Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for document update generation. In one aspect, a method includes identifying pairs of content nodes where a first content node in each pair is in a first hierarchical representation of a first document and a second content node in each pair is in a second hierarchical representation of a second document, in which the content nodes represent visible content and in which identifying comprises selecting the first and second content nodes such that a cost based on structural differences between the first and second hierarchical representations and a content difference between the first and second content nodes is minimized; associating rendered layout information related to the first content node with the second content node; and determining whether to generate a snippet for the content difference between the first and second content nodes based on the rendered layout information.