Recommender systems are the backbone of personalized services that provide tailored experiences to individual users. Still, data sparsity remains a common challenge, especially for new applications where training data is limited or unavailable. We'll present a combinatorial optimization problem that formalizes the selection of item universe for experimentation in recommender systems. On one hand, a large set of items is desirable to increase diversity. On the other hand, a smaller set enables rapid experimentation and minimizes the time and the amount of data required to train machine learning models. We show how to optimize for such conflicting criteria using a multi-level optimization framework. Our approach integrates techniques from discrete optimization, unsupervised clustering, and latent text embeddings. Experimental results on well-known recommendation benchmarks demonstrate the benefits of optimized item selection.