Currently, many people share news, links, or videos, without being aware of the impact they can have on people's decisions or ways of acting. A clear example, recently experienced in Colombia, corresponds to the national strike which happened at the time of this research. Due to these unexpected circumstances, colombians experienced the influence news have on decision making that can affect the country, not only economically but politically, and socially. It showed how news can generate fear in people, or even misinform, as is the case of fake news. For these reasons, it is key to determine the relevance a story can have. Predicting the impact, will allow us to pay more attention to those news that can affect people more, avoiding misinformation and fake news. However, the problem is that there is no way of predicting the impact that a press article can have. Therefore, the aim of this work is to implement a machine learning model that allows us to predict, with the best possible accuracy, the virality of online press articles (defining virality as the amount of clicks that an article receives when it is opened). In order to achieve this goal, we followed the CRISP-DM methodology, which focuses on machine learning projects. The best obtained result corresponds to the model where the core of the architecture was based on BERT, a pre-trained model, which, for a pair of press articles headlines, predicted whether the first headline would be more viral than the second one. On the other hand, the evaluation was carried out by comparing the amount of clicks for a pair of articles. For a practitioner point of view, digital marketers can use our results to select the best words for their online marketing campaign. For a theoretical point of view, our results present an innovative natural language processing approach based on one of the best breed of Neural network models (BERT).