This paper suggests a novel Vietnamese segmentation approach for text categorization. Instead of using an annotated training corpus or a lexicon which are still lacking in Vietnamese, we use both statistical information extracted directly from a commercial search engine and a genetic algorithm to find the optimal routes to segmentation. The extracted information includes document frequency and n-gram mutual information. Our experiment results obtained on the segmentation and categorization of online news abstracts are very promising. It matches near 80 % human judgment on segmentation and over 90 % micro-averaging F1 in categorization. The processing time is less than one second per document when statistical information is cached.