Many neural-based recommender systems were proposed in recent years and part of them used Generative Adversarial Networks (GAN) to model user-item interactions. However, the exploration of Wasserstein GAN with Gradient Penalty (WGAN-GP) on recommendation has received relatively less scrutiny. In this paper, we focus on two questions: 1- Can we successfully apply WGAN-GP on recommendation and does this approach give an advantage compared to the best GAN models? 2- Are GAN-based recommender systems relevant? To answer the first question, we propose a recommender system based on WGAN-GP called CFWGAN-GP which is founded on a previous model (CFGAN). We successfully applied our method on real-world datasets on the top-k recommendation task and the empirical results show that it is competitive with state-of-the-art GAN approaches, but we found no evidence of significant advantage of using WGAN-GP instead of the original GAN, at least from the accuracy point of view. As for the second question, we conduct a simple experiment in which we show that a well-tuned conceptually simpler method outperforms GAN-based models by a considerable margin, questioning the use of such models.