Start with observed data
Results are built from recorded exposed visitor cohorts and matching outcome events. Cohort dates use UTC. Counts can be empty when no eligible events have been recorded; Varypilot does not populate illustrative performance data.
Interpret the selected goal, date range, currency, allocation version and conversion window together. A row from an incomplete reporting page must not be presented as a complete experiment total.
Conversion analysis
The current binary comparison uses a normal approximation for the difference in conversion proportions and 95% intervals. Per-arm conversion intervals use the Wilson method. A comparison is unavailable when either arm has no visitors or insufficient success/failure counts.
Intervals and p-values are pairwise; a multi-variant design needs its own multiple-comparison plan. An interval is uncertainty about an estimate, not a guaranteed future uplift.
Revenue analysis
Revenue comparisons use a visitor-level normal approximation and require at least 30 exposed visitors per arm. Revenue sums and sum-of-squares are needed for variance. Heavy-tailed revenue, immature outcomes and weak sample sizes can make a result unreliable.
Make the decision
Check the planned sample size, planned duration, primary metric and data-quality diagnostics before deciding. Statistical significance alone does not establish business value, a healthy integration or a finished stopping horizon.