Mining of biomedical data increasingly relies on utility of knowledge repositories. In gene expression analysis, these are often used for gene labeling with an assumption that similarly annotated genes have similar expression profiles. In the paper we use this assumption to craft a method with which we scored six different annotation sources (e.g., Gene Ontology, PubMed, and MeSH annotations) for their utility in gene expression data analysis. Experiments show that the sources that include manual curation perform well and, for instance, score better than automatic annotation from gene-related PubMed abstracts. We also show that there is no clear winner, pointing at the need for methods that could successfully integrate annotations from different sources.