Geographic routing protocols exploit node positions on the Euclidean plane to route packets in multi-hop networks. Their potential lies in forgoing the overhead of building routing structures for instance in the form of link-state tables. At the same time their performance suffers from local dead ends and missing or inaccurate node coordinates. These issues have been addressed by various means, most notably by falling back to face routing on a planarized network graph or to a backup routing infrastructure such as a minimum spanning tree. Existing solutions however are not geared towards finding efficient paths or result in a considerable - often upfront and global - communication overhead. We propose assisting geographic routing adaptively by resorting to an on-demand route discovery only when and where it becomes necessary to sustain its greedy forwarding functionality. Simulating practically relevant qualities of dead ends and location inaccuracies, we show how the overhead of our method to deal with such cases remains localized and limited. We claim that our results make geographic routing more practicable particularly in wireless sensor networks that are deployed in immobile large-scale meshes and characterized by locally varying network conditions.