Runs and answers
Every figure traces back to a stored run. What a run holds, how to find the one behind a number, and how to export runs as CSV or XLSX.
Updated October 20, 2026
A run is one prompt, on one assistant, at one moment, with the full answer kept. Every chart in the product is an aggregation of these, and every aggregation opens back to them.
What a run holds
- The prompt as asked, and the model that answered.
- The complete answer text.
- Whether search was available, and whether the model actually fired one.
- The fan-out searches it ran.
- The sources it retrieved, and which of those it cited.
- The brands recommended in the answer, in order, and how yours was matched.
- The token and search cost of the probe.
- The batch it belonged to, and the configuration in force at the time.
Finding the run behind a number
Open Runs and filter by prompt, assistant or date, or click through from a point on any chart. This matters more here than in a traffic tool because the underlying data is prose: when a number looks wrong, the answer text usually explains it in one read.
Reading answers deliberately
Early in a project, read answers instead of charts. Ten answers tell you whether the prompts are good, whether the brands are mapped, and what models currently believe about your market. The third of those is often the most valuable thing you take to a meeting.
Later, read answers when a number moves. The question "what changed in the answer" is almost always more productive than "what changed in the chart".
Failed runs
A probe can fail on the provider side: no credit, a revoked key, a rate limit. Transient rate limits are retried automatically. A persistent failure is surfaced with the provider's own reason rather than dropped, so a gap in the data is explained instead of silent.
Errored and degraded probes are counted separately from clean ones and are excluded from the rates, which is why the probe count on a rate can be lower than the total runs in the range.
Export
Runs export as CSV or XLSX, filtered the same way the screen is. The export carries the prompt, its external id, the engine and the citation detail per row, so a study or a client report does not have to open runs one at a time.
For anything recurring, the reporting API reads the same records, so a dashboard you build once stays current.
Before a client call, export the range you are presenting. Being able to open the actual answer behind a number, in the meeting, changes the conversation more than any chart does.
Questions about runs and answers
- What is the difference between a run and a batch?
- A batch is one measurement of a prompt; a run, or probe, is one answer inside it. Rates are computed per probe, and the confidence band depends on how many there were.
- A run says degraded. What does that mean?
- It completed but came back incomplete, so it is counted separately rather than being mixed into a clean rate.
- How far back do runs go?
- All of them, for the life of the project. Pausing or cancelling stops new runs and deletes nothing.