The instrument, versioned
Every change to how we measure, dated
A number is only checkable if the ruler that produced it is versioned. This page lists every change to the measurement itself, including the uncomfortable entries: when our own benchmark found defects in our detector, the fix made numbers before and after that date incomparable, and that belongs here more than any improvement does.
First-mention position and the Answer Instability Index
August 16, 2026Each stored answer now also records where the first clean brand mention lands (as a fraction of the answer), and repeat samples of the same question are compared to surface questions the engine flips on.
Comparing across this change: Position exists only on answers measured from this date onward; earlier answers report it as not measured, never as position zero. Instability is computed retroactively, because the repeat samples were already stored.
Owner exclusions shrink the denominator, with the reason published
August 15, 2026A site owner can exclude a question as a legitimate non-fit. The exclusion requires a written reason, and that reason prints on the coverage receipt beside every share it changed.
Comparing across this change: A share computed after an exclusion covers fewer questions than one before it. The receipt states the exclusion and the reason, so the two numbers are never silently mixed.
Answer framing labels
August 13, 2026Each answer that names the brand is labelled with HOW it framed the brand (recommended, listed, mentioned, criticised, inaccurate), written at measurement time from fixed rules.
Comparing across this change: Answers measured before this date carry no label. A null label means not labelled, never neutral.
Free checkup sample floor
August 12, 2026The free checkup refuses to publish a per-engine rate below a minimum answer count. The first live run produced a 0% built on a single answer, which reads as a finding and is actually an outage; the floor exists so that cannot happen again.
Comparing across this change: Thin samples render as a stated sample-count note instead of a rate. No published checkup number sits on fewer answers than the floor.
Brand-mention detector: three defects fixed
July 29, 2026Our own accuracy benchmark found three defects in the mention detector: it inspected only the first occurrence of a name, multi-word brands failed across line wraps, and the word-boundary test was ASCII-only (over-reporting brands adjacent to accented characters). All three are fixed and preserved as permanent regression cases.
Comparing across this change: THIS IS THE BIG ONE: mention-based numbers before and after this date come from two different instruments. A trend spanning July 29 should be read as two series, not one.
Every reported share carries a 95% Wilson interval
July 17, 2026Shares moved from bare percentages to count-over-denominator with a Wilson score interval, and change detection began requiring a movement to clear its own interval before rendering as a change.
Comparing across this change: Older surfaces showed the same counts without intervals; the underlying counting did not change.
citation-1.0.0: the baseline aggregation
July 2026The versioned baseline: a citation counts only when the engine itself returned the site as a source; a brand mention counts only on a clean word boundary; every metric row persists the version it was computed under.
Comparing across this change: The version string rides on persisted rows, so any future definition change becomes a visible version bump rather than a silent redefinition.
How each metric is computed lives at how we measure; the accuracy of the detector itself is published at detector accuracy.