This work investigates the significance of choosing appropriate recurrent neural networks (RNNs) architecture for a spatiotemporal next location prediction framework. Dockless shared micro-mobility sharing programs provide spatial trajectory data that entails essential information for city planners and developers. The study compares (i) the variable-sized geohash tessellation and (ii) two common RNN architectures: Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), using bike/scooter location data for Washington DC, USA. LSTM and GRU networks are used for modeling and incorporating information from spatial neighbors into the model. The study suggests that the LSTM model yields slightly better performance than the GRU model based on the same tessellation. However, geohash size might play a significant role in model performance. The study highlights the need to explore hyperparameter tuning, multiple spatial partitioning techniques especially with the Google S2 library, and more trip data for improving the prediction performance in neural network models.