E-commerce agents with Bayesian learning were first proposed by Zeng and Sycara in their Bazaar automated bargaining system . Many studies have directly applied or extended Bazaar to agent learning. In Bayesian learning, it is critical to construct the conditional probabilities for new events in order to obtain an accurate estimation of the posterior probability. The construction of such conditional probabilities requires domain knowledge of the target problem and an appropriate translation of this knowledge into a corresponding set of conditional probabilities. Unfortunately, such issues have either been ignored or over-simplified in previous studies. Accordingly, the present study aims to enhance the performance of Bayesian learning by developing a new formulation for the conditional probabilities during the learning process. An online used-textbook store is built and used as the basis for a series of experiments to evaluate the performance of the proposed approach. The experimental results demonstrate that the prediction accuracy of Bayesian learning using the proposed conditional probability formulation is superior to that of a previous approach that uses a simpler formulation of conditional probabilities.