Cell state: one prompt on one engine
A cell is read from the latest reading in the window that has at least one successful answer on that engine. Failed answers do not count; a reading where every answer failed is skipped and the one before it is used. Withn successful answers in that reading:
The rules are checked in that order, and the first that holds wins.
- At least half, not at least one. A cell is a label on
nanswers. At sampling depth 3, a label that one answer in three could flip would move inside the sampling noise. Ties go to you: named in 1 of 2 answers is Named. - Lost needs a rival in the words. A competitor counts only when the answer names it in its text. A competitor that appears only among the links is a source, not a name the answer chose. A lost cell says which rival, the one named in the most answers (ties broken alphabetically), and when the reading was taken.
- Absent is a measurement. It means the answers were read and nobody you track was named in enough of them. Not measured means there was nothing to read. The grid draws them differently, and never one in place of the other.
Prompt state: one prompt across its engines
The prompt’s state is folded from its cells. Only measured cells count: let M be the number of engines with a measured cell.
The precedence is won, then contested, then lost, then absent. A prompt named on one engine and lost on two is Contested, not lost: you are in the answer somewhere.
An engine that is not measured is left out of M. It neither helps nor hurts the prompt, because counting it would turn a gap in our coverage into a claim about your brand. A prompt read on four engines, where one failed for the whole window, is judged on the other three.
Each prompt also counts the engines it is lost on. That number orders the red questions on the Scoreboard.
Where you see it
- The Scoreboard: each question’s colour on each engine is its cell, and its colour across engines follows the prompt rule above.
- In a chat: the Questions by engine card draws each question’s cell on each engine. See What the agent shows.
- A prompt’s detail: each engine shows its state with the
nit was read over and when, from the same rule. - Over MCP:
query_evidencecounts prompts by state (prompt_state_count) and filters any metric bystate. See MCP tools.
Reading it
State is a word, not a trend. It describes the latest successful answers in the window, so it moves as soon as one reading moves, and a single reading at sampling depth 1 can only ever be all or nothing. Read then beside a cell before you act on it, and use the rates and their significance test for whether anything changed.
Next: Share of voice for the rate behind Named, and Competitors for the rivals a Lost cell names.