Metrics

Every number in Geonimo is defined here, with its formula and its failure modes. A metric you cannot define is a metric you cannot defend in a review.

How metrics are produced

Every metric is an aggregate over answers — one capture of one prompt on one engine, at one location, on one day.

answers → mentions (extracted)   → visibility, share of voice, position
        → citations (captured)   → sources, domains, quoted vs listed
        → surface units          → shopping and ad placements

Aggregates are stored per day × engine × entity, which is why every screen can slice by engine and date instantly, and why changing a date range never re-runs collection.

Confidence, not just point estimates

This is the part that separates a measurement from a number.

AI answers are non-deterministic: the same prompt asked twice can name different brands. Each prompt is therefore run several times per engine per day, and the spread across those runs produces a 95% Wilson confidence interval on every rate.

Two consequences you will see throughout the interface:

• Every rate carries an interval. A visibility of 46% with a ±2 point band is a result; the same 46% with a ±22 point band is a rumor.
• Changes are tested against a noise floor. Comparing two independent estimates carries about 1.4× the single-point half-width, so the floor quoted next to a delta is wider than the interval on either endpoint. The interface always captions the change floor, never the narrower per-point one, because the change is what you are about to act on.

A move inside the floor is not marked significant, and should not start a project.

Interval width is driven by how many answers back the estimate. That is why prompt allocation is fixed per subject rather than spread thin — see Prompts.

Three rules for reading them

1. Rates over counts. Mention counts inflate whenever you add prompts. Visibility and Share of Voice are ratios over the same answer set, so they stay honest when the instrument changes.

2. Significance over movement. Check the marker before believing a delta.

3. Segments over averages. A blended number across brand and product prompts, across every engine, tells you almost nothing. Read the groups separately.

Unmeasured is not zero

Across every screen, a value that has not been measured renders as a dash — never as a zero. A product added today has no reading; a product measured across 200 answers and named in none has a reading of zero. Only one of those is bad news, and conflating them would manufacture bad news out of new data.

What moves a metric — and what only looks like it

Real movement: content you shipped once engines re-crawled it; a new citation on a trusted domain; a competitor gaining or losing a key source; an engine changing how it selects sources.

Instrument changes, which are not movement:

• Adding prompts — changes the denominator of every rate
• Adding or removing competitors or products — changes the Share of Voice denominator
• Changing location — changes which brands engines surface at all

Record the date whenever you change the instrument. Half the confusing charts in analytics are a configuration change nobody wrote down.

Not yet measured

Sentiment and recommendation strength require the enrichment pass, which is not yet running in production. Where a screen depends on it, the value is captioned as sample data or hidden rather than estimated. See Product status.