- Published
- Byline
- GoHighLevel Partners Editorial
- Filed under
- GoHighLevel Automation
The allocation decides how fast you get an answer and what the experiment costs while you wait, and the two pull against each other.
Split test traffic distribution: how much of the audience each variant gets
Once a test is running, the allocation decides two things: how quickly you get an answer, and how much the experiment costs while you wait. An even split answers fastest. An uneven one protects the business from a variant that might be worse, and pays for that protection in time. Neither is correct in general, and choosing without noticing the trade-off is how tests end up running for months without resolving.
Setting the test up is what split testing is and is not and choosing what to change on a step; ending it is ending a test and declaring a winner. This is the middle part, which gets the least attention and quietly determines whether the other two were worth doing.
Even is the default because it is the fastest
A test resolves when the smaller arm has accumulated enough conversions, so the arm receiving less traffic is what sets the clock. Splitting evenly makes both arms grow at the same rate and reaches a conclusion sooner than any uneven allocation of the same traffic. Unless you have a specific reason to protect one side, even is the right starting point.
The reason people drift away from it is discomfort: the new variant might be worse, and half the audience is seeing it. That instinct is reasonable, and the cost of acting on it is a longer test — which also means longer exposure to whatever the current page is doing badly.
Three allocations and what each is for
| Allocation | Use it when | What it costs |
|---|---|---|
| Even | The default; both variants are acceptable to show | Nothing — this is the fastest option |
| Weighted toward the control | The variant is risky, or the page carries high-value traffic | A longer test, often much longer |
| Ramped: small, then even | You want an early sanity check before committing | Complexity, and a mixed sample if you analyse it carelessly |
Weighting hurts more than the numbers suggest. Moving from an even split to roughly ninety against ten does not make the test ten per cent slower; it makes it several times slower, because the small arm now needs the same number of conversions from a fraction of the traffic. If the reason for weighting is fear rather than a concrete risk, the honest answer is usually to run even and watch it closely.
Ramping, and how to do it without poisoning the data
Starting a variant at a small share for a day is a reasonable safety check — it catches a broken page, a form that does not submit, or a checkout that fails, before many people see it. The rule is that the ramp is a smoke test, not part of the measurement.
Once the variant is confirmed to work, move to an even split and start the measurement window from there. Blending the ramp period into the result mixes two different sampling rates and, worse, mixes a period when the page may have been broken with one when it was not.
Assignment has to be sticky
A visitor who sees one variant and returns to see the other has been in both arms, and their behaviour cannot be attributed to either. Worse, they have had an inconsistent experience: a price or a headline that changed between visits reads as a site that cannot be trusted.
Assignment is normally stored in the browser, which means it is subject to the same limits as anything else stored there. It does not survive a different device, it does not survive a cleared browser, and it does not cross a domain boundary — so one funnel on several domains will re-assign a returning visitor arriving on the other host. Where a consent banner and its stored decision gates storage, a visitor who declined may be re-assigned on every visit. None of these invalidate a test, but they all add noise, and they explain why arm sizes rarely land exactly on the ratio you set.
Do not change the split mid-test
Changing the allocation while a test runs makes the two arms cover different periods in different proportions, so any difference between weekday and weekend traffic, or before and after a campaign, is baked into the comparison as if it were an effect of the variant. The result becomes uninterpretable rather than merely noisy.
If the split has to change — because the variant is clearly broken, or a business decision forces it — treat that as the end of the test rather than an adjustment to it. Record what happened, and restart cleanly if the question still matters.
More than two variants costs more than it looks
A three-way test with equal shares gives each arm a third of the traffic, so each grows at two-thirds the rate it would in a two-way test, and you need more data overall to separate three things than two. On the traffic most agency funnels actually have, a three-way test is often a test that never resolves.
Run the strongest challenger against the control, decide it, then run the next. Sequential two-way tests reach conclusions; a wide test reaches the end of the quarter. And keep the changes off anything shared — a variant built into a global section shared across pages appears everywhere the block does, so the change is not confined to the page under test.
Checking that the split is behaving
- Compare actual arm sizes against the configured ratio after a day. A large, persistent gap points at assignment rather than chance.
- Confirm both variants are being served on the live URL, not only in the builder.
- Check both arms on a phone, since a variant that fails on mobile shows up as an allocation problem first.
- Confirm the outcome event fires identically for both — the thank-you step counting on one variant and not the other produces a false winner.
- Verify neither variant is materially slower, because a page speed score differences between arms measure load rather than content.
- Watch that a campaign has not started mid-test — a paid social funnel changing the traffic mix affects both arms but not necessarily equally.
For reading the result rather than the split, reading a funnel exploration shows where each arm lost people, tracking clicks that leave the site covers the exits, and pipeline conversion keeps the judgement anchored to a booked outcome rather than a click. Where the test lives inside an automation rather than on a page, split testing inside a workflow covers the same allocation question in that context.
Frequently Asked Questions
What is the best traffic split for an A/B test?
Even, unless there is a concrete reason to protect one side. A test resolves when the smaller arm has accumulated enough conversions, so an even split reaches a conclusion sooner than any uneven allocation of the same traffic.
How much slower is a weighted split?
Much slower than the ratio suggests. Moving from even to roughly ninety against ten does not add ten per cent to the duration — it multiplies it, because the small arm needs the same number of conversions from a fraction of the traffic.
Should I start the variant at a small percentage?
As a smoke test, yes: a day at a small share catches a broken page or a form that does not submit before many people see it. Then move to even and start the measurement window there, rather than blending the ramp period into the result.
Can I change the split while the test is running?
Avoid it. Changing allocation mid-test means the arms cover different periods in different proportions, so weekday and weekend differences get baked into the comparison as if they were effects of the variant. If it must change, treat that as the end of the test.
Why are my arm sizes not exactly the ratio I set?
Assignment is normally stored in the browser, so it does not survive a different device, a cleared browser, or a crossing between domains, and it may not persist where a consent banner gates storage. Small persistent gaps are normal; a large one points at assignment rather than chance.
Will a returning visitor see the same variant?
They should, and the stored assignment is what makes that happen. A visitor who sees both has been in both arms and cannot be attributed to either, and they have also experienced a headline or price that changed between visits, which reads as a site that cannot be trusted.
Can I test three versions at once?
You can, and on the traffic most agency funnels have it usually means a test that never resolves. Each arm grows more slowly and more data is needed to separate three things than two. Run the strongest challenger against the control, decide, then run the next.
We build your GoHighLevel account for you
CRM, pipelines, funnels, SMS and automation — configured for your business, not left for you to figure out.
- Setup (one-time)$850
- Ongoing (per month)$38.80
No obligation quote · reply within 1 business day
This site is independently operated and is not officially affiliated with, endorsed by, or sponsored by HighLevel LLC. We do not receive referral commissions — this is our own flat-fee implementation service.