Techniques for distributing online ads by targeting online ad requests using test data to predict performance. The techniques can target ad requests in automated online advertising systems in which ad requests are generated by an ad exchange server and bids are placed by marketer devices in real time. The techniques aggregate bid units and compare bid unit characteristics to select bid units to target in ways that address data sparsity, variance, and volume issues. Data sparsity issues are addressed by aggregating bid units to avoid using bid units having insufficient data. Data variance issues are addressed by computing stability metrics for bid units that enable discounting the effect of outliers. Data volume and processing efficiency issues are addressed by grouping similar bid units based on similar metrics (e.g., normalized ROI) and/or similar stability scores, and then ranking the bid units and selecting the top ranked bid units to target.