We outline and test a Bayesian model of the effects of evidence sampling on property induction. Our model assumes that people are sensitive to the effects of different sampling frames applied to sampled evidence. Two studies tested the model by comparing property generalization following exposure to samples selected because they belong to the same taxonomic category or because they share a salient property. Both studies found that category-based sampling led to broader generalization than property-based sampling. In line with model predictions, these differences were attenuated when a mixture of positive and negative evidence was presented (Experiment 1) and when category-property relations were probabilistic rather than deterministic (Experiment 2).