The most common way a marketing number fails is not that it is wrong. It is that nobody in the room can say what it is a proportion of. This guide is about making our numbers survive that question.
It is for anyone who presents our measurement — to leadership, to a client, or to someone who does this for a living and is not inclined to be generous. You should not need to read a spec to use it.
Two questions every number must answer
Both are recorded in the metric registry, so neither is a judgement call in the moment.
What is it over? Every rate has a denominator and ours are written down. "Sixty-one per cent" means nothing until you can say sixty-one per cent of what — of tracked answers in a window, of prompt-engine cells, of classified citations. A rate without a stated denominator is the commonest way a number misleads, and it usually misleads its own author first.
What is it NOT valid for? Every metric has uses it does not support, and ours are recorded as a required field — a blank would mean nobody had thought about it. Share of Answer is not valid across categories, because prompt universes differ. Retrievability is not a content-quality judgement. Our accuracy work is not a causal claim about revenue. Saying so before you are asked is what separates a measurement from a sales figure.
Disclosure and evidence are different things
This is the distinction that trips people up, and it is worth ten minutes.
Evidence class answers how strongly do we know this — observed, reported, derived, inferred.
Disclosure answers how much have we published about how it is computed — documented, measured, inputs named, modeled, proprietary.
They are independent. A metric can be measured on evidence and proprietary on disclosure at the same time. You can know something precisely and have published nothing about the calculation. You can also publish a full formula for a number resting on a weak inference.
When someone asks how confident you are, they are usually asking both questions at once, and the honest answer separates them. We measure it directly and we have published the formula is a different sentence from we measure it directly and the weighting is proprietary. Both are defensible. Conflating them is how people overclaim by accident.
The label that does the most work
Inputs named is where most composite scores in this category actually sit: the contributing factors are listed, the weighting is not.
That makes a metric legible but not auditable. A reader understands what goes in. They cannot reproduce the output, cannot tell whether a five-point move came from sentiment or position, and cannot detect a weighting change between releases.
It is a real advantage over publishing nothing. It is not defensibility, and one quantitative person in the room will find that gap immediately. Our own composite index sits there today, and the registry says so.
Why you cannot compare our numbers to a competitor's
There is no industry-standard visibility score. Every vendor defines it differently. In a same-domain, same-window test, one tool reported 1.9 times more citations than another for the same brand — not because either was wrong, but because each counted differently.
The mechanism is not mysterious, and at least one vendor documents it: their headline score is the most recent run, while their mention rate is aggregated across every run in the window. Those answer different questions. Over thirty days they will diverge, and the divergence is not an error.
What to say: these scores are valid for trend over time within one instrument. They are not valid for cross-vendor comparison, and anyone presenting them that way is comparing two different measurements that share a name.
The trap inside citation data
A brand can be cited as a source and never named in the answer. The engine uses the page, extracts what it needs, and answers without attribution. Industry research puts this at 62% of all AI citations.
The consequence people miss: citation share and mention rate can move in opposite directions. A brand watching only citations can believe it is winning while being invisible in the answers.
So the ratio between them is itself a diagnostic. High citations with low mentions means you are feeding the answer and getting no credit — a specific, fixable content-structure problem, not a visibility problem.
Why one reading is never enough
Answer engines are non-deterministic. The same prompt, the same engine, the same day, produces different answers. A single reading is not a weak measurement — it is not a measurement.
We run each prompt-engine pair repeatedly, report the frequency with a confidence interval, and treat the series rather than any single point as the thing to interpret.
When asked why the number moved: ask first whether it moved beyond the sampling interval. Movement inside the interval is noise, and reporting it as change is how a measurement programme loses credibility in its second month.
Two things we deliberately refuse
Both refusals are more defensible than the alternatives, and both will be questioned.
Share of voice. Uninterpretable when computed from retrieval. Say share of retrieved conversation, or omit it.
Prompt volume. Every vendor models it. No engine publishes it and no vendor discloses their model. It is the least verifiable number in the category and among the most used to justify budget. We do not produce our own, and saying why is a stronger position than having one.
The number that needs no assumptions
When a room gets sceptical, anchor on the metric that requires no weighting at all.
Our headline Share of Answer is a commercially weighted proportion of answers that recommend a product. Published formula, stated denominator, no pillar weights. The composite index sits beside it and is explicitly the contestable one.
That pairing is the posture: an assumption-free anchor, reported next to the richer number whose assumptions are published. A challenge to the index does not take the anchor down with it.
What to say when challenged
Four sentences that hold up, roughly in order of how often you will need them.
That is a directional index, not a measurement. It is valid for trend within one instrument and not for cross-vendor comparison. Here is the raw number, with its formula.
The inputs are confirmed; the weighting is proprietary. We use it as a directional signal, not a precision measurement.
That movement is inside the sampling interval, so we are not treating it as a change yet.
We do not report that, because it would imply a measurement we cannot stand behind.
None of those is a retreat. Each positions the speaker as the most rigorous person in the room — the only durable place to stand when the category has no agreed standard.
The one habit that matters
If you take a single thing from this: state the denominator and the invalid uses before anyone asks.
A number volunteered with its limits reads as confidence. The same number defended after a challenge reads as backpedalling — even when the words are identical.