"Brand visibility in AI search" is now a line item in enterprise plans, and almost everywhere we look it is being tracked as a single number — a percentage of prompts where the brand appears, sometimes dressed up as a visibility score.

That number moves. It just does not tell you why, and it cannot tell you what to do, because it is three unrelated measurements added together. We have watched programs run for two quarters against a rising score while the thing that actually determines revenue stayed flat.

This piece separates the three, gives the measurement for each, and states what we have seen move each one.

Measurement 1 — Presence

The question: when a buyer asks a question in your category, does your brand appear in the answer at all?

This is the measurement most tools report, and it is the least informative of the three, because it is binary and saturates early. Most established enterprises pass it for their obvious queries within weeks of paying any attention at all. A large company with a Wikipedia entry, trade press coverage, and a decade of history is generally present.

How we measure it: a fixed set of category questions — not brand-name questions, which are trivially passed and tell you nothing — run across ChatGPT, Claude, Gemini, and Perplexity, coded simply for whether the brand is named.

The distinction between category and brand-name prompts is where most measurement goes wrong. "What does Acme do" will name Acme. "Which vendors handle multi-currency invoice reconciliation for mid-market finance teams" is the question a buyer actually asks, and it is the one worth counting.

What moves it: corroboration breadth. Independent sources describing you, in enough places that retrieval finds something regardless of which way the query leans. This is largely a communications and industry-presence problem rather than a website problem, and it is slow — measured in quarters.

Where it misleads: a rising presence rate on category questions you chose yourself is easy to manufacture. Broaden the question set until it includes prompts you were always going to pass and the number goes up while nothing changed.

Measurement 2 — Description accuracy

The question: when your brand does appear, is what the system says about you the thing you would have said?

This is where we find the largest gap in large organizations, and it is the measurement almost nobody runs. Enterprises with long histories, acquired product lines, and material published across a decade are describing themselves in a dozen partially inconsistent ways, and the systems reading that material produce a reconciliation of it rather than a description of it.

Characteristic findings when we run this:

The last one is the most consequential and the easiest to misread as neutral. Hedging is what a system produces when its sources do not agree — including when the disagreeing sources are all yours.

How we measure it: for each answer where the brand appears, code the description against a short list of facts you consider load-bearing — primary category, primary customer, the two or three differentiators you would want stated. Score each as correct, outdated, absent, or wrong. This requires human coding; it does not automate well, and the fifty answers a quarter it takes are worth more than any dashboard.

What moves it: the conversion of positioning language into structured, checkable facts, applied to your highest-authority pages. This is the intervention with the shortest latency of the three — we have seen description accuracy shift within weeks, because it depends on what your own material says rather than on what anyone else does. It is also the one that stalls organizationally, since the language belongs to brand, the facts belong to product, and the approval belongs to legal. Our four-layer framework covers those stall points in detail.

Measurement 3 — Shortlist inclusion

The question: when the answer takes the form of a recommendation — a list of vendors to consider — are you on it?

This is the measurement that corresponds to revenue, and it is the one least correlated with the first. A brand can be present in most answers, described accurately, and still be absent from every recommendation, because being described and being recommended are produced by different parts of the process. We take that mechanism apart in how engines choose.

How we measure it: a separate question set, phrased the way a buyer near a decision phrases things — "best options for," "which vendors should I consider," "compare providers of." Code for inclusion, position within the list, and the reason given. The reason is the most useful field and the one most often not captured.

What moves it: comparative evidence. Specific, attributable claims about fit — conditions under which you are the right choice, stated with enough precision to be lifted. Systems assembling a shortlist need a reason to put each name there. Absent a specific reason attached to you, a generic-but-well-covered competitor gets the slot.

The uncomfortable finding: stating what you are not for tends to improve inclusion rather than harm it. A named boundary — "for teams above 200 seats," "not suited to single-jurisdiction operations" — gives a system the condition it needs to place you confidently. Unbounded claims are harder to act on than bounded ones.

Why the composite score misleads

The three measurements move at different speeds, respond to different interventions, and are owned by different functions:

Average them and you get a number that rises when the easy one rises. The common pattern: presence improves because coverage broadened, the composite goes up, leadership is satisfied, and inclusion — the one tied to pipeline — has not moved at all. Two quarters later the program is questioned on the grounds that the metric went up and nothing else did.

Report three numbers. They are not hard to explain, and each points at a different owner.

The minimum viable measurement

What we would consider the floor for an enterprise program:

  1. Two question sets. Twenty category questions and ten recommendation questions. Fixed. Changing them between runs destroys the comparison, and it is the most common error we see
  2. Four platforms. ChatGPT, Claude, Gemini, Perplexity. They disagree enough that one platform is not a sample
  3. Three codings per answer. Present or not; description correct, outdated, absent, or wrong; included in shortlist or not, with the stated reason captured verbatim
  4. Monthly, human-coded. Roughly a day of work per cycle at this size
  5. Raw text retained. The transcripts are more valuable than the scores — they tell you which sources are being drawn on, which is what you act on

Our sixty-day tracking write-up gives the full procedure with the numbers from an actual run, including how much week-to-week variance to expect before you start reading noise as movement.

Where this framework comes from

The three-way split is a measurement expression of the staged model in GEO: The New Enterprise Gateway in the AI Era — being discovered, being understood, being shortlisted. The book's argument is diagnostic: the three stages fail for different reasons and are fixed by different work. Ours is the operational consequence — if the stages are distinct, a single score that spans them cannot tell you which one you are failing.

The book also supplies the CEO self-audit checklist, which is a faster way to get an initial read than a full measurement cycle, and the 90-day pilot structure for the program that follows. Chapter-level notes in our summary, and the wider shelf in the reading list.

Book details

Related research

See also the six properties of sources that get named, what to check when a brand stops appearing, and the four-week executive reading path.