Every A/B test answers two questions. Which variation wins is the one everyone talks about. The one that quietly decides how much the test costs you is the other: how much traffic does each variation get while you wait?
In Splitlab that second question is one dropdown, Traffic Split, with three options: Equal Distribution, Custom Distribution and Multi Arm Bandit. They look like a small setting. They're actually three different philosophies about risk, speed and how much revenue you're willing to spend to learn something.
Key takeaways
Equal gives every variation the same share. It's the fastest route to a trustworthy answer and the right default for most tests.
Custom lets you choose each share yourself. Use it to limit risk on a page that already makes money, or to roll a winner out gradually.
Multi Arm Bandit moves traffic toward the variation earning more while the test is still running. It earns more during the test and gives a less precise answer at the end.
Pick the split before launch and leave it alone. Changing it mid-test changes who sees what and muddies the result.
What a traffic split actually controls
A traffic split is the share of visitors each variation receives. If you're testing a control page against two new versions, the split decides whether that's 34/33/33, 80/10/10, or something that changes every day.
It doesn't change what you're measuring. Your primary metric, ideally revenue per visitor rather than conversion rate, stays the same. What the split changes is:
How fast you get an answer. A variation with 10% of traffic collects data ten times slower than one with 100%.
How much revenue the test costs. Every visitor sent to a losing page is a sale you probably didn't make.
How much you can trust the answer. Shares that move during the test make the final comparison less clean.
Every option below is a different trade between those three.

Equal Distribution: the clean experiment
Equal Distribution splits traffic evenly across every active variation, control included. Two variations get 50/50. Three get 33.3% each. Pause one and the remaining variations share its traffic evenly. It's the default for every new Splitlab test.
Why it's the default: an even split gives every variation the same amount of evidence at the same time. That’s the fastest way to reach a confident answer for the whole test. No variation is starved, nobody has to argue about why one arm got less data, and the result reads exactly the way the statistics assume.
What it costs: while the test runs, the losing variations get as much traffic as the winner. If Variant B is clearly worse by day three, it still gets a third of your visitors until you stop the test.
Choose Equal when:
You're testing a genuinely new idea and have no strong guess which version wins.
You'll make a permanent decision from the result, such as a new offer, a new price or a redesigned page.
You want the shortest time to a result you'd defend in a meeting.
Traffic is modest. Splitting it unevenly would leave some variations waiting weeks for enough data.
Custom Distribution: you set the shares
Custom Distribution lets you type the share for each variation yourself. Any variation can get between 0% and 100%, and the shares always add up to 100%. You can lock a variation's share so it stays put while you adjust the others, and the unlocked ones rebalance automatically. A variation set to 0% stays in the test but receives no new visitors.

Why you'd use it: the control page is often the one paying the bills. Custom lets you test boldly without betting the business on it. Send 80% to the proven page and 10% each to two challengers, and a bad idea can only ever touch a tenth of your traffic.
What it costs: the variations with small shares collect data slowly. In an 80/10/10 test, each challenger needs about eight times longer to collect what a variation in a 50/50 test collects. The answer is slower, not wrong.
Choose Custom when:
The page is a top earner and a bad variation could cost real money: a checkout, an order form, a high-traffic sales page.
You're trying something risky, like a much higher price, a different offer or a stripped-down page, and want to cap the downside.
You're rolling out a winner gradually, for example 20%, then 50%, then 100%, while you watch refunds and support tickets.
One variation needs to sit out for a while without being deleted. Set it to 0%.
Don't choose Custom to "help" a variation you like. Giving your favourite more traffic doesn't make it win. It only means the others take longer to prove you wrong.
Multi Arm Bandit: the split that learns
Multi Arm Bandit starts even and then moves traffic toward whichever variation is earning more on your primary metric. Splitlab checks the results every hour and recomputes the split once 20 new conversions have come in, so busy funnels move within the hour and quiet ones wait for real evidence. Variations that are winning get more visitors; variations that are losing get fewer.
It doesn't go all-in on an early leader. Every Splitlab bandit keeps at least 10% of traffic on the control so the lift stays measurable, spreads a 5% exploration share across all variations, and cuts any variation that is clearly losing badly (40% worse than control with 95% confidence after 100 visitors). You also choose the algorithm. The default, Splitlab Adaptive, never moves a share by more than 15 points in a single update. Our multi-armed bandit vs A/B testing guide covers all five algorithms in detail.
Why you'd use it: it earns while it learns. If one variation is clearly better, most of your traffic is already on it long before a fixed test would have finished.
What it costs: precision. Because the losing variations get fewer visitors, you learn less about them. You'll know which one is best with confidence. You'll know less about by exactly how much, and less about the runner-up.
Choose Multi Arm Bandit when:
The test has a deadline: a launch week, a seasonal sale, a webinar replay that runs for five days.
You care more about revenue during the test than about a precise number at the end.
You're testing several variations (three or more) and expect some of them to be clearly worse.
The test never really ends, like an always-on headline or offer rotation you want to keep self-optimising.
Equal vs Custom vs Multi Arm Bandit at a glance
Equal Distribution | Custom Distribution | Multi Arm Bandit | |
|---|---|---|---|
Who sets the split | Splitlab, evenly | You, by hand | Splitlab, from live results |
Does it change during the test? | Only if you pause a variation | Only when you edit it | Yes, after every 20 new conversions |
Time to a confident answer | Fastest | Slower for small shares | Fast for the winner, slower for the rest |
Revenue lost to losers while testing | Highest | Capped by the shares you set | Lowest |
How precise the final lift is | Most precise | Precise, given enough time | Least precise |
Best for | New ideas, permanent decisions | Risky changes, gradual rollouts | Short windows, many variations, always-on tests |
Which one should you pick?
Ask two questions, in order.
1. Is the cost of a bad variation unacceptable? If a losing page would genuinely hurt (it's your main checkout, or the change is risky), use Custom and give the challengers a small share. You're trading speed for safety.
2. Do you need to know the exact lift, or just make more money? If you'll make a big, permanent decision from the number, like a new price, or you need a figure for a board or a client report, use Equal. If the goal is simply to send more people to whatever works, especially on a deadline, use Multi Arm Bandit.

For most funnel businesses this settles into a simple habit: Equal for new ideas, Multi Arm Bandit for promotions and anything with three or more variations, and Custom for the pages you can't afford to break.
A worked example: same test, three splits
Say a sales page gets 9,000 visitors over a test. The control earns $3.00 per visitor, Variant B earns $3.60 and Variant C earns $2.10. Same page, same visitors, only the split changes:
Split | Visitors to Control / B / C | Revenue during the test |
|---|---|---|
Equal | 3,000 / 3,000 / 3,000 | $26,100 |
Custom 80/10/10 | 7,200 / 900 / 900 | $26,730 |
Multi Arm Bandit (shares drift to about 15/75/10) | 1,350 / 6,750 / 900 | $30,240 |
The bandit makes about $4,100 more than the equal split during the test, because most visitors land on the winning page early. The custom split makes a little more than equal, because it kept the losing Variant C small, but it kept the winning Variant B just as small. The equal split ends with the strongest evidence about all three variations. The numbers are illustrative, but the pattern holds: Custom caps your downside, the bandit captures your upside, and Equal buys you the cleanest answer.
Mistakes to avoid
Switching the split mid-test. Changing from Equal to Custom halfway through means the two halves of the test were run under different rules. Splitlab warns you on live campaigns for this reason: changing a running test's settings can distort the data. Pick before launch.
Starving a challenger. A 95/5 split can take months to show anything. If a variation deserves a test, it deserves enough traffic to finish one.
Using a bandit to settle a debate. A bandit is built to make money, not to produce a precise lift for a slide. If the number matters more than the revenue, run an Equal test.
Optimising the bandit for the wrong metric. A bandit chases whatever your primary metric rewards. Point it at conversion rate and it will happily favour a cheap offer over a profitable one. Use revenue per visitor.
Forgetting that 0% is still in the test. A variation at 0% receives no new visitors but keeps its history. Pause or remove it if you're done with it.
How to set the traffic split in Splitlab
Open your test and go to the Variations step.
Next to Traffic Split, open the dropdown and choose Equal Distribution, Custom Distribution or Multi Arm Bandit. The choice saves immediately.
For Custom, click the pencil next to a variation's percentage to open the Allocate Traffic window, set the shares, lock any you want fixed, and save.
For Multi Arm Bandit, pick an algorithm. Splitlab Adaptive is a sensible start. The split then updates on its own as new conversions arrive.
Make sure your primary metric is the one you actually want to grow, ideally revenue per visitor, then launch.
Frequently asked questions
What is the best traffic split for an A/B test?
For most tests, an equal split. It gives every variation the same evidence at the same time, so you reach a confident answer fastest. Use a custom split when a losing variation would be costly, and a multi-armed bandit when earning more during the test matters more than a precise result.
Does an uneven traffic split make an A/B test invalid?
No. A test with an 80/20 split is still a valid comparison, as long as the split stays fixed and visitors are assigned at random. It just takes longer to reach the same confidence, because the smaller variation collects data more slowly.
Is a multi-armed bandit better than an equal split?
It's better at earning during the test and worse at measuring the exact difference between variations. If you need a precise lift for a permanent decision, use an equal split. If you want to send more visitors to the winner while the test runs, use a bandit.
Can I change the traffic split after a test has started?
You can, but you shouldn't unless you have to. Results gathered under one split and results gathered under another are no longer a clean comparison. If you must change it, note the date, and treat results from before and after the change separately.
What does 0% traffic mean for a variation?
The variation stays in the test with its existing data but receives no new visitors. It's a way to rest a variation without deleting it.
How does a multi-armed bandit decide how much traffic to give each variation?
It looks at each variation's performance on your primary metric and gives more traffic to the ones likely to be best. Splitlab offers five algorithms that make that call differently. Our multi-armed bandit vs A/B testing guide compares them.
Should I use a custom split to roll out a winner?
Yes, that's one of its best uses. Once a test has a winner, a custom split lets you move it from 20% to 50% to 100% while you watch for side effects like refunds or support load, before you make it permanent.
My final thoughts
The traffic split is the setting most people leave on default, and it quietly decides how much a test costs you.
My short version: use Equal when you want a clean answer you can trust and reuse. Use Custom when a new version is risky and you want to limit the damage, or when you’re rolling out a winner. Use Multi Arm Bandit when earning during the test matters more than a precise result.
None of them is wrong. Picking without thinking is. Book a demo if you want to see the three options on your own test.
Choose the split on purpose, before the test starts.
Built for revenue
Test for revenue, not just clicks.
Run A/B, multivariate, and multi-armed bandit experiments — with real revenue tied to every variation.
We’ll increase your conversion rate by 30% or you don’t pay. Terms apply.
✓
Revenue per variation
✓
A/B & MVT
✓
Multi-armed bandit
Test for revenue, not just clicks.
A/B, multivariate and multi-armed bandit experiments, with real revenue tied to every variation.


