Splitlab home

Book a Demo

Splitlab home

Book a Demo

Splitlab home
The Splitlab dashboard in a desktop browser, with total revenue, recent purchases and two active A/B tests, between two versions of a landing page

November 18, 2026

By

Bishal Katuwal

·

16 min

read

How to A/B Test Landing Pages: A 10-Step Process From Baseline to Decision

November 18, 2026

By

Bishal Katuwal

·

16 min

read

How to A/B Test Landing Pages: A 10-Step Process From Baseline to Decision

In the ever-evolving business environment, achieving a competitive edge demands innovative solutions beyond conventional methods. Organizations require smart systems that can optimize workflows, enhance customer interactions, and reveal new avenues for growth. Enter Biotix, the cutting-edge AI-powered CRM management tool.

The exact 10 steps I follow to A/B test one landing page, from pulling the baseline to calling the result, with a webinar page, a $47 course page and real sample-size math.

Bishal Katuwal

Bishal Katuwal

Splitlab Founder

The Splitlab dashboard in a desktop browser, with total revenue, recent purchases and two active A/B tests, between two versions of a landing page
The Splitlab dashboard in a desktop browser, with total revenue, recent purchases and two active A/B tests, between two versions of a landing page

To A/B test a landing page, you split visitors at random between your current page and one changed version, then compare which one earns more. The process that works is ten steps: get your baseline, find where the page leaks, write a hypothesis, pick one change and one main metric, work out how many visitors you need, set the test up, check it on a phone, run it without touching it, read the result, and write down what you learned. Skip any one of those and you can get a “winner” that doesn’t make you more money.

I’ve run these tests on my own money for years. My publishing business did over $20M in revenue on more than $10M in ad spend before I sold it, and every funnel in it was split tested. This post walks through one test, start to finish, the way I run them. The examples are a webinar registration page, a book-a-call funnel and a $47 course sales page, because that’s what most of our customers sell.

If you want the reference side instead (the types of tests, what’s worth testing and the common mistakes), that’s in my landing page split testing guide. This one is the procedure.

Key takeaways

  • Start with numbers, not ideas. At least your last 30 days, and up to 90 days if you have them, tell you where the funnel loses people and what “better” has to beat.

  • Find the leak in the whole funnel, not just one page. Install Microsoft Clarity (it’s free) to see where people click, scroll and get stuck.

  • Write the hypothesis before you build anything: “Because X, changing Y will raise Z.”

  • Judge on the metric that pays: revenue per visitor first. If revenue comes later, use leads per visitor or calls per visitor until it does.

  • Decide the sample size first, run full weeks, and don’t stop early. Peeking and stopping on a good day is the most common way tests lie.

  • Test bold changes on small traffic. Small tweaks need far more visitors than most info-product pages get.


The 10 steps of one landing page A/B test: 1 baseline (last 30 days, by device), 2 find the leak, 3 hypothesis (Because X, Y raises Z), 4 one change and one metric (RPV), 5 sample size (full weeks, no peeking), 6 set it up (page, variants, metrics), 7 QA on a phone, 8 run it hands off, 9 read and decide (RPV, then guardrails), 10 document and go again

Step 1: Get your baseline numbers

Pull at least the last 30 days for the page you want to test: visitors, conversions, conversion rate, revenue and refunds. If you have longer data, use it. Up to 90 days is better, because one good or bad week can’t fool you. Then work out revenue per visitor, which is net revenue divided by visitors. Split it by device and by traffic source if you can.

Why: the baseline is what every result gets compared to, and it tells you what’s realistic. It also tells you how much traffic you have, which decides what kind of test you can afford in step 5.

For example, say your $47 course sales page had 15,000 visitors from Meta ads last month and 600 sales. That’s a 4% conversion rate, $28,200 in sales, and after 12 refunds ($564), $27,636 net. Revenue per visitor is $1.84. (These are illustrative numbers, and I use them in the worked example below.)

Know what normal looks like for your kind of page, too. In the funnels I run, a $27 product usually converts around 5% and a $97 product 1 to 3%. A simple name-and-email opt-in page should do 20% to 25%, and 30% or more is good. If you’re far below that, you probably have something broken to fix before you test anything. My squeeze page conversion rate post goes deeper on opt-in numbers.

Step 2: Find the leak

Before you decide what to change, find where people stop. Look at the whole funnel first, then the page.

Take a typical book-a-call funnel:

Ad → opt-in landing page → video → application page → pick a time → thank-you page

Write down the conversion rate from each step to the next. Say 35% of visitors opt in, 60% of those watch the video and click through, but only 8% of them fill in the application. The opt-in page isn’t your problem. The video-to-application step is the leak, so that’s where your next test goes. Testing the opt-in headline would be fixing the part that already works.


Find the leak in a book-a-call funnel: ad, opt-in page 35% opt in, video 60% click through, application page only 8% apply, pick a time 70% book, thank-you page. The biggest drop is between the video and the application, so that is the step to test first

Then look inside that page:

  • Heatmaps and recordings. Install Microsoft Clarity. It’s free, and it shows you where people click, what they ignore, how far they scroll and where they get confused, with recordings of real visits. Ten minutes of watching recordings will give you better test ideas than an hour of guessing.

  • The click from page to checkout. Do people click “Buy” and then not pay?

  • Device split. Does the page convert half as well on phones as on desktop?

  • Traffic source. Does your email list buy while cold ad traffic doesn’t?

Why: a test aimed at the leak can move the number. A test aimed at something people never see can’t. If most visitors leave on the first screen, testing the FAQ at the bottom is a waste of a month.

In my paid-ad funnels, 90% or more of the traffic is on phones, so I always start by looking at the page on a phone. My landing page CRO checklist is a quick way to spot fails here. Anything broken (a form that errors, tracking that doesn’t fire) gets fixed, not tested.

Step 3: Write a hypothesis

Write one sentence before you build anything: "Because X, changing Y will raise Z."

  • X is what you saw in step 2.

  • Y is the one change you’ll make.

  • Z is the metric you expect to move.

Why: a written reason stops you from testing random ideas, and it means you learn something even when the test loses. If the hypothesis was “the form is below the first screen” and moving it up doesn’t help, you’ve ruled out a cause. You know to look somewhere else.

And sometimes your “sure thing” loses. On one of my $5 book funnels, the order form said no money back guarantee. Every funnel I saw had one, so I was sure adding a 30-day guarantee would win. It lost, badly, in two different funnels. The opt-in step on the order form dropped from 69% to 44%, and checkouts from 47% to 31%. If I hadn’t tested it, I’d have shipped it everywhere.

I Thought This Would Increase Sales… It Didn’t

Here’s one for a webinar registration page where most ad clicks come from phones and the form sits below the first screen:


Example hypothesis card for a webinar registration page: because the form sits below the first screen on phones, where most ad clicks land, changing the first screen to show the date, the promise and a Save my seat button will raise registrations per visitor, with the guardrail that show-up rate must not drop

Notice the guardrail. A page that gets more registrations by hiding the webinar date might get more sign-ups who never show up. Writing the guardrail down now stops you from calling that a win later.

Step 4: Pick one change and one main metric

Change one thing per test, or one clear idea. "A new headline that repeats the ad’s promise" is one idea, even if it touches two lines of copy. "New headline, new price, new video and a new button color" is four, and when it wins you won’t know why.

Then pick the one metric that decides the test. For almost everyone, that’s revenue per visitor. But sometimes the revenue comes late, like a sales call that closes two weeks after the opt-in. Then you judge on the step that leads to it, per visitor:

  • Pages that sell (sales pages, order forms): revenue per visitor (RPV). It counts both how many people buy and how much they spend, minus refunds.

  • Opt-in and webinar registration pages: revenue per visitor if your tool can credit the later sale back to the page the visitor saw. If not, leads per visitor, with a guardrail on show-up rate or later purchases.

  • Book-a-call funnels: revenue per visitor once deals close. Until then, calls booked per visitor, because a booked call is the step right before the money.

Add one or two guardrails, metrics the winner must not hurt: refund rate on sales pages, show-up rate on webinar pages, conversion rate if the main metric is revenue.

Why RPV over conversion rate: a cheaper price or a pushier page can raise conversion rate and still lose money. I go through real cases in revenue per visitor vs conversion rate. Why pick it now: if you choose the metric after you see the results, it’s too easy to call whatever moved a win.

If your test is a headline, my post on how to test landing page headlines covers how to write the variants.

Step 5: Work out the sample size and the duration

This is the step most people skip, and it’s why so many "winners" don’t hold up. You need to know, before launch, how many visitors each version needs to tell a real difference from luck.

You need three inputs:

  • Baseline conversion rate from step 1.

  • The smallest lift worth finding. Be honest: on small traffic, only big lifts are findable.

  • How sure you want to be. The usual standard is 95% confidence and 80% power, which means a 5% chance of calling a winner that isn’t real, and an 80% chance of catching a real lift of the size you picked.

Here’s what that standard gives, using the normal two-proportion sample size formula. You can check any of these in Evan Miller’s sample size calculator; his rounding can differ by a little.

  • Webinar registration page at 20%, looking for 25% (a 25% relative lift): about 1,100 visitors per version, 2,200 in total.

  • $47 sales page at 4%, looking for 5% (a 25% relative lift): about 6,750 per version, 13,500 in total.

  • Same sales page, looking for 4.4% (a 10% relative lift): about 39,500 per version, 79,000 in total.

That last line is the reason I tell small funnels to test bold changes. At 500 visitors a day, finding a 10% lift on that sales page would take over five months. Finding a 25% lift takes about four weeks.

These numbers are for conversion rate. When the main metric is revenue per visitor, treat them as a floor. Revenue jumps around more than a yes/no conversion does, especially with order bumps and upsells, so RPV tests usually need at least as many visitors, often more.

Then turn visitors into time:

  • Divide the total by your daily traffic, then round up to full weeks. People buy differently on a Saturday than on a Tuesday, and a test that covers only part of a week can be skewed by the days it caught.

  • Run at least two full weeks, even if you hit the number sooner, so one unusual day can’t decide the result. With low traffic, plan on a month or more.

  • Write the end date down.

Don’t peek and stop early. Early in a test, the lead swings wildly. Evan Miller shows that if you peek at a test ten times, what you think is 1% significance is actually just 5%. In his example of stopping the first time a test hits 5% significance (or calling it off after 150 observations), the real false-positive rate is 26.1%. His fix is simple: decide the sample size in advance and wait until the test is over. I explain how early stopping produces "winners" that never show up in the bank account in why A/B test winners don’t make money.


Example chart of a variant's revenue-per-visitor lead over four weeks: it looks like a winner on day 3, a loser on day 6, and settles slightly ahead by the planned end. Below, Evan Miller's figure: stopping at the first 5% significant peek gives a 26.1% real false-positive rate

Step 6: Set the test up

In Splitlab, a test has three setup steps. I’ll use the $47 sales page.

Pages. Tell Splitlab which page runs the test with a URL match: the URL of your sales page in ClickFunnels, HighLevel or wherever it lives. Why: the URL rule is what makes sure only visitors to that page enter the test.


Splitlab app, test setup Pages step: Page Selection with a URL matches rule pointing at a ClickFunnels sales page URL

Variants & Traffic. Your current page is the control. Add the variant and build the change in the visual editor (for this test, the new headline).

Then choose the split. Equal Distribution, 50/50, is the clean choice for a test you want to learn from. Custom Distribution lets you send less traffic to a risky variant. Multi Arm Bandit shifts traffic toward the version earning more while the test runs, which suits times when earning during the test matters more than a clean read. I compare those options in multi-armed bandit vs A/B testing.

Metrics. Revenue per visitor, average order value and conversion rate come built in. Star RPV as the primary metric, then add your guardrails, such as refunds. Why: the starred metric is the one that decides the test, and you’ve now committed to it before any data comes in.


Splitlab app, Metrics step: Revenue per Visitor (net revenue divided by unique visitors) starred as the primary metric, with Average Order Value, Conversion Rate and a custom Order Page Visit metric

One more piece matters for pages that sell: the sale usually happens on another step or another platform. Splitlab’s integrations with ClickFunnels 2.0 and Classic, GoHighLevel, Cal.com, Calendly and stores credit each order, upsell and booking to the variant the visitor saw. Without that, the variant that sold more gets no credit. If you build in ClickFunnels, the full setup is in how to A/B test ClickFunnels 2.0. On HighLevel, see GoHighLevel split testing.

Step 7: QA both versions on a phone

Before you send traffic, open both versions on your own phone and check:

  • You see the right version. Use a private window to get assigned, and reload to make sure you stay on the same version.

  • Nothing flickers. The original shouldn’t flash before the variant appears.

  • The change looks right on a small screen. A headline that fits on a laptop can push the button off the first screen on a phone.

  • The buy button works, and the sale is tracked. Make a real test purchase (or a 100% coupon), then check it shows up under the right variant. Refund it after.

Why: a broken variant doesn’t lose because the idea was bad. It loses because it was broken, and you’ll throw away a good idea. And with most paid traffic on phones, a variant that only works on desktop is broken.

Step 8: Run it and leave it alone

Turn it on and don’t touch anything until the end date. That means:

  • No edits to either version mid-test.

  • No new traffic sources halfway through. If you add a new ad audience in week 2, you change who sees the test.

  • No early stopping, as in step 5, unless something is broken.

Why: every change mid-test mixes two experiments into one, and the result stops meaning anything. If you must fix a bug, fix it and restart the count.

Here’s a real one from my funnels. On an upsell page I tested a version with a short video against the same page without it. I let it run from September 25 to October 29 without touching anything. The video version converted at 43% against 38%, and that one change added over $2,000 in revenue on that upsell alone. In the first few days the lead flipped back and forth, which is exactly why you don’t stop early.

Step 9: Read the result and decide

At the end date, read the numbers in this order:

  1. The main metric. Is the variant’s revenue per visitor higher, and is the difference past the confidence bar you set before launch?

  2. The guardrails. Did refunds, show-up rate or anything else you named get clearly worse?

  3. The segments, only as a check. Did it win on phones and desktop, or only on one? Don’t go hunting for a segment where it won. With enough slices, something will always look like a winner by chance.

Then decide. There are only three outcomes:

  • Clear win, guardrails fine: ship the variant.

  • Clear loss: keep the control. You’ve still learned the hypothesis was wrong.

  • No clear difference: keep whichever is simpler, and test a bolder change next. "No difference" usually means the change was too small for your traffic.

A worked example, start to finish

These numbers are illustrative, not from a real client. They’re built to show the steps with math you can check.

  • Page: a $47 video editing course sales page, cold traffic from Meta ads, about 500 visitors a day.

  • Baseline (step 1): 15,000 visitors, 4.0% conversion, $1.84 revenue per visitor.

  • Leak (step 2): most phone visitors leave on the first screen. The ad promises "Edit your first Reel in 10 minutes", but the page opens with "Welcome to the Creator Course".

  • Hypothesis (step 3): because the headline doesn’t match the ad, changing it to repeat the ad’s promise will raise revenue per visitor. Guardrail: refund rate.

  • Size (step 5): to catch a lift from 4% to 5%, we need about 6,750 visitors per version. At 500 a day that’s 27 days, so four full weeks, 7,000 per version.

  • Run (step 8): in the first week the variant was ahead, then behind. Nobody touched it.

After four weeks:

  • Control: 7,000 visitors, 280 sales, 6 refunds. Net revenue $12,878, so $1.84 per visitor.

  • Variant: 7,000 visitors, 343 sales, 10 refunds. Net revenue $15,651, so $2.24 per visitor, about 22% more.

  • Conversion: 4.0% vs 4.9%. A standard two-proportion test puts that gap at p ≈ 0.01, inside the 95% bar we set.

  • Refunds (guardrail): 2.1% vs 2.9% of sales. With only 16 refunds in total, that difference could easily be chance, so it doesn’t block the decision, but it goes on the watch list.


Illustrative result read-out for a $47 course sales page after 4 weeks with 7,000 visitors per version: revenue per visitor $1.84 for the control vs $2.24 for the ad-promise headline variant, conversion 4.0% to 4.9% (clears the 95% bar), refunds 2.1% to 2.9% (too few to tell). Decision: ship the variant and keep watching refunds for 30 days

Decision: ship the variant and check refunds again in 30 days. The price didn’t change, so here revenue per visitor and conversion rate agree. They often don’t. If the variant had been a $37 price, conversion rate could have gone up while revenue per visitor went down, and RPV would have caught it.

Step 10: Document it and go again

Write a short record of every test, even the losers:

  • The hypothesis, word for word.

  • Dates, traffic sources and visitors per version.

  • The result on the main metric and the guardrails.

  • The decision, and what you’ll test next.

Why: after a dozen tests, this log is worth more than any single winner. It tells you what your buyers respond to (proof, clarity, price framing) and stops you from re-testing ideas that already lost. Then go back to step 2, find the next leak, and start again. Each winner becomes the new control.

How Splitlab helps

Splitlab runs A/B, multivariate and multi-armed bandit tests on funnel pages, including pages built in ClickFunnels and HighLevel. Every test has revenue per visitor, average order value and conversion rate built in, and the integrations credit each order, upsell and booking to the variant the visitor saw, so steps 4, 6 and 9 of this process are set up for you. We’ll increase your conversion rate by 30% or you don’t pay (terms apply). See the features or book a demo.

FAQ

How long should a landing page A/B test run?

Until it reaches the sample size you worked out before launch, rounded up to full weeks, and at least two full weeks. For a $47 sales page at 4% looking for a 25% lift, that’s about 6,750 visitors per version. At 500 visitors a day, that’s four weeks.

How much traffic do I need to A/B test a landing page?

It depends on your conversion rate and the size of lift you want to find. At 95% confidence and 80% power, a 20% opt-in page needs about 1,100 visitors per version to detect a jump to 25%. A 4% sales page needs about 6,750 per version to detect 5%. Smaller lifts need far more, which is why low-traffic pages should test bold changes.

Can I stop an A/B test early if one version is clearly winning?

Not if you want to trust the result. Leads swing a lot in the first days, and stopping the first time a test looks significant raises the false-positive rate well above what the tool reports. Stop early only if a variant is broken.

What should I measure in a landing page A/B test?

For pages that sell, revenue per visitor, with refunds as a guardrail. When the revenue comes later, measure the step before it per visitor: leads per visitor on opt-in pages, calls booked per visitor in book-a-call funnels.

What tool should I use to find what to test?

Start with Microsoft Clarity. It’s free and shows heatmaps, scroll depth and recordings of real visits, so you can see where people get stuck before you decide what to change.

Can I test more than one change at once?

You can test one idea that touches several elements, like a headline and subheadline rewritten around one promise. If you want to know which of several separate changes caused the result, use a multivariate test, which needs much more traffic.

My final thoughts

Ten steps sounds like a lot for one test. In practice most of them take minutes, and the ones people skip (the baseline, the sample size and leaving the test alone) are the ones that decide whether the result is real.

If you only adopt two habits from this process, make them these: write the hypothesis before you build anything, and decide the sample size before you launch. Those two stop most of the fake winners I see.

Then keep the log. After a dozen tests, it tells you more about your buyers than any single winner. If you’d like help setting up your first test, book a demo and we’ll walk through it on your own page.

One honest test a month beats ten rushed ones.

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.


Start optimizing for revenue today not just conversion rates.

Run experiments, track revenue per variation, and scale what actually makes you money.

We’ll increase your conversion rate by 30% or you don’t pay. Terms apply.


Start optimizing for revenue today not just conversion rates.

Run experiments, track revenue per variation, and scale what actually makes you money.

We’ll increase your conversion rate by 30% or you don’t pay. Terms apply.


Start optimizing for revenue today not just conversion rates.

Run experiments, track revenue per variation, and scale what actually makes you money.

We’ll increase your conversion rate by 30% or you don’t pay. Terms apply.