How we measure AI visibility

Four KPIs, computed deterministically from stored answer transcripts. A “scored answer” is an answer an engine actually returned for a tracked prompt this period — the denominators below count real answers, never the configured prompt count.

We measure each client’s configured answer engines through their APIs, routed via the OpenRouter aggregator. Engines that exist only as a user interface expose no API to query: they are not measured, and every dashboard view and report names them as an explicit coverage gap — never a silent omission.

What we measure is each engine's API, not its consumer app: the app may personalise, remember earlier turns, use your location, or run a different retrieval stack, so the two can cite different sources for the same question. We query logged out, with no location and no memory, so a reading describes the engine rather than a person using it.

Automated interaction with a consumer UI surface may conflict with that provider's terms of service. Where an operator explicitly opts in, we capture AI Overviews / Copilot at low volume, from the operator's own signed-in browser session, with no captcha solving, proxying or fingerprint evasion, and every such answer is disclosed as UI-captured — never presented as API-measured.

Each client is measured on a fixed cadence — weekly or monthly, set per client. Every KPI, comparison, and delta is scoped to its measurement period, and each report states the period it covers.

Where a prompt is asked more than once in a period, the rate pools every repeat; a per-prompt interval is computed over that prompt's own repeats; an engine's overall rate carries a prompt-clustered interval, which counts each prompt once however often it was repeated, while a cross-engine interval for a single prompt stays withheld because a handful of engines is too few to estimate its spread from; and the Recommended reading is judged on the first repeat only.

AI answers are non-deterministic: the same prompt on the same engine can return a different answer minutes later. A single run is a sample, not a truth.

Repeat sampling is available for core prompts — the highest-value buyer queries — at up to R runs per engine (default 3) where a client has it configured. Where more than one sample per (prompt, engine) pair was actually stored, we report how many of those runs mentioned the brand together with a 95% Wilson confidence interval: the uncertainty range the true mention rate plausibly falls within given that sample size.

Sample size: a single run per (prompt, engine) pair — n=1 — is the default; repeat sampling runs only for prompts a client has flagged as core. Every dashboard and report states the sample size its own figures were computed from, so a confidence interval is never implied where only one sample was stored.

A metric from a single run carries no interval at all — it is labeled “directional — single run”. When a budget stop leaves fewer samples than configured, we state the actual n of R configured runs used — a partial sample is never presented as a full one, and a prompt with no samples yet is marked explicitly rather than shown as 0%.

Independent research documents large run-to-run variance in AI recommendations — the reason a single run is reported as directional rather than as a precise figure, and the reason repeat sampling exists for core prompts at all:

Collection method: API-based. We query engine APIs and store the raw transcripts we score — no synthetic extrapolation, and no UI capture other than the opt-in, disclosed case described in the engines section. Any number we publish can be traced back to a stored transcript.

Some campaigns measure a person, such as a founder or an expert, instead of an organisation. For a person the headline is the Known stage: whether the engines recognise them when asked who they are, and whether they get the facts right. No new metric is introduced for this; the definitions below say how the existing figures read for a person.