Avatars are now appearing as online assistants on transactional websites, yet their scope is still limited. This paper explores their potential roles in providing assistance, a friendlier interface and purchase recommendations. As avatars are at early stages of implementation, the conceptual framework draws upon human-computer interaction research, plus cognate literature on salesperson roles and the use of synthetic characters in other contexts. The empirical study involved two longitudinal panels of online buyers and an international, online survey of 2114 internet users. Following split-sample principal components analysis and k-means clustering, four categories of role preference are identified. The results inform decisions on the appropriateness of avatars, their adaptation to customer needs and buying contexts, and their possible roles. Hypothesised relationships with age, gender and online buying experience are tested, suggesting scope for avatar role segmentation. Suggestions are offered for marketers and website designers, considering deploying avatars, and for future research directions.