In one embodiment, techniques herein determine a plurality of resources loaded during rendering of a web page on a client device, and determine a duration of time taken for each of the plurality of resources to fully load. Accordingly, the techniques herein may then cluster the plurality of resources into clusters, comprising a first cluster consisting of resources having the shortest durations of the plurality of resources and a second cluster consisting of resources having the longest durations of the plurality of resources. Those resources of the first cluster may then be classified as cached resources, while those resources of the second cluster may be classified as non-cached resources.