What is the minimum detectable effect?

The smallest change between two variants that the test is sized to catch. Here you enter it as a relative lift: 50% on a 2% sale rate means telling 2% apart from 3%. Halve it and the traffic you need roughly quadruples.

Why 95% confidence and 80% power?

They are the usual planning defaults. The first caps how often a test crowns a winner that is only noise. The second sets how often a lift of the size you entered gets caught. Raising either one makes the test longer.

Should a presell test count clicks or sales?

Sales, whenever you can afford the traffic, because the presell exists to sell. Click-through to the offer is a quicker read and a fair first cut between very different headlines or formats.

How long should a test run once it reaches its sample?

Until it has covered at least a full week of traffic, even when the count arrives sooner. A test that saw only weekdays or only a weekend has measured part of your audience.

Testing and measurement

Split-test sample size planner

At 95% confidence and 80% power, each variant needs about 7.84 × [p₀(1 − p₀) + p₁(1 − p₁)] ÷ (p₁ − p₀)² visitors, where p₀ is your current rate and p₁ the rate you want to be able to detect. At a 2% sale rate with a 50% lift target, that comes to about 3,800 ad clicks and 76 sales per variant. Put your own rates into the planner below.

By Presellbook editors. Published 13 October 2026.

Summarize with:

Drawn diagram of two page variants side by side, each with a column of dots reaching the same marked line, and an hourglass between them

Most split tests on presells get called too early. One variant pulls ahead after a few hundred clicks, the buyer pauses the other, and a week later the "winner" earns no more than the page it replaced. The fix is to decide before launch how much traffic each arm gets, and this planner works that number out from two things you already know: the rate you convert at now and the smallest lift worth finding.

It gives a planning estimate from a standard formula. Whether to test, what to test first and how to rotate variants in your tracker are covered in landing page split testing for presells.

The planner

What the test is judged on

Sales per ad click Click-through to the offer

Baseline rate (%) Sales ÷ ad clicks on the presell you run now. Between 0.1 and 50.

Smallest lift to detect (%) Relative change: 50 means 2% rising to 3%. Between 5 and 200.

Variants in the test Traffic is split evenly between them.

Ad clicks a day (optional) Total daily clicks into the test, all variants together. Leave it empty to skip the day count.

Fixed for every result: 95% confidence, two-sided (z ≈ 1.96), and 80% power (z ≈ 0.84).

3,819Ad clicks each variant needs
7,638Ad clicks across the whole test
76Sales per variant at your baseline
NoneDays to finish: add daily clicks

Send 3,819 ad clicks to each of the 2 variants before you compare them, and do not stop sooner because one looks ahead.

The planner splits traffic evenly and gives every arm the same count. With three or four variants it makes no allowance for comparing several challengers against one control, so treat its figure as the floor for each arm.

How the numbers are worked out

Fixed-horizon A/B tools compare two conversion rates with a two-sided z-test. Optimizely's notes on sample size calculations for experiments state that Optimizely uses this Wald test in its fixed-horizon frequentist mode. The same notes give the per-group formula for a yes-or-no metric, and this planner uses it:

p₀ is your baseline rate and δ the relative lift you entered
Δ = δ × p₀, the absolute gap, so p₁ = p₀ + Δ
V = p₀(1 − p₀) + p₁(1 − p₁)
n per variant = (z_α + z_β)² × V ÷ Δ², rounded up

Each p(1 − p) term is the variance of a single visit that either converts or does not, which Evan Miller sets out in How Not To Run an A/B Test. Optimizely's own estimator refines the relative-lift case further. This page keeps to the absolute-difference version so you can check every result by hand.

Confidence and power are the two dials in the formula. Unbounce's A/B test duration calculator describes a 95% confidence level as being 95% sure the result is reliable, and power as the likelihood that the test detects a real difference when one exists. Optimizely's defaults of α = 0.05 and β = 0.2 give the 1.96 and 0.84 used here, and (1.96 + 0.84)² is the 7.84 in the formula above. Google Ads uses the same threshold in its experiments: it reports a variation as statistically significant when it is at least 95% likely to have caused the change.

Worked example: a sales test on a 2% baseline

Your presell turns 2% of ad clicks into sales and you want to know if a new lead can lift that by half, to 3%. Δ is 0.01. V is 0.02 × 0.98 + 0.03 × 0.97 = 0.0487. Multiply by 7.84 and divide by 0.0001 and you get 3,818.1, which rounds up to 3,819 ad clicks per variant and 7,638 for a two-arm test. At the 2% baseline each variant collects about 76 sales along the way, which falls within the sixty to eighty conversions per variant that Binom's course on split testing landers and offers recommends for this scenario.

Lower the target to a 20% lift (2% to 2.4%) and the count jumps to about 21,000 ad clicks per variant. At a low sale rate, small improvements take a lot of traffic to prove.

Worked example: a click-through test

Switch the planner to click-through. Say 25% of presell visitors click to the offer page and you want to detect a 20% lift, to 30%. Δ is 0.05 and V is 0.1875 + 0.21 = 0.3975, so each variant needs 1,247 visitors, about 312 of whom click through at the baseline rate.

That is a third of the traffic the sales test needed while hunting a smaller lift. The catch is what it measures: a presell that sends more readers to the offer can send colder ones, so a click-through winner still has to earn its place on sales per ad click. The EPC and max CPC calculator shows how both rates combine into earnings per click.

Checking the result against presell rules of thumb

Affiliate playbooks set their own minimum counts, and comparing them with the planner's output is a quick sanity check. CB-Stats' guide to A/B testing affiliate landing pages quotes roughly 200 to 300 clicks per variation before a change in click-through rate is worth trusting. Binom's sixty to eighty conversions apply to a sales test at 2% with a 50% lift. Both are rules of thumb from people who run traffic, and neither comes out of the formula. If your inputs differ from the case they were written for, the formula's answer is the one that fits your funnel.

Where the two disagree, take the larger. An AffiliateFix member who wrote up the truth about most split tests pre-committed to 2,500 clicks per side and watched the variant leading at around 200 clicks finish behind. Sizing the test up front is what stops that early leader from being declared the winner.

Turning the sample into a calendar

Enter your daily ad clicks and the planner divides the whole test's traffic by them. ROAS365's guide to landing page A/B test sample size makes the same point about order: decide the sample size and the end date before launch, then work the days out from daily traffic.

A test that reaches its sample in a few days should still run for a full week. Unbounce recommends at least one to two weeks so the sample covers how visitors behave on different days, and a test that reaches its count by Wednesday has not yet seen a weekend.

A test that needs months is badly designed for its traffic. If the calendar runs into months, test a bigger change (a new angle or format instead of a button colour), raise the lift you are looking for, or judge the test on click-through first and confirm the winner on sales later. Stopping halfway because a dashboard shows significance throws away the protection the plan bought. Evan Miller's advice is to decide the sample size in advance and wait until the experiment is over before believing the "chance of beating original" figure.

Our take

We'd start every presell test by writing down three things: the per-variant count from this planner or your own floor, whichever is larger, the end date, and the metric that decides it. We'd judge the result on earnings per ad click once the sample is in, and treat click-through as the early warning.

On low-volume sources we'd run two arms only and pit the control against a page that differs a lot, such as a new angle or a different format. Each extra arm needs the full per-variant count on top of the others, and proving a headline tweak at a 1% sale rate can take longer than the offer runs.

The rest of the routine (what to test first, rotating arms behind one ad URL, reading results by source) is in the guide to landing page split testing for presells, and presell-to-sale conversion by traffic source sets up the tracking that makes the counts trustworthy. More measurement guides sit on the testing and measurement hub, and all presell guides are listed newest first.