Given the difficulty of setting up large-scale experiments with real users, the comparison of content-based image retrieval methods using relevance feedback usually relies on the emulation of the user, following a single, well-prescribed strategy. Since the behavior of real users cannot be expected to comply to strict specifications, it is very important to evaluate the sensitiveness of the retrieval results to likely variations of users behavior. It is also important to find out whether some strategies help the system to perform consistently better, so as to promote their use. Two selection algorithms for relevance feedback based on support vector machines are compared here. In these experiments, the user is emulated according to eight significantly different strategies on four ground truth databases of different complexity. It is first found that the ranking of the two algorithms does not depend much on the selected strategy. Also, the ranking of the strategies appears to be relatively independent of the complexity of the ground truth databases, which allows to identify desirable characteristics in the behavior of the user.