Embodiments are directed towards modifying a distribution of writers as either a push writer or a pull writer based on a cost model that decides for a given content reader whether it is more effective for the writer to be a pull writer or a push writer. A cache is maintained for each content reader for caching content items pushed by a push writer in the content writer's push list of writers when the content is generated. At query time, content items are pulled by the content reader based on writers a content reader's pull list. One embodiment of the cost model employs data about a previous number of requests for content items for a given writer for a number of previous blended display results of content items. When a writer is determined to be popular, mechanisms are proposed for pushing content items to a plurality of content readers.