- Published
- Byline
- GoHighLevel Partners Editorial
- Filed under
- GoHighLevel Automation
A variant can win on its own metric and lose money — which is why the judging number sits further down the funnel than the step you are testing.
How to A/B test a funnel step without fooling yourself
Testing one step of a funnel is different from testing a standalone page, because the step is not the outcome. A variant can win on its own metric and lose money: an opt-in page that collects more addresses by promising more, a checkout that converts more orders by hiding a cost, a qualification step that passes more people through by asking less. Each of those improves the number directly under it and damages the number that pays you.
So the first decision in any funnel test is not what to change. It is which number you will judge it on, and that number almost always sits further down the funnel than the step you are testing. The general mechanics are in what split testing is; what follows is what changes when the page is one step of several.
Choose the step before you choose the change
Test where the loss is. That means measuring the funnel as a series of transitions first, in the way measuring conversion by stage describes, and picking the step with the largest drop that you can actually influence. A step losing a large share of visitors has room to improve; a step already performing well has little, and a test there consumes traffic for a result too small to detect.
| Step | Judge it on | The trap |
|---|---|---|
| Landing or opt-in | Qualified leads, not submissions | A broader promise raises submissions and lowers quality |
| Qualification | Booked or accepted opportunities | Asking less passes more people and wastes sales time |
| Offer page | Completed purchases that are not refunded | Overstatement converts now and refunds later |
| Checkout | Orders, and their average value | Removing a step can remove the add-on revenue with it |
| Post-purchase steps | Revenue per buyer across the whole path | A winning upsell that raises refunds on the main order |
One variable, and a real difference
Change one thing at a time, and make the change large enough to matter. Two variants differing by a button colour will need an implausible amount of traffic to separate; two variants offering genuinely different propositions will not. Small changes are not more scientific — they are just harder to detect, and funnel steps rarely carry the volume to detect them.
The changes worth testing on a funnel step are usually structural: the headline claim, what is asked for, the order of the argument, whether a price appears before or after the explanation, whether a step exists at all. Removing a step is one of the most informative tests available and one of the least run.
Decide the rules before you start
- Write down the metric, including where in the funnel it is measured, before the test goes live.
- Decide how long it runs, and let it run. Stopping when a variant looks ahead is how noise becomes a decision — early results swing wildly and settle.
- Run whole business cycles. A test spanning Monday to Thursday compares different populations than one spanning a weekend.
- Split traffic randomly and keep visitors on their variant across the whole path, so someone who sees variant A at step one does not meet variant B at step three.
- Do not change anything else during the test — not the ads feeding it, not the offer, not the follow-up sequence.
- Accept an inconclusive result. “No detectable difference” is a real finding and usually means the variable was not the constraint.
The measurement problems specific to funnels
Attribution has to survive the step. If the variant a visitor saw is not recorded on the resulting contact or order, you cannot attribute the downstream outcome to it, and the test only ever tells you about the step itself. Carry the variant through the same way you carry UTM parameters, and store it on the record.
Traffic mix moves. A funnel fed by a paid social funnel sees its audience change as the platform optimises, so a test running while campaigns are being edited is comparing variants across different populations. Freeze the campaigns or accept the noise.
Speed is a confound. If one variant is materially heavier than the other, part of any difference is load time rather than content — worth checking with what a page speed score measures before concluding anything about the copy.
Downstream steps interact. A change at the offer page alters who reaches the checkout, so the checkout’s conversion rate will move even though nothing about it changed. This is why the judging metric belongs at the end of the path rather than at each step, and it applies particularly to where an order bump belongs and downsell branching, whose performance depends on who arrived.
What to test, in rough order of likely payoff
Start with the proposition rather than the presentation. On an offer page, what is promised and to whom moves results more than layout — the questions in what an offer page has to answer are a reasonable list of candidates. On a qualification step, the number and nature of the questions is the lever, and whether a a qualifying quiz outperforms a plain form is a genuine test rather than a matter of preference.
On a form, the number of required fields is the classic test and one where the downstream metric matters most, because fewer fields reliably produce more submissions and not always more customers — the trade set out in required-field validation. And on any step, removing it entirely deserves a place in the queue.
Record what you learn
Most funnel tests produce a small effect or none, and their value accumulates only if the results are written down. Keep a log of what was tested, on which step, over what period, with what result, in something durable like a durable record outside the platform — a platform’s own testing panel rarely survives a funnel being rebuilt.
Then re-verify after any structural change. Test configuration is among the things that quietly fails to carry when a funnel is copied between accounts, as what moves and what breaks on import describes, and a test still reported as running against a variant that no longer exists produces confident, wrong conclusions. Check that both variants are actually being served before you trust any result — the most common reason a test shows no difference at all is that everyone was seeing the same page.
Picking a funnel builder: GoHighLevel vs ClickFunnels and GoHighLevel vs Systeme.io compare plan limits and usage billing side by side.
Frequently Asked Questions
What should an A/B test on a funnel step be judged on?
A metric further down the funnel than the step itself. An opt-in page can collect more addresses by promising more, and a checkout can convert more orders by hiding a cost — both improve the number directly under them while damaging the one that pays you.
Which step should be tested first?
The one with the largest drop that you can actually influence. That requires measuring the funnel as a series of transitions first. Testing a step that already performs well consumes traffic for a result too small to detect.
How big should the change between variants be?
Large enough to detect. Two variants differing by a button colour need an implausible amount of traffic to separate; genuinely different propositions do not. Small changes are not more rigorous, just harder to measure, and funnel steps rarely carry the volume.
Why should the run length be decided in advance?
Because early results swing widely and settle, so stopping when a variant looks ahead turns noise into a decision. Set the duration and the metric before the test goes live, run whole business cycles, and change nothing else while it runs.
Why does a downstream step’s conversion rate move during a test?
Because a change earlier in the funnel alters who reaches it. The later step’s rate shifts even though nothing about it changed, which is why the judging metric belongs at the end of the path rather than at each step.
What is the most common reason a test shows no difference?
That both visitors were seeing the same page. Test configuration frequently fails to carry when a funnel is copied between accounts, and a test reported as running against a variant that no longer exists produces confident, wrong conclusions. Verify both variants are being served before trusting any result.
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.