Experiments
Predict what a change does to a goal before you ship it, on people built from your real sessions, and check the prediction against what really happened.
How a prediction is made
Describe the change in a sentence (“make the sign-up button twice as big”, “move pricing above the fold”, “add a free tier”), pick how many simulated visitors to ask (10 is quick, 50 is the most accurate) and press Run.
- Grounded on your sessions. Each simulated visitor is built from one real recorded session on your site: what they did, where they clicked, what broke for them, their device.
- Sampled at your real mix. Visitors are drawn in the proportions your segments actually occur, and each rates the goal under the current flow and under your change.
- Checked, not trusted. Every prediction is logged with the people it ran on. When you ship, log the real result and the calibration adjusts to your workspace.
Reading a prediction
A prediction is a goal rate before and after, overall and per segment, with the reasons the visitors gave. A change that helps one segment and hurts another shows as such, which is usually the interesting part. Predictions made earlier stay on the page with what happened when you shipped, once you told it.