How every number is calculated

Every figure Citerva displays is defined here. If you see one in the product that is not on this page, that is a bug.

Vocabulary

A run is one prompt asked of one model at one provider. A 20-prompt set across three providers is 60 runs.

A succeeded run is one that returned an answer we could store and parse.

The denominator rule

Every rate is computed over succeeded runs only, and the product always shows the denominator.

A run that timed out, was rate-limited, was declined by the provider, or was never attempted is excluded — not counted as "no mention".

This matters more than it sounds. If one assistant fails on 10 of 20 prompts and we counted those as zeroes, your visibility would appear to fall because of our infrastructure rather than anything happening in your market. Failures are shown separately as scan health, never folded into a visibility number.

Counting rule

A mention is counted once per run per entity.

If a model writes "Acme is great. Acme integrates well. We recommend Acme." that is one mention, not three. Otherwise a rambling answer would outscore a concise one and the metric would measure prose style rather than market presence.

The individual positions are still recorded and shown as evidence — they are just not the unit of the rate.

The metrics

Mention rate

answers that named you / answers received

Reported overall, per provider and per topic. The per-provider split is never averaged away: 30% overall made of 40% on one assistant and 0% on another is a completely different problem from an even 30%, and the average hides the part worth acting on.

Citation rate

answers citing a domain you own / answers from providers that return sources

Note the denominator. Perplexity and Gemini return the sources they used. OpenAI and Anthropic answer from the model unless web search is enabled, so they return none — counting them would make this number measure your provider mix rather than your attribution.

Citation rate is always lower than mention rate, and the two move independently. Being described from third-party reviews without ever being linked to is common, and it is a different problem from not being named.

Competitor mention rate

answers naming competitor C / answers received

Computed over the identical denominator as your own mention rate, so the two are directly comparable.

Share of voice

your mention-runs / (your mention-runs + every competitor's mention-runs)

An answer naming you and two rivals contributes 1 to your numerator and 3 to the denominator — a third, not all of it.

Share of voice is only comparable across scans while the competitor list is unchanged. Adding or removing one changes the denominator, and the trend chart marks the discontinuity rather than drawing through it.

Prompt coverage

prompts with at least one succeeded run / approved prompts

A data-quality metric, not a visibility one. It exists so a half-executed scan can never be mistaken for a complete picture.

Absence, and why it is not the same as unknown

A prompt is absent when an assistant answered and did not name you. That is actionable.

A prompt where nothing answered is unknown. We did not find out. Conflating the two is the easiest way for a tool like this to say something false, so Citerva keeps them apart everywhere.

Branded prompts are excluded

A prompt naming your brand will nearly always produce a mention of your brand. Including those would drive mention rate towards 100% while measuring nothing about discovery, so they are excluded from mention rate, citation rate, share of voice, coverage and absence — and counted separately so the exclusion is visible rather than silent.

Reading a change

Two scans of the same prompt set will differ simply because the models are probabilistic. Citerva labels a change as movement only when the underlying run counts differ by more than one; anything smaller is shown as within sampling noise.

Scans below ten prompts carry a small-sample warning. Treat those numbers as directional.