Recommendation systems have become essential across domains, while large language models (LLMs) have become central to human-like, user-centered interaction. However, both share a common limitation: their decision-making processes are opaque and difficult to interpret. Neural recommenders make accurate predictions but provide little insight into why those predictions are made. LLMs can generate convincing explanations, yet these explanations are often unfaithful or hallucinatory. We introduce a framework that combines their strengths by keeping LLMs for fluent communication and grounding them with counterfactual evidence derived from the recommender itself. This design aligns the explanation with the model’s internal reasoning while maintaining natural and persuasive presentation.