generalizing in a small population is often more accurate than generalizing over a giant userbase
That's actually not true at all. In a small population each individual node has much more weight. For example, in a population of 10, you can't look at 3, see that one is a guy and two are girls, and assume that 2/3 of the population are girls, even though you took a 33% sample. However, in a population of 1,000,000 if you find that in a sample of 333,333 that 2/3 are girls and 1/3 are guys, you can much more accurately assume that 2/3 of the total population are girls and 1/3 are guys.