Control prompts
A control prompt never receives optimization work. Why that makes it the only way to attribute a move to your work rather than to model drift.
Updated October 20, 2026
A control prompt is one you deliberately never optimize. It is tracked, it runs on the same cadence as everything else, and no work is ever aimed at it.
Why you need one
Model behaviour drifts. Providers change models, change how aggressively they search, change which sources they weight. When that happens, every number in your project moves at once, and from inside your project it looks exactly like the result of what you shipped last week.
The control is what separates those two cases. When your worked prompts climb and the control stays flat, the difference is attributable to the work. When both climb together, the market moved and you are reading a tide, not a result.
Picking one
Three properties:
- Representative. Same kind of question, same market, same sort of buyer as the prompts you are working. A control from an unrelated topic does not drift in the same way.
- Genuinely untouched. No page written for it, no source claimed for it, no mention of it in the work queue. If you optimize it even once, it is not a control any more.
- Stable. Not about a product you are about to launch or a market that is about to be reshaped.
The easiest honest control is a question adjacent to your core set that you have decided not to pursue this quarter.
Setting it
Set Prompt type to control when you create the prompt, or change it on an existing one that you have never worked. It is a flag rather than a separate kind of object: a control carries the same answers, sources, fanouts and history as any other prompt.
Reading it
Put the control on the same chart as the metric you are reporting.
| Worked prompts | Control | Reading |
|---|---|---|
| Up | Flat | Your work. This is the result you can claim. |
| Up | Up | Market or model drift. Claim less. |
| Flat | Down | You held a position while the ground moved. Often the best result of a bad quarter. |
| Down | Flat | Something you did, or a competitor did. Go to the cited sources. |
The reporting value
A client asking "how do we know this worked" is asking for a counterfactual, and the control is the only one you have. One extra prompt, measured all along, is the difference between a claim and a demonstration.
Set the control up on day one. A control started three months into a project has no baseline over the period you are being asked about, which is exactly when you need it.
Questions about control prompts
- How many controls do I need?
- One is enough to catch drift. Two or three, spread across topics, tell you whether the drift is market-wide or specific to one part of it.
- Does a control cost the same as a tracked prompt?
- Yes. It runs like any other prompt, which is the point: it has to be measured the same way to be comparable.
- Can I promote a control to a worked prompt later?
- You can, and then it stops being a control from that date. Retire it as a control and start a new one rather than losing the baseline.