Analysis Engine

The engine preprocesses your data
and selects the right method — automatically

No need to decide which test to use or how to clean the data. Upload results, get analysis.

Try free
Co-developed with statistics professors from Russian and British universities
All methods are peer-reviewed and published in statistical literature
Validated by IT analysts with 10+ years of A/B testing practice

Preprocessing

Data is clean before analysis even starts

The preprocessor runs automatically on every metric — analyzes the distribution and picks the right transformation. No config needed.

What the preprocessor analyzes
Distribution shape
Skewness Kurtosis Modality

Direction and strength of data bias, tail heaviness, and whether multiple peaks exist.

Spread metrics
IQR — interquartile range CV — coeff. of variation Std deviation

How widely values are spread — both absolute (IQR, σ) and relative to the mean (CV).

Outlier detection
% beyond 3 IQR % beyond 1.5 IQR % extreme values

What share of the data falls outside normal boundaries — determines severity of the outlier problem.

Additional checks
Zeros & negatives Median / mean ratio

Zero and negative values affect log applicability; median/mean gap is an independent skewness signal.

Transformation decision
Log-transformation
Strong skewness + high kurtosis + heavy tails. Compresses large values without losing them.
high skewness heavy tails
IQR3 clipping
Many extreme outliers beyond 3 IQR. Clips to normal range boundary — no data is dropped.
many > 3 IQR outliers
Winsorization 5%
Moderate outliers, elevated CV. Replaces top and bottom 5% with nearest boundary values.
moderate outliers high CV
Winsorization 1% default
Stable distribution, minimal deviations. Light protection against rare anomalies.
stable distribution
+ Delta linearization
Always applied first for ratio metrics — converts numerator/denominator into a numeric column before any transformation.
ratio metrics only
Full preprocessing documentation →

Statistical Tests

Three metric types — the right test for each

3 metric types × 2 test modes = 6 method groups. The engine picks the correct one automatically.

A / B — two variants
A / B / C — 3+ variants
Conversion
Binary: did / didn't
signup, purchase, click
Global test
Two-proportion Z-test
Pooled standard error, two-sided. Unequal group sizes handled automatically.
Confidence interval
Newcombe–Wilson method
Global test
GLM Binomial + Likelihood Ratio
Tests whether any variant differs from control across all groups simultaneously.
Pairwise post-hoc
Z-test per pair + Benjamini–Hochberg FDR
Numeric
Continuous values
session time, LTV, pageviews
Test
Welch's t-test
Unequal variance assumption — more robust than Student's t-test for real-world distributions.
Global test
Welch ANOVA
Unequal variances across groups — does not assume homoscedasticity.
Pairwise post-hoc
Games–Howell (no equal-variance assumption, already multiplicity-corrected)
Ratio Δ-method
Rate of rate
AOV, revenue/user, CTR
Preprocessing
Delta method linearization
Y − θ·X where θ = sum(numerator) / sum(denominator) in control. Converts ratio to a numeric metric.
Then
Welch's t-test on linearized values
Preprocessing
Delta method linearization
Same linearization applied per group before the multi-variant test.
Then
Welch ANOVA + Games–Howell pairwise post-hoc

Adaptive Engine

The engine adapts to your data automatically

Not a checklist of options — a system that adjusts to the situation.

Monte Carlo resampling

Activated automatically on small samples when parametric assumptions don't hold. Resampling iterations replace the normal distribution — accurate p-values and CI without over-reporting significance.

FDR correction

Benjamini–Hochberg correction applied across your selected primary metrics — not blindly to everything. Controls the false discovery rate without over-penalising.

Docs →

Sequential testing

With each daily dataset upload the Bayesian forecast is recalculated and priors are updated from the previous run. The posterior narrows as data accumulates — probabilistic sequential analysis without fixed stopping rules.

Success Forecast

Bayesian success probability alongside frequentist results — Beta–Binomial for conversion, Normal simulations for numeric. Shows P(Δ>0) and credible intervals per variant.

Docs →

Analysis Results

Everything you need in one results view

Upload your data, configure segments and significance level — the engine handles the rest.

SRM check
Preprocessing label per metric
Effect (Δ) with CI and p-value
Sample progress tracking
workbench.ab-labz.com
Filtering by Segments
device
Not selected
country
Not selected
cohort
Not selected
channel
Not selected
Alpha (α)
0.05

Analysis Results

SRM not detected
Traffic distribution matches expected. p-value: 0.4823 (α = 0.01)
Metric Treatment Effect (Δ) CI Low CI Up p-value
cr_landing — 0.42 (+0.91%) −0.21 1.04 0.2146
cr_signup — 0.18 (+0.34%) −0.33 0.69 0.5912
cr_activate P — 0.55 (+1.48%) 0.08 1.02 0.0214
cr_pay P — 0.31 (+1.12%) −0.04 0.66 0.0387
sessions log 0.01 (+0.72%) −0.01 0.02 0.2671
revenue log 0.19 (+2.14%) −0.11 0.49 0.1988

Aggregated Metrics

Metric A B Effect (Δ)
users 48 210 48 044 −166 (−0.34%)
cr_landing 22 418 (46.5%) 22 571 (47.0%) 0.42 (+0.91%)
cr_signup 25 812 (53.5%) 25 901 (53.9%) 0.18 (+0.34%)
cr_activate P 17 904 (37.1%) 18 142 (37.8%) 0.55 (+1.48%)
cr_pay P 13 244 (27.5%) 13 381 (27.9%) 0.31 (+1.12%)
sessions 91 440 92 108 668 (+0.73%)
revenue 1 284 620 1 312 104 27 484 (+2.14%)

The full engine is available on the free Solo plan

No credit card. No setup. Upload your first dataset and get results in minutes.

Register free →