When you configure an A/B test node with a defined ratio between variants (for example, 50/50 or 70/30), the number of customers in each variant might not come out exactly matching that ratio. This article explains why this happens and how to get closer to an exact split.
Understand why the split isn't exact
Think of it like flipping a weighted coin, where heads and tails have an equal chance on every flip:
If you flip a fair coin 1000 times, you're unlikely to land on exactly 500 heads and 500 tails.
The same principle applies to variant selection. Each customer's assignment is an independent random event, just like a coin flip - it's simply weighted according to the ratio you configured.
Because every preview involves a new round of random selection, the split you see can look different each time you check it.
Get a precise split
Add a Customer Limit node after each variant path of your A/B split node to enforce a precise, deterministic split:
Place a separate Customer Limit node right after each variant branch - one on the A path, another on the B path, and so on if you have additional variants.
Set each node's limit to the exact number of customers that variant should receive, based on your configured ratio. For example, for a 70/30 split across 1000 customers, set the limit on variant A to 700 and on variant B to 300.
Each Customer Limit node then stops passing customers through once its configured count is reached, so only that specific number continues along the path for that variant. This works because the Customer Limit node counts occurrences and deterministically cuts off the flow, unlike the underlying random variant assignment.
If your total audience size changes between runs, recalculate and update the limit values accordingly - a limit set for one audience size won't automatically scale to a different one and needs to be reset after each run.
Even with Customer Limit nodes in place, you may still see a small difference between variants. This happens because not all customers get successful enqueues or deliveries, so the number of customers who complete each variant's path doesn't always match the number who entered it.