Traffic allocation: when to use 50/50 and when to use 90/10

The traffic split is the setting people change most and think about least. It looks like a preference. It's actually a trade between speed of learning and exposure to risk, and the right answer depends entirely on what you're testing.

Why 50/50 is the default#

An even split reaches statistical significance fastest for a given total traffic. That isn't a convention, it's arithmetic: the precision of your comparison is limited by the smaller group, so any imbalance wastes sample.

Going from 50/50 to 90/10 roughly doubles the time you need, because your variant group — the one that determines how precisely you can measure the difference — is now a fifth of the size.

For a small brand this matters a lot. If a 50/50 test takes four weeks, the 90/10 version takes about eight, and eight weeks is long enough that your traffic mix, your ad spend and possibly your product range will have changed underneath the experiment.

So: default to 50/50, and require a reason to deviate.

The reasons to deviate#

You're managing blast radius on a risky change. A full checkout restructure that might break something should not go to half your customers on day one. Start at 90/10, confirm nothing is on fire — error rates, completion rate, support tickets — then move to 50/50 for the actual measurement.

Note that this is two phases, not one long uneven test. The first phase is a safety check, not a measurement. Trying to measure at 90/10 is where the cost lands.

The variant costs you money per exposure. If the variant offers a discount, every visitor who sees it and would have paid full price is margin you gave away. Here the uneven split is a business decision and the slower result is the price.

You're testing something with permanent consequences. Anything that changes what a customer is promised — pricing, delivery terms, guarantees — you can't unsay to the people who saw it.

What an uneven split does not fix#

It does not make a test safer to interpret. People sometimes run 90/10 thinking it's more conservative statistically. It isn't; it's just slower. The confidence threshold is what governs how likely you are to be wrong, and that doesn't move with the split.

It also doesn't reduce the total number of people affected by a wrong decision, because the decision gets applied to everyone afterwards. It only limits exposure during the test window.

Ramping, and the trap in it#

A common pattern is starting at 90/10 and ramping to 50/50 over a few days. This is sound for risk management but has a real statistical catch: the visitors from the early low-exposure period saw the variant under different conditions — different day of week, possibly different campaigns running — than the later ones.

If you pool everything, you're comparing a control measured mostly in week one against a variant measured mostly in week two. Seasonality and campaign changes then show up as variant effects.

The clean approach: treat the ramp as a safety phase, discard its data, and start counting when you reach your final split. You lose a few days of sample and you keep a result you can actually defend. Bixlow lets you mark the measurement start separately from the launch for exactly this reason.

Three variants and the split#

With A/B/C, an even three-way split is again the fastest route to comparing all three. But if one variant is a long shot you're including out of curiosity, giving it a smaller share doesn't help — you'll simply fail to reach a conclusion about it while slowing the ones you care about.

If a variant isn't worth an equal share of traffic, it isn't worth including. Drop it and run it later.

Key takeaways#

• 50/50 reaches significance fastest. Every deviation costs time, and 90/10 roughly doubles it.
• Deviate for blast radius, for margin cost per exposure, or for changes you can't unsay — not for statistical comfort.
• An uneven split doesn't make a result more reliable. Only the confidence threshold does that.
• If you ramp, treat the ramp as a safety phase and start measuring at the final split.
• If a variant doesn't deserve an equal share, it doesn't deserve to be in the test.