Since Wide and Deep Learning for Recommender Systems appeared in 2016, multiple architecture models have been created around this idea of jointly train a wide and deep neural networks as this architecture allow the model to learn both memorization and generalization, which are critical for recommender systems. It may be possible that these kind of architecture change forever the way recommendation systems predict the preference of a user with respect to an item? In the spirit of answering this question from our own experience, we explore, design, and reproduce a deep learning-based model recommender system, and trained it with the Camper's e-commerce dataset. We wanted to validate in our own experience how good a wide and deep model can be, and how much could improve the accuracy of different baseline models. We have explored two different experiments. The first model was trained to predict the potential rating with which a user would evaluate his preference for a certain category of shoes, on a [1,5] scale, whereas the second model was trained to determine whether a user would, or would not, have a interaction with a specific category of shoes. Our experiment's results reveal that wide and deep models present slightly better but similar performance with respect to other deep learning models, however, for small to medium size dataset instances, or for those datasets that do not have the most suitable feature variables for a recommendation problem, then it would be better to use classic algorithms. Wide and deep models have a nice theoretical basis, but in practice the results only improve under certain circumstances, and with huge instances of data, even so the improvement could not be that significant. Our results are an invitation to don't neglect or ignore the nature of the data. Although deep learning models are considerably improving multiple algorithms, they do not always perform better than simpler and well-known machine learning models which require less data pre-processing.