This study addresses the sensitivity of short-term flow forecasting in the Seine River basin (43,800 km2, France) to the spatial distribution using a semi-distributed model (Transfer with GR, TGR). The basin was decomposed into intermediate basins depending on the gauging stations selected for this study. A lumped hydrological model was applied on each intermediate basin and a routing model was used to propagate the discharge through the river network. Discharge data at the gauging stations were assimilated using a Kalman filter and tests for flow forecasting were performed with a lead time up to 72 h. Several spatial configurations, defined by a selection of one or several gauging stations, were tested and the performances were compared to a reference lumped model currently used operationally by the regional flood forecasting centre. Results showed that the forecasting performance improves with an increase in the degree of spatialization. Nevertheless this improvement was not systematic and the integration of some particular gauging stations degraded the model performance. In addition, it was shown that integrating some other stations (generally the most upstream) led to a negligible improvement. This suggests that in an operational context, where the model has to be robust and computationally efficient, some efforts should focus on finding the optimal spatial distribution, which is not necessarily the one using all the available stations.