And if it were 1,000? 5,000? 25,000? You're still talking epsilon-level percentages here.

"Epsilon-level percentages" isn't really a useful concept. As a sample size grows, phenomena become significant at smaller percentages. The set of all tweets is gigantic, so we can still make significant observations based on a "small" subset.

The point of this article is not to say "look at how many racists there are on Twitter." The point is to say "Let's look at the racist people on Twitter. What is the geographical distribution?"

395 geotagged tweets is absolutely enough to start making some interesting observations. Those observations can only be about relatively broad regions at that sample size, but they are interesting nonetheless.

In some previous twitter work, I've seen geo-located tweets at about 2-3% of all tweets (assuming no bias), so charitably 19750 racist tweets here.

Which is to say, I agree with you.