Recent research has amply documented that even preschoolers learn selectively from others, preferring, for example, reliable over unreliable and competent over incompetent models. It remains unclear, however, what the cognitive foundations of such selective learning are, in particular, whether it builds on rational inferences or on less sophisticated processes. The current study, therefore, was designed to test directly the possibility that children are in principle capable of selective learning based on rational inference, yet revert to simpler strategies such as global impression formation under certain circumstances. Preschoolers (N = 75) were shown pairs of models that either differed in their degree of competence within one domain (strong vs. weak or knowledgeable vs. ignorant) or were both highly competent, but in different domains (e.g., strong vs. knowledgeable model). In the test trials, children chose between the models for strength‐ or knowledge‐related tasks. The results suggest that, in fact, children are capable of rational inference‐based selective trust: when both models were highly competent, children preferred the model with the competence most predictive and relevant for a given task. However, when choosing between two models that differed in competence on one dimension, children reverted to halo‐style wide generalizations and preferred the competent models for both relevant and irrelevant tasks. These findings suggest that the rational strategies for selective learning, that children master in principle, can get masked by various performance factors. Preschoolers are capable of rational inference on the basis of trait reasoning in their selective trust. Yet, they fall back on simpler cognitive strategies (e.g., global impression formation) due to limiting performance factors, such as a lack of relevant trait knowledge, insufficient executive function, or task structures that facilitate simpler strategies.