Prior studies on species-specific responses to habitat alteration have demonstrated that niche breadth is positively associated with patch occupancy rates in landscapes fragmented by agriculture. However, these studies generally have focused on vertebrates and relied upon data collected at a single point in time, neglecting dynamic processes that could alter inferences. We studied the effects of host selection and forest fragmentation on population dynamics of acorn weevils Curculio, the primary insect seed predators of oaks in North America. Detection/non-detection data were collected from 174 red and white oaks in 19 forested fragments from 2005-2008. We used dynamic multi-season site-occupancy models within a Bayesian framework to explore variation in patch (tree-level) occupancy dynamics of three species of weevils that vary in their specialization, i.e. their relative selection of red and white oak as hosts: C. pardalis (white oak specialist), C. sulcatulus (generalist) and C. proboscideus (generalist). Contrary to expectations, the specialist exhibited greater estimated rates of occupancy than generalists. However, red oak trees occupied by the white oak specialist appeared to function as sink populations maintained by frequent colonization following local extinction. Specialists also exhibited greater relative variation in occupancy and relative abundance on their host trees among years. Generalists exhibited lower local extinction and colonization rates than the specialist. Occupancy and vital rates of weevils on a host tree increased with acorn production and were significantly influenced by neighborhood forest density. Our results suggest that across much of their range in the eastern United States acorn weevils exist in fragmented, temporally dynamic landscapes, with generalists occurring on a lower proportion of usable trees but buffered by access to more suitable patches and greater patch-specific survival. More generally, our results demonstrate that estimates of specialization derived from occupancy data may be misleading in the absence of patch-specific information on vital rates.