Run experiments even with 30 users. The underlying trends you get won't change much at larger scale — only the precision does. Zero data points is always worse than 30.