In digital platforms, recommender systems (RecSys) are essential for aligning content with viewer preferences. This work considers the evaluation of RecSys updates, referred to as “treatments”, by analyzing their “global treatment effect” (GTE) – the expected overall benefit of universally applying these treatments. Our focus is on treatments targeting content creators. We utilize A/B experiments on the creator side and identify that the conventional difference-in-mean estimator is biased for GTE estimation, due to interference among creators competing for visibility. To address this challenge, we introduce a semi-parametric model that combines a parametric choice model, designed to streamline the recommendation process, with a nonparametric component that employs machine learning to account for the heterogeneity among viewers and content. Using this model, we approximate GTE with a doubly robust estimator that satisfies Neyman orthogonality, ensuring consistency and asymptotic normality, and supporting hypothesis testing. We show the robustness and semiparametric efficiency of our estimator even under model mis-specification. We demonstrate the efficacy of our method through simulations and practical applications on a leading short video platform. This is joint work with Shichao Han, Yuchen Hu, and Zhenling Jiang.