Both real-world category knowledge and instance-based sample data are often available as sources of inductive inference. In three experiments using natural social categories, we test the influence of general category knowledge on the use of category instances to make property inductions both to other category members and to others in the population. We find that a category's coherence--the extent to which its features are interrelated through prior knowledge (Murphy & Medin, 1985)--influences inductions positively to new category members and negatively to the population. This effect of coherence is strongest with small as compared with large samples of instances. The results are interpreted from both similarity and explanation-based perspectives.