Methods, systems, and apparatus for determining content item quality based on content item presentation data are presented. In one aspect, a method includes storing a first count of occurrences of a pair of content items being presented; storing a second count of second occurrences of a first content item of the pair being interacted with when it was presented in a lesser position in a resource relative to a second content item of the pair when the second content item of the pair was not interacted with; determining that the first count meets a first threshold; determining that a ratio of the second count to the first count meets a second threshold; and storing correlation data for the second content item based on the ratio.