Statistics you can defend
Lift and confidence intervals
Two intervals, doing two different jobs. The one that settles an argument is the second.
The first interval is for a variant’s own conversion rate: the range its true rate most likely falls in. It is a Wilson interval rather than a naive one, which means it stays inside nought and a hundred per cent even at small samples, where a simpler calculation cheerfully reports a lower bound below zero.
The second is the one that decides things. The lift interval is the range for the difference against the control. If the whole range sits above zero, the variant is genuinely ahead. If it straddles zero, "no difference" is still consistent with what you have measured, however encouraging the point estimate looks.
That distinction is why both are shown. Two overlapping per-variant intervals do not tell you whether the difference between them is real, and reading them as if they did is one of the most common ways an experiment gets called early.
The lift interval is also what the rest of the platform reads. The attention queue only calls a winner when the interval stays on one side of zero, the implementation queue applies the same gate, and the automated stop refuses without it. One definition of "ahead", used everywhere.