How analysts test whether a new version actually beats the old one \u2014 and the traps that fake a win.