Predicting vehicle mobility is crucial in domains such as ride-hailing, where the balance between offer and demand is paramount. Since city road networks can be easily represented as graphs, recent works have exploited graph neural networks (GNNs) to produce more accurate predictions on real traffic data. However, a better understanding of the characteristics and limitations of this approach is needed. In this work, we compare several GNN aggregated mobility prediction schemes to a selection of other approaches in a very restricted and controlled simulation scenario. The city graph employed represents roads as directed edges and road intersections as nodes. Individual vehicle mobility is modeled as transitions between nodes in the graph. A time series of aggregated mobility is computed by counting vehicles in each node at any given time. Three main approaches are employed to construct the aggregated mobility predictors. First, the behavior of the moving individuals is assumed to follow a Markov chain (MC) model whose transition matrix is inferred via a least squares estimation procedure; the recurrent application of this MC provides the aggregated mobility prediction values. Second, a multilayer perceptron (MLP) is trained so that—given the node occupation at a given time—it can recursively provide predictions for the next values of the time series. Third, we train a GNN (according to the city graph) with the time series data via a supervised learning formulation that computes—through an embedding construction for each node in the graph—the aggregated mobility predictions. Some mobility patterns are simulated in the city to generate different time series for testing purposes. The proposed schemes are comparatively assessed compared to different baseline prediction procedures. The comparison illustrates several limitations of the GNN approaches in the selected scenario and uncovers future lines of investigation.