Research prompts
Run a set of prompts once for a study rather than tracking them: how research batches are budgeted, how to follow one, and how to pull the results.
Updated October 20, 2026
A research prompt runs once, for a study, and is never scheduled. It is how you answer a question like "which sources get cited across 200 buyer questions in this category" without turning 200 questions into 200 tracked prompts.
How it differs from a tracked prompt
| Tracked | Research | |
|---|---|---|
| Schedule | Runs on a cadence | Runs once |
| Plan limit | Counts as a tracked prompt | Does not |
| Budget | Charged per run as it goes | Whole batch costed and reserved up front |
| Purpose | Measuring a position over time | Answering a question once |
Both produce the same kind of data: runs, answers, fan-out searches, cited and retrieved sources. The research grain just arrives all at once.
Running a batch
Research batches are submitted through your agent over MCP, because that is where the prompt list usually comes from in the first place. The shape is:
- Submit the prompts, with a sample count, and an idempotency key.
- The batch is costed as prompts multiplied by samples and checked against your monthly budget. If it does not fit, the whole batch is refused rather than partly run.
- You get a batch id back immediately. The work happens in the background.
- Poll the batch status rather than re-submitting.
- When it is done, pull everything by batch id.
Give every prompt an external id
An external id is your own identifier for a prompt, something like f001. It comes back on every run, every citation and every export row, which is what lets you join the results to your own study spreadsheet without matching on prompt text.
It is also a deduplicator: a prompt whose external id already exists in the project is reused rather than duplicated, so a re-submitted study does not create a second copy of every question.
Following a batch
Batch status reports prompts and probes completed, the per-engine outcome, the spend so far, and, if it stalled, which prompts never ran. Poll that rather than resubmitting; resubmitting without an idempotency key is how a study gets paid for twice.
Pulling the results
Export by batch id, as CSV or XLSX, or read the same records through the reporting API with the batch filter. Every row carries the prompt, its external id, the engine and the citation detail, so a study never has to open runs one at a time.
Samples per prompt drive both the cost and the quality of a study. A 200-prompt batch at one sample each is 200 units and a wide band on every prompt; at three samples it is 600 units and a usable band. Decide which question the study is answering before setting that number.
Questions about research prompts
- Do research prompts count against my tracked-prompt limit?
- No. They are never scheduled, so they do not occupy a tracked slot. Their runs are charged to the monthly budget up front instead.
- What happens if the batch is too big for my budget?
- The whole batch is refused rather than half of it running. You find out before paying, not afterwards.
- Can I retry a batch safely?
- Yes, with the same idempotency key. The retry returns the original batch instead of running and charging it again.