Before a test starts, decide how small a change is worth detecting: the minimum detectable effect (MDE). Then work out the sample size that gives a good chance, the power (usually 80 percent), of detecting it at alpha = 0.05.
For conversion rates, statsmodels works on a standardized effect size: proportion_effectsize(new_rate, base_rate). Then NormalIndPower().solve_power(effect_size=es, alpha=0.05, power=0.8) returns the sessions needed per group. For means, divide the change by the SD and use TTestIndPower.
Leave exactly one of effect size, sample size, alpha and power unset, and solve_power solves for it.
Relative MDEs ("a 10 percent lift") and absolute ones ("one point") give very different answers on a 5 percent rate. Small changes on small rates need very large tests: that is the most common reason A/B tests come back inconclusive.