Retrovirus genomes contain a sequence at the 5' end which directs their packaging into virions. In Rous sarcoma virus, previous studies have identified important segments of the packaging signal, Psi, and support elements of a secondary-structure prediction. To further characterize this sequence, we used an in vivo selection strategy to test large collections of mutants. We generated pools of full-length viral DNA molecules with short stretches of random sequence in Psi and transfected each pool into avian cells. Resulting infectious virus was allowed to spread by multiple passages, so that sequences could compete and the best could be selected. This method provides information on the kinds of sequences allowed, as well as those that are most fit. Several predicted stem-loop structures in Psi were tested. A stem at the base of element O3 was highly favored; only sequences which maintained base pairing were selected. Two other stems, at the base and in the middle of element L3, were not conserved: neither base pairing nor sequence was maintained. A single mutation, G213U, was seen upstream of the randomized region in all selected L3 stem mutants; we interpret this to mean that it compensates for the defects in L3. Randomized mutations adjacent to G213 maintained the wild-type base composition but not its sequence. The kissing-loop sequence at end of L3, postulated to function in genome dimerization, was not required for infectivity but was selected for over time. Finally, a deletion of L3 was constructed and found to be poorly infectious.