Experiment lifecycle and stopping¶
An A/B experiment prepares both versions, runs the selected shopper sessions, then finalizes the available evidence. You can follow progress while it runs or stop it yourself.
The phases of a run¶
When you click Run experiment, the run starts and moves through these phases on the experiment page.
- Warming up. Squoosh reserves the synthetic shoppers and prepares both versions of your page — Control page (A) and Variant page (B). No shoppers are on the floor yet. While the run is in this phase, the stop control reads Cancel launch.
- Shoppers entering. The first synthetic shoppers begin to shop both versions, and the floor starts to populate. The stop control changes to Stop.
- Live. The full run is underway. The shopper floor fills in as more shoppers finish, alongside a combined funnel and what individual shoppers did and said. The status reads Live.
- Completed. The shopper sessions have finished or otherwise ended, and Squoosh finalizes the available evidence. Completion does not guarantee enough usable evidence for a winner.
What you see while a run is Live¶
While a run is Live, the page shows the shopper floor, the combined funnel across both versions, and feedback from individual shoppers. It does not show a winner, a lift, or a confidence figure.
Those numbers are a finish-line reveal. They appear only once the run completes, so a partial result can't mislead you mid-run. For how to read them after the run, see How Squoosh decides a winner and Reading lift and confidence.
How long a run takes¶
A/B experiments run shoppers in parallel, up to 20 active sessions per experiment, as shared capacity allows. Each shopper runs on both versions, and higher rigor runs more shoppers. Each free slot can admit the next session; the largest rigor levels can take hours or days.
The launch footer shows Execution: Priority instead of a fixed time estimate. Duration depends on the rigor level, site behavior and available capacity. Results arrive as shoppers finish.
Automatic completion¶
A/B experiments do not currently stop when an interim confidence score crosses a threshold. The selected sample runs to completion or another terminal condition before the result is finalized. Live progress is descriptive; it is not an early winner call.
This also applies to the Village conversion lift result that a 100-shopper run in Two URLs mode produces. It uses the sample and evidence requirements set at launch: a paired shopper counts only when a valid outcome was observed on both versions, a pair that fails is retried once, and missing pairs make the result insufficient rather than a guess. Auto-stop is always off for these runs. See Conversion-evidence results.
Conversion Reports have their own panel and completion rules. See Run a Conversion Report.
How a run ends¶
Every run reaches one of these final states. The state shows as the status on the experiment page and in the experiments list.
| Status | What it means | What to do |
|---|---|---|
| Completed | The run finished and Squoosh filed a verdict. | Read the verdict: ship the variant, keep the original, or re-test if it was too close. |
| Undecided | The run finished, but the two versions never separated enough to call a winner. | Re-test with a sharper change between A and B. This is not a failure. |
| Terminated | You stopped the run before it finished. | Launch a new experiment when you're ready. |
A run you haven't launched yet shows as Draft.
Stopping a run yourself¶
You can end a run early from the top of the experiment page:
- Before any shoppers are live, the control reads Cancel launch.
- Once the run is Live, it reads Stop.
Either way, the run ends in the Terminated state. It does not file a verdict.
Note
Squoosh attempts to recover stalled runs. An ended session may have no usable conversion evidence; read the final coverage and any failure reason rather than treating every ended session as a successful observation.