To gain insights into the biological function and relevance of genes using serial analysis of gene expression (SAGE) transcription profiles, one essential method is to perform clustering analysis on genes. A successful clustering analysis depends on the use of effective distance or similarity measures. For this purpose, by considering the specific properties of SAGE technology, we modeled the SAGE data by Poisson statistics and developed two Poisson-based measures to assess similarity of gene expression profiles. By employing these two distances into a K-means clustering procedure, we further developed a software package to perform clustering analysis on SAGE data. The software implementing our Poisson-based algorithms can be downloaded from http://genome.dfci.harvard.edu/sager. Our algorithm is guaranteed to converge to a local maximum when Poisson likelihood-based measure is used. The results from simulation and experimental mouse retina data demonstrate that the Poisson-based distances are more appropriate and reliable for analyzing SAGE data compared to other commonly used distances or similarity measures.