As a consequence of significant advances in information technology, the marketing community has become increasingly interested in the possibilities afforded by interactive media. The explosion of the World Wide Web is the most notable example of such interest. Interactive media allow the marketer to 1) identify the consumer and characteristics of the consumer, 2) decide on the marketing message real-time, and 3) capture response to marketing communications. In contrast to some traditional media (e.g., television, radio, print), which send one standardized message to all consumers, interactive media allow the marketer to deliver customized messages tailored to the individual consumer. In addition, unlike most other marketing environments which require media planning decisions to be made in advance, interactive media allow the marketer to make decisions “on the fly,” using information about previous decisions to guide the current decision. In other words, marketing decisions through interactive media can be truly dynamic. In this paper, we formulate a unique procedure to exploit the benefits of interactive media. We study the general problem of a marketer whose objective is to maximize expected return (e.g., response rate) over the course of a direct response marketing campaign. The marketer has the ability to dynamically allocate two or more unique marketing messages (e.g., ads, Web pages) to achieve this objective. Consumer response to a particular message may depend on a set of covariates (e.g., demographic characteristics). In ignorance of the true relationship between response and the covariates, the marketer can use regression techniques to learn about the parameters of the response functions. Moreover, because the marketer can continually update the parameter estimates, he can also continually adapt the decision of which message to select. The theoretical framework we draw from is the multi-armed bandit problem in statistical decision theory in which there is a fundamental dilemma between “information,” such as the need to learn about the parameter values governing response for each ad, and “control,” such as the objective of maximizing response rate. In such problems, it may be wise to sacrifice some potential early payoff for the prospect of gaining information about consumers that will allow for more informed decisions later. An important difference in our problem is that we incorporate covariates (although our solution can handle the standard non-covariate case). Suppose that the marketer has two or more unique marketing messages available and weak prior beliefs as to the effectiveness of each message. The marketer can randomly assign the messages for a brief initialization period to collect training samples for each of the messages. From then on, the marketer can estimate response functions, which can be periodically updated. In deciding which message to select for any particular consumer, the marketer can compare the regression estimates and select the message with the highest predicted response. We refer to this approach as the myopic rule. However, since the predictions are estimated with imprecision, it may be worthwhile for the marketer to calculate “uncertainty adjustments,” and incorporate these adjustments in its message selection decisions. The uncertainty adjustments reflect the imprecision in the parameter estimates of the regression models. The adjustments are larger for consumers that have extreme characteristics (covariates) than for typical consumers because an extreme consumer will provide more information about the nature of the response function than a typical consumer will. Our proposed procedure is easy to implement, and will enable marketers to increase the success of their marketing campaigns. In addition, the approach can be used in several different marketing applications, including direct response advertising, Web page design, and product development.