Pagerank works on the nodegraph of a series of pages and the directed edges formed by their respective inward and outward links. Thus the rank of a particular page is broadly a locally-induced effect in the nodegraph.
SVD, on the other hand, works on a whole matrix of values, and has no directionality - a link between site A and site B would only register as a 1 on the correct matrix element. It is a global system, so ranking is a global effect.
Given the extreme sparseness of web-derived matrices, I would expect SVD to be a bad performer here, as it requires a complete dataset, and has significant memory requirements.
Is that true? Does Pagerank outdo SVD largely because it is a nodegraph-based algorithm? How can Pagerank infer semantic relevance from a page beyond the number of times a word is mentioned? Or would that be a second step, performed after Pagerank has ranked the pages?