On the challenge of measuring spam prevalence externally, because historical data tends to be cleaner: https://cyber.fsi.stanford.edu/io/news/content-moderation-survivor-bias