Mr Agrawal said the company used a combination of public and private data to determine which accounts were real, reviewing random samples every few months. The margins of error are well within its estimate of spam accounts representing less than 5% of daily users, he said.
