The A/B testing step that divides people into branches by weight: adding branches, how weights become percentages, how it behaves in the simulator, and how to read the results.
"Which message gets more replies?" is not a question you answer by guessing. The step divides incoming people into branches at the ratios you set, so you can test two wordings in the same week on the same audience and see in numbers which one works. This guide covers building branches, what the weights mean, and how to read the outcome.
Random split
What it does
The step sends every person who reaches it into a random branch. The randomness is weighted: with two branches of equal weight, people split roughly in half; make the weights 3 to 1 and three quarters go to the first branch.
Once somebody has been split, they continue down that branch; the rest of the flow behaves normally.
It works best for:
Testing text — comparing two wordings of the same message.
Testing an offer — "10% off" against "free delivery".
Testing a route — is it better to ask a question first or show the product first.
Spreading load — dividing incoming requests between two team branches.
The step sends nothing to the person; it has no relationship with the 24-hour window and works on both channels.
Settings
The step has one section: the branch list. Add branches with Add branch and remove them with Remove branch. Each branch has two fields.
Branch name
The name shown on the canvas, up to 40 characters. This is both the output's name and what is written to the run log — so this is the name you will read your test results by.
Write meaningful names: A: short text / B: long text, or Discount / Free delivery. "Branch 1" and "Branch 2" mean nothing three weeks later.
If a branch name is left empty, the pre-publish check raises a warning.
Weight
The branch's share. A number between 1 and 100, with the panel showing that branch's percentage next to it.
The weight is not a percentage but a share: whatever the total, each branch gets its own weight divided by that total.
Two branches, 1 and 1 → 50% / 50%
Two branches, 3 and 1 → 75% / 25%
Three branches, 1 / 1 / 1 → roughly 33% each
Its outputs
Every branch opens its own output, named after the branch. With no branches the step shows a single Next output — but in that case publishing is blocked anyway.
You need at least two branches. A single-branch split means nothing and the check reports an error. There can be at most 5 branches.
How it behaves in the simulator
In the simulator on the Test tab the split is not random: the first branch is always taken. That is deliberate — a test should give the same result every time so you can step through the flow.
To try the other branches, temporarily reorder them, or select the branch directly and review its steps one by one.
How to read the results
The split shows up in two places:
On the Runs tab, each run records which branch it took (by the branch name).
On the Stats tab you see how many people passed each step; comparing the numbers under the two branches tells you which one carried more people forward.
For a sharper measurement, put a different Conversion step at the end of each branch: "sale-a" and "sale-b". The report then shows both conversions side by side.
Tagging the branches helps too: put an Action step at the start of each branch applying test-a / test-b, then filter on the Contacts screen.
An example
A gym wants to test two wordings of its first message.
They add the keyword membership to the Start step.
They add a Random split step with two branches: A: price first (weight 1) and B: trial class first (weight 1). The panel shows 50% next to each.
Under branch A, a Message: "Monthly membership is £70 with unlimited access. Want the details?"
Under branch B, another Message: "Your first class is on us. Which day suits you?"
Under each branch they add an Action step applying test-a / test-b.
At the end of each road, a Conversion step: membership-a and membership-b.
Two weeks later they check the report, see which wording brought more memberships, and remove the losing branch.
Common mistakes
Ending the test before a week is out. A sample of twenty people tells you nothing. Let at least a few hundred pass through.
Changing two things at once. If branch A changes both the text and the buttons, you cannot tell which one worked. Test one variable.
Leaving branch names as "Branch 1" and "Branch 2". Those names appear in the runs and the stats; make them meaningful.
Forgetting to connect a branch. Anybody landing on an unconnected branch gets nothing.
Starting with more than two branches. Three branches divide your sample three ways and triple the time to a result. Make your first test a two-branch one.
Assuming the step is broken because the simulator never shows the other branch. The simulator always takes the first branch; that is expected.
Leaving the step in the flow after the test is over. Once you have a winner, remove the split; a pointless step makes the flow harder to read.
Limits
At least 2, at most 5 branches.
Weight between 1 and 100; a weight is not a percentage but a share of the total.
Branch name at most 40 characters.
The simulator always takes the first branch.
The step sends nothing to the person; it has no relationship with the 24-hour window.
The split happens per person; somebody entering the flow again is split again.
Related concepts
To measure the branches' outcome, see The Conversion step; to mark people, see Tag actions. There is also a simpler way to send a few wordings of the same message at random: the text variants in The Message step.
You cannot improve a message you do not measure, and the split step is the easiest way to measure one.