A/B Test Design Workbench
Plan feature-launch experiments across metric types, power, duration, multiple testing risk, A/A checks, and CUPED variance reduction.
The test math changes depending on whether the metric is a proportion or an average.
Current conversion rate before launching the new feature.
Minimum meaningful change worth detecting.
False-positive rate for one comparison.
Beta is 20% at this power target.
Higher correlation means pre-experiment data can remove more noise.
For a 0.5 pts / 5.0% effect at 5% alpha and 80% power, the clean two-variant test needs 57,645 users per variant.
Because you are checking 12 comparisons, the chance of at least one false positive rises to 46.0%. Bonferroni keeps the family-wise error rate controlled by using 0.42% alpha per comparison.
CUPED reduces noise when pre-experiment behavior predicts the metric. With correlation 0.40, it saves about 4 days after the multiple-testing adjustment.
7 days
- Sample / variant
- 35,000
- Detectable effect
- 0.6 pts
- Relative MDE
- 6.3%
Fast, but usually underpowered and noisy.
14 days
- Sample / variant
- 70,000
- Detectable effect
- 0.4 pts
- Relative MDE
- 4.5%
Often a practical minimum with two weekly cycles.
28 days
- Sample / variant
- 140,000
- Detectable effect
- 0.3 pts
- Relative MDE
- 3.2%
More stable read with seasonality and novelty risk reduced.
56 days
- Sample / variant
- 280,000
- Detectable effect
- 0.2 pts
- Relative MDE
- 2.2%
Best for small effects or low-traffic segments, but slower learning.
Correlation of 0.40 reduces variance by 16%, saving about 4 days at current traffic.
Even with no real effect, repeated checks can produce false positives.
Expected false positives
0.50
At least one false positive
40.1%
- Instrumentation is stable and logging every assignment.
- Randomization happens before users can self-select into variants.
- Metric definitions are frozen before the test starts.
- Eligibility, exposure, and analysis populations are clearly defined.
- The team agrees not to keep slicing until something looks significant.