Metric · Changes
Material change
Definition
A movement large enough to outrank routine noise, measured between seven-day windows of runs rather than between two single runs: 5 points for mention and citation rates, 4 points for share of voice, 5 points for sentiment shares, and a quarter of a rank for average position.
How it is calculated
AI answers are non-deterministic, and one run holds about one answer per cell, so a day-to-day comparison is mostly sampling wobble. Pooling a week of runs multiplies the answers behind every number, which is what lets these thresholds sit low enough to catch real movement. Two spans are reported. A shift compares the last seven days with the seven before, and catches an abrupt break. A drift compares the newest week with the oldest of four, and fires only when every week in between moved the same way, so a slow slide that no single week ever breaches still surfaces while a bounce does not. Windows are compared only over the cells (one prompt on one surface) that every compared window actually answered, so a partial run can never fabricate a change. That count is usually smaller than prompts times surfaces, because a surface that returned no answer for a prompt leaves nothing to compare. Share of voice, position, and competitor comparisons pause when the tracked competitor set changed anywhere in the compared span, because those metrics are relative to that set.
Where to see it
This metric appears in the refd dashboard and in the no-signup demo, and every value links back to the AI answer it was scored from. A number you cannot trace to an answer cannot be checked, so refd keeps the raw response behind each one.
The definition above is the one the product uses. It is read directly from the same source the dashboard and the MCP server read, so this page cannot drift from the software.
See your evidence