A single AI rank check is a coin flip. We don't sell coin flips.

Why isn't one check enough?

Because the same question, asked twice, gives two different answers. AI engines are non-deterministic by design: they generate a fresh set of search queries almost every time, and the businesses they name shift from run to run. Research on repeated prompting has found that only a small fraction of citations stay consistent across three runs of the identical prompt. One check tells you what happened once. It does not tell you what your customers see.

So we don't ask once. Try it — here is one real money question, run three times on the same engine, minutes apart:

Same engine, same question, same day. Three answers. Now imagine deciding your strategy off the first one. Anonymized from a real panel.

What does Cited measure — the API or the app?

The surface your customers actually use. An API call and the consumer app can return different model versions, different search behavior, and different citations; an API request without live search returns the model's memory, not what a person sees on screen. So we collect money questions from the real consumer interface. Manual browser runs are ground truth by definition — the same standard behind the published local-search studies this field is built on.

How do you score an answer?

Scoring is pre-registered — defined before the baseline runs, and never reinterpreted mid-engagement. Every run is graded against four fixed events, for your business and for each competitor:

• Mention. Your business is named anywhere in the answer text.
• Recommendation. You appear in a recommended list — and we record your ordinal position.
• Citation. Your URL is linked as a source the engine drew from.
• Sentiment. How the answer treats you: positive, neutral, or negative.

Why report every engine separately?

Because each engine runs on a different source stack, and winning one says nothing about the others. Studies of cross-model agreement put it near 40% — the engines name the same top business fewer than half the time. Blend them into one score and you hide the exact thing you're paying to see. So we never blend. Every fix we prescribe is aimed at a named engine.

What's the statistics behind it?

A citation rate is a proportion, so we treat it like one. We report the rate across runs with a confidence interval, not a bare percentage. Repeated-generation measurement with panel-level means is now the published standard for evaluating language models; the variance research is clear that single-run estimates swing by several points even at temperature zero, and that prompt wording is a dominant source of error. That last finding is why the panel is frozen: we never edit a prompt mid-engagement, because changing the words changes the number.

Common questions