They probably have some kind of text classifier that's trained on `tweet_contents -> engagement`, with some suitable engagement metric (dwell time, likes, replies, etc).