Using a computational model for the population growth and dispersal of a model species in a fluctuating environment, we test three nature reserve geometries (one large, many small, and a self-similar distribution of reserve sizes) to determine which geometry maximizes species longevity. The self-similar distribution is a close approximation to the distribution of managed areas in the conterminous United States. We consider models with and without migration from or between reserve fragments and both short- and long-range dispersal mechanisms. The optimal geometry depends on the type of dispersal and on the relative probability of survival in protected and non-protected areas. When no migration is allowed from or between reserve fragments of the three geometries, many small equally sized reserves are the optimal geometry. When migration is allowed, the optimal geometry is a single large reserve when the survivability in non-protected areas is low and a self-similar distribution when the survivability is high.