A/B Split Testing
A box with an arm on each side and arithmetic deciding which one somebody takes. Because it is arithmetic and not a coin, asking twice gives the same answer twice.
Which arm a lead takes, and why it never changes
arm B- Send: case study
- Wait 3 days
- If replied
- Send: short ask
- Wait 5 days
- Exit
Product figures from the platform’s own defaults - not customer averages
How it works.
The arm is computed, never rolled
The run identifier goes through a small fixed hash and the remainder against the total weight picks the arm. Nothing random is drawn and nothing is stored to remember the choice, so re-running the same step lands the same side every time.
The weights are what divide the range
Each arm carries a weight and the engine cuts the range in proportion to them, so a ninety-ten division is two numbers on the box rather than a different kind of step. A box you drag on arrives at fifty apiece, which is the even split most people want first.
Each arm is a port with a whole subtree behind it
Whatever hangs off one side is a full branch of the cadence in its own right, with its own sends, pauses and questions. The two sides never have to meet again, and one of them may simply stop.
The assignment is recorded the moment it is made
Before moving on, the box writes down which side it chose against that person's run. So the population either side can be counted from the event record rather than reconstructed from who received what.
One moment in every run.
Every run passes through the same seven. A/B Split Testing is the lit one, and everything either side of it is a different page in this category.
- 01Trigger
the thing that happened first
- 02Enrol
how somebody gets onto it
- 03Wait
the pause, and what governs its length
- 04Branch
the fork, and which side is taken
- 05Act
the mail, the text, the task that goes out
- 06Measure
what counts as it having worked
- 07Exit
how somebody comes off it
The specifics.
7 facts- Arms
- Two in the box you place. The stored format allows more, each with a port name of its own
- How the side is chosen
- A remainder of the hashed run identifier against the weights added together. An arm with no weight written on it counts as one
- Stability
- Deterministic. Nobody changes sides part-way through a cadence
- Weights
- Fifty and fifty on a new box, held on the arms themselves and changed one arm at a time
- Control groups
- A different mechanism. A percentage set on the sequence parks those people before they take any step
- What it will not do
- No significance test, no winner declared, and no promoting the better side afterwards
- Not the same as
- This divides an audience by weight. Reacting to what somebody did is Branch Logic
What starts this, and what it starts.
Automation is only ever a middle. These are the things that set it running and the things that run because of it.
More in Automation & Flows
13 capabilitiesSequences, dispositions and webhooks that act without being asked.
A unique URL per integration with HMAC signing, rate limits, payload caps and idempotency - plus a log of every delivery it ever accepted or refused.
Enrol a lead once and the sequence carries them: sends, waits, branches and exits, with a ledger that makes a duplicate send structurally impossible.
Hold a lead for minutes or days between steps, clamped to the sending window you set, so that an overnight wait never lands at four in the morning.
Set the hours a sequence may send inside, as a fixed offset you enter yourself, and every wait is pushed forward into them before it is stored.
One do-not-contact roll per workspace, checked when a lead is enrolled and again inside the send, straight after that send has taken its own claim.
New leads enrol themselves the moment they are created, and the four guards that can refuse an add are read on every single route into a cadence.
Drop a call step into a sequence and it raises an assigned task on the lead's owner, with the outcome the rep logs coming back as a branchable answer.
Twenty-six behaviours attach to wrap-up codes, and twenty codes ship already pointing at one, so ending a call the right way books the follow-up.
Six event types a CRM rule fires from: stage change, field update, new record, reassignment, schedule and inactivity, agreed in four places at once.
Send the email, assign the owner, raise the task, set the field, move the stage or call a webhook. Six modules, run in the order you stored them.
Eight comparison operators guarding each action, combined with AND and only AND, so a rule meaning either of two things is built twice, as two rules.
A text is a step on the same canvas as the email and the call task, written into the real thread and carried out by the number it is going out from.
A versioned path, a key minted once and kept as a hash, and the same role and profile model the interface runs on deciding what that key may touch.
The rest of the platform.
Five more categories, all on the same record and the same bill. Each card names three of its capabilities, so you can tell from here whether it is worth opening.
Automation & Flows
See a/b split testing on your own floor.
Thirty minutes, your numbers and your data. We will set a/b split testing up live and you can decide from the thing itself rather than from this page.
14-day trial · no card · migration included