---
description: "How a Dalton experiment runs: an even A/B split over a planned window, the data it needs, and when a result can be called."
---

# How Experiments Run

Every Dalton experiment is a straight A/B test. Traffic splits evenly between your original page and each variant, the split never moves, and the result is read at the end of a window fixed before launch.

![A running experiment counting down to its conclusion](/v2/experiment-running.png)

## The planned window

Dalton sets a conclusion date before the test starts, based on the traffic the targeted pages get.

| Window | Given when |
| --- | --- |
| **2 weeks** | The pages can send about 5,000 visitors to each variant in that time |
| **4 weeks** | They need longer |
| **6 weeks** | They cannot reach it. Also the hard ceiling |

Windows are always whole weeks, so a run counts the same number of Mondays as Saturdays. For the first three days the plan shows as two weeks and is provisional, until there is enough traffic to project from.

Days only count while the test is collecting. Pausing does not advance the countdown.

::: tip Check the window before you launch
The targeting picker estimates **time to results** before you commit. Widening the targeting is cheaper than waiting. See [Targeting Groups](/experiments/targeting).
:::

## Why no leader appears early

While a test runs, **Leading variant** stays empty even when the numbers look convincing. Lift and chance to beat are shown throughout, but no leader is named until the window completes.

Stop a test the moment it looks good and you will stop it on the days it happened to look good. Two exceptions, both using a calculation that stays valid however often it is checked:

- **A clear loser** is flagged for early stopping once every variant has 50 conversions and all are decisively behind.
- **A clear win** can be called early, but only after it has held for two days.

## What Dalton needs before it reports

Figures show as a dash until **every** variant, including the original, clears both bars:

- **350 visitors**
- **50 conversions**

Conversions are the binding one. A page converting at 0.5% clears 350 visitors on two orders, and one more order then swings the lift by tens of percent.

## Calling the result

| Chance to beat | Verdict |
| --- | --- |
| **95%+** | **Winner.** Deploy it, or conclude it |
| **80-95%** | **Promising.** Worth another week if under the six-week ceiling |
| **5-80%** | **No effect** |
| **Under 5%** | **The original won** |

A test that runs its full window without clearing the data bars is reported as inconclusive. The page is too quiet for that question, and running it longer will not fix it.

## While it runs

- Do not stop a test on a bad day three unless Dalton flagged it.
- Do not edit the original page mid-test.
- Pause if something is broken. Changes revert instantly.

Next: [Reading Your Results](/experiments/reading-results).