Control questions: how to prove an AEO change worked
In short
Control questions are the tracked questions a change didn't aim at. When you publish a page for some questions, you compare how those moved against how the untouched ones moved over the same weeks. If both rose, something else changed; if only the targets rose, beyond the noise band, the page is the likely cause.
Keats, from PressRoom. Updated Read as markdown
Say you publish a comparison page aimed at three buying questions, and three weeks later you're named more often on them. Did the page do that? Maybe. Assistants update, rivals publish, a roundup adds you, seasons change what people ask. Any of those would lift your numbers without your page.
Control questions separate the two. They're the questions you track but didn't touch. The test is the difference:
- Targeted questions up, controls flat: the change probably worked.
- Both up by about the same: the market moved; your page may have done nothing.
- Targeted flat, controls down: your page may have held ground the rest lost.
Keats does this for every finished action card. When a change is reported live, it records the targets and the controls, then reads both at day 21 and day 35. A verdict needs the gap between them to be wider than the noise band. On Prepco, the made-up brand in our demo, a comparison page moved its targets +35 points by day 35 against +2 points on the controls.
The same logic works without Keats: hold some questions back from every change, and compare.
Sources
- Keats: llms.txt, the operations and workflows for agents, read November 12, 2026