Accurate urban road traffic flow prediction is crucial in many smart city applications. This study presents a codesign approach to traffic sensor location problem (TSLP) and traffic flow prediction using graph neural networks (GNNs). The zone-based TSLP algorithm maximizes coverage by dividing the city into representative districts that balances randomization and expert opinion building on the SOTA methodologies of TSLP. The GNN utilizes these partitions to improve prediction accuracy. By aligning the spatial logic of sensor placement with the information propagation capabilities of GNNs, this approach creates a robust framework where machine learning predictions are rooted in strategically gathered data, achieving a balance between prediction accuracy and cost-effective sensor coverage. The algorithm was validated based on simulated traffic flow data utilizing SUMO traffic simulator on the road network of the city Győr. Simulation results suggest that the proposed algorithm is capable of both traffic estimation and prediction even with only 10 % of the roads covered by sensors.