame they already own — competitive position, and where the gap is.
Put in the language of a board meeting: when patients and buyers ask AI about the services your organization provides, what proportion of those answers do you appear in — and who is appearing instead?
Three measures, never combined
The most common error in commercial AI visibility reporting is collapsing everything into a single "visibility score." Three things are happening in a generative answer, they carry different commercial meaning, and they must be reported separately.
| Measure | What it means | Why it matters |
|---|---|---|
| Presence | The organization is named anywhere in the answer. | Baseline participation. Includes neutral and negative mentions. |
| Recommendation | The organization appears in the recommended or shortlisted set — not in passing, contrast or caveat. | The only measure with direct acquisition value. |
| Citation | A domain the organization owns is used as a source for the answer. | Leading indicator. Citation share tends to precede recommendation share. |
An organization mentioned in the sentence "Memorial is another option in the area, though Regional Medical Center is generally considered stronger for complex cases" has presence and does not have recommendation. Reporting that as visibility is not measurement, it is flattery.
Presence Rate and Share of Discovery are different numbers
These get used interchangeably in the market and they are not interchangeable.
Presence Rate is absolute: the percentage of answers in which you appear. It ranges from 0 to 100% and it does not sum across organizations, because a single answer can name five of you.
Share of Discovery is relative: your slice of all competitive-set appearances. It sums to 100% across the set.
Both are useful and they answer different questions. Presence Rate tells you whether you are participating at all. Share of Discovery tells you your position against the competitors you actually have. A hospital can hold a 40% presence rate and still be third in its market, because the competitors are at 60% and 55%.
Volatility, and why it may be the most useful number in the report
Generative systems are non-deterministic. Ask the same question five times and you may get five different sets of organizations. Most reporting treats this as noise to be averaged away. It is signal.
Volatility(q) = distinct organization sets returned ÷ number of runs 1.0 = every run returned a different set — fully contested 0.2 at five runs = every run returned the same set — entrenched
A question where the model returns the same three hospitals every single time has a settled answer. Displacing an incumbent there is expensive and slow. A question where the answer scrambles run to run is unsettled — the model has no confident view, and comparatively small changes in the underlying source landscape can move it.
High-volatility questions are the movable ones. That is where the work should go first.
This converts the assessment from a scorecard into a prioritization tool. Two service lines with identical presence rates can have completely different prospects, and volatility is what separates them.
A worked example
One question, one platform, five runs. Question: best hospital for hip replacement in Broward County.
Client Health appears in runs 1, 2 and 4 — a presence rate of 60%. It sits in the recommended set in runs 1 and 4 — a recommendation rate of 40%. Its own domain is cited as a source in run 4 — a citation rate of 20%.
Across those same five runs, the competitive set accumulates eleven presence events in total: Client Health 3, Competitor A 5, Competitor B 2, Competitor C 1. Client Health's Share of Discovery is 3 ÷ 11 = 27%.
Four distinct organization sets appeared across the five runs, giving a volatility of 0.8. The answer is contested. This question is worth working on.
Note how much information the single 60% would have hidden.
Why one screenshot is not a measurement
The prevailing practice in this young market is to run a few prompts, screenshot the results, and present them as a visibility finding. It is persuasive in a pitch and it does not survive contact with a competent skeptic.
Because outputs vary across runs, sessions, accounts, geography and model versions, anyone who re-runs those prompts will get different answers. If the client's agency does that — and eventually one of them will — the finding collapses, and so does the credibility of everything presented alongside it.
A defensible measurement requires a minimum of five runs per question per platform, fresh unauthenticated sessions with personalization disabled, controlled and recorded geography, model versions stamped on every run, documented scoring rules that two analysts would apply identically, and results reported as ranges with the observed variance disclosed rather than smoothed away.
Reporting a range instead of a clean number feels like a weakness. It is the opposite. It is the single clearest signal that a real measurement happened, and it is the thing a sophisticated reader checks for.
The full specification is published
Every element above — question construction, sampling requirements, scoring rules with edge cases, the calculations, reporting standards and the stated limitations — is published as an open protocol, versioned, and free for any party to apply.
That is deliberate. A measurement standard only its author can run is not a standard, and a visibility figure that cannot be independently reproduced is an assertion rather than a measurement.
Related
Get your Share of Discovery measured
Twenty-five to fifty strategically selected questions, measured across four platforms, scored against your named competitive set and reported by service line, market and competitor.