A growing collection of practical decision tools built with analytics, statistics, and AI.
Each tool runs entirely in your browser. No account, no installation, no data leaves your machine.
https://gracege.com/tools/ab-test-calculator
Four calculators for common experimentation decisions:
- Mean Comparison — Welch's t-test for continuous metrics (revenue, order value, time on page)
- Proportion Comparison — z-test for conversion rates, with Newcombe (Wilson score) confidence intervals
- Sample Size — separate modes for A/B tests (baseline + MDE + power) and finite-population surveys
- Statistical Power — work out whether your test can actually detect the effect you care about
Each tab includes a "how to read this & common mistakes" section covering the failure modes that matter: peeking mid-test, confusing statistical significance with practical importance, post-hoc power, multiple comparisons.
- Confidence intervals for proportions use the Newcombe method (Wilson score hybrid) rather than the simple Wald interval. Wald breaks down with small samples or rates near 0%/100% — it can produce impossible negative bounds. Newcombe doesn't.
- The significance test uses a pooled z-test (assuming equal proportions under the null); the interval uses the unpooled Newcombe approach. These are built on different frameworks, so in extreme small-sample cases they can appear to disagree. That's a known property, not a bug.
- Welch's correction is applied by default for mean comparison — no assumption of equal variances.
- All statistical functions (erf, incomplete beta, t-distribution CDF, Wilson score) are implemented directly in plain JavaScript. No external libraries, no CDN dependencies. The files work offline.
Most A/B test calculators give you a number and stop. The harder part is knowing what the number means and when not to trust it.
Built by Grace Ge. Building systems for better decisions.
More tools coming.
MIT — see LICENSE.