The negative impact of false information on social networks is rapidly growing. Current research on the topic focused on the detection of fake news in a particular context or event (such as elections) or using data from a short period of time. Therefore, an evaluation of the current proposals in a long-term scenario where the topics discussed may change is lacking. In this work, we deviate from current approaches to the problem and instead focus on a longitudinal evaluation using social network publications spanning an 18-month period. We evaluate different combinations of features and supervised models in a long-term scenario where the training and testing data are ordered chronologically, and thus the robustness and stability of the models can be evaluated through time. We experimented with 3 different scenarios where the models are trained with 15-, 30-, and 60-day data periods. The results show that detection models trained with word-embedding features are the ones that perform better and are less likely to be affected by the change of topics (for example, the rise of COVID-19 conspiracy theories). Furthermore, the additional days of training data also increase the performance of the best feature/model combinations, although not very significantly (around 2%). The results presented in this paper build the foundations towards a more pragmatic approach to the evaluation of fake news detection models in social networks.