Existing peer-to-peer overlay approaches for location-based search have proven to be a valid alternative to client-server-based schemes. One of the key issues of the peerto-peer approach is the high churn rate caused by joining and leaving peers. To address this problem, this paper proposes a new location-aware peer-to-peer overlay termed Geodemlia to achieve a robust and efficient location-based search. To evaluate Geodemlia, a real world workload model for peer-to-peer location-based services is derived from traces of Twitter. Using the workload model, a system parameter analysis of Geodemlia is conducted with the goal of finding a suitable parameter configuration. In addition, the scalability and robustness of Geodemlia is compared to a state-of-the-art tree-based approach by investigating the performance and costs of both overlays under an increasing number of peers, an increasing radius of area searches, an increasing level of churn as well as for different peer placement and search request schemes. The evaluation results reveal that in contrast to the tree-based approach, Geodemlia provides on average a 46% better success ratio as well as a 18% better recall at a moderate higher traffic overhead of 13 bytes/s and an increased average response time of 0.2 s.