A/B/N testing from what I heard. One of the examples I know of it that they split the users into about ten groups, and very slightly changed the blue hyperlink colour between them. They found that even a slight variation made a difference to the click-through rate.