Can Explanations Improve Recommendations? Evidence from Prediction-Informed Explanations

Recommender systems are central to how modern digital platforms match consumers with content, yet state-of-the-art models are often black-box, limiting their interpretability and adoption in many business settings. Existing explainable AI approaches typically treat explanations as separate or post-hoc tasks, often at the cost of predictive accuracy. We challenge this practice and propose that explanations, when incorporated as an \emph{integral} part of the learning process, can improve \emph{both} explainability and predictive performance. We introduce RecPIE (Recommendation with Prediction-Informed Explanations), a framework that jointly learns personalized natural-language explanations and recommendation predictions. Central to RecPIE’s design is our proposed \emph{usefulness} reward, which uses downstream recommendation accuracy to guide explanation generation, contrasting existing literature that focus on explanation quality objectives alone and therefore the generated explanations do not necessarily improve predictive accuracy. Specifically, guided by this reward, RecPIE fine-tunes an LLM using LoRA through reinforcement learning (PPO) to generate \emph{prediction-informed explanations}, which are then embedded into a discriminative recommender to improve predictions (\emph{explanation-informed predictions}), creating a mutually reinforcing learning loop. We evaluate RecPIE on large-scale point-of-interest recommendation from Google Maps, a challenging recommendation setting where consumer preferences span diverse place categories and are long-tailed, requiring the recommender system to generalize across category and to data sparse scenarios. RecPIE improves predictive accuracy by 1.7—3.6\% over state-of-the-art baselines and outperforms the strongest full-data baselines using only 25\% of the training data. The gains are largest exactly where black-box AI systems struggle most: consumers with niche tastes and sparse histories, where firms face a particularly difficult targeting problem with limited information. In a human study with 566 participants, RecPIE’s explanations are preferred 61.5\% of the time (compared with 16.6\% for the strongest baseline), and are rated closer to consumer-elicited reasons. Together, these results suggest that insights about consumer behavior and AI decisions need not be imposed externally or generated post hoc; they can be learned as an integral component of the AI systems, allowing the systems themselves to discover which insights are useful for improving the objective the platform ultimately cares about.