Minimum detectable effect
The smallest change the experiment is designed to distinguish from noise. Asking the test to detect tiny lifts dramatically increases the required sample.
A/B Test Feasibility Calculator
Estimate the sample, traffic and time your landing-page test needs to detect a worthwhile improvement—before weeks of inconclusive data make the decision harder.
The thinking behind the forecast
Sample size is not a universal traffic threshold. It changes with your starting conversion rate, the smallest win worth detecting, the risk of a false winner, the risk of missing a real winner and how traffic is divided.
The smallest change the experiment is designed to distinguish from noise. Asking the test to detect tiny lifts dramatically increases the required sample.
Confidence limits false winners. Power limits missed real winners. The 95% confidence and 80% power defaults are a common planning balance.
The model converts the sample into calendar time using eligible traffic, then adds a launch check and a two-week observation guardrail.
Experiment notes
Statistical settings cannot rescue weak tracking, changing rules or a test that was too small from the beginning.
It is the smallest conversion-rate increase that would make the winning page worth implementing. Set it from commercial value and implementation effort, not from the lift you hope to see. Smaller improvements need much larger samples.
Relative lift describes the improvement compared with the current rate. Moving from 5% to 6% is a 20% relative lift but a one percentage-point increase. The calculator supports both and shows the resulting challenger target.
A test that fills its sample in a few days can still overrepresent particular weekdays, campaign conditions or sales patterns. Two complete weeks are a practical minimum guardrail for most lead-generation landing pages, though highly seasonal or considered purchases may need longer.
Not safely if that was not the stopping rule planned in advance. Repeatedly checking and stopping on a favourable result increases the chance of a false winner. Complete the planned sample and observation window, or use a properly configured sequential method.
Use a two-sided test by default because a challenger can perform meaningfully better or worse. A one-sided test needs less traffic but should only be chosen before launch when the decision genuinely cares about improvement in one direction and the downside is handled separately.
Test a larger proposition or journey change, choose a higher-volume behaviour as the primary metric, aggregate a genuinely comparable audience, or use structured qualitative research and a time-based or holdout design. Do not keep an underpowered test running indefinitely and call the direction a result.
No. The calculation addresses sample size under a two-proportion normal approximation. Validity also depends on clean assignment, reliable tracking, stable traffic, no interference between groups, a precommitted metric and stopping rule, and a test experience that remains consistent.
No. The calculator runs in your browser. A shareable result puts the values in the link itself, so only create or send one when you are comfortable sharing those assumptions.
When the answer is “do not A/B test”
We Know Growth can connect customer research, analytics, page strategy and experiment design so your team learns without waiting months for a result the site could never support.
Book a free discovery call