A method includes receiving a training metric, the training metric indicating a parameter to be optimized by content recommendations. A machine learning algorithm maybe used to determine a plurality of different combinations of the recommendation algorithms in the experiment, and each of one or more of the plurality of combinations may be used to generate a content recommendation for one or more visitors. The statistical performance of each of the one or more combinations in optimizing the parameter (based on the content recommendations generated by those combinations) may be monitored and a higher percentage of visitors may be allocated to a combination that generates content recommendations that are the most effective at optimizing the parameter.