Pedestrian Dead Reckoning (PDR) plays an important role in many (hybrid) indoor positioning systems since it enables frequent, granular position updates. However, the accumulation of errors creates a need for external error correction. In this work, we explore the limits of PDR under realistic conditions using our graph-based system as an example. For this purpose, we collect sensor data while the user performs an actual navigation task using a navigation application on a smartphone. To assess the localisation performance, we introduce a task-oriented metric based on the idea of landmark navigation: instead of specifying the error metrically, we measure the ability to determine the correct segment of an indoor route, which in turn enables the navigation system to give correct instructions. We conduct offline simulations with the collected data in order to identify situations where position tracking fails and explore different options how to mitigate the issues, e.g. through detection of special features along the user’s path or through additional sensors. Our results show that the magnetic compass is often unreliable under realistic conditions and that resetting the position at strategically chosen decision points significantly improves positioning accuracy.