Three-dimensional (3D) point clouds have become a popular means of geometric modeling in many applications, but the variability of these due to the noise in the scan and environmental factors can undermine their accuracy and, further, its usability. More importantly, it is critical to understand the relationship between these factors, especially the ones related to the manufacturing processes, and the variability embedded in the point cloud data. Conventional methods require either a large labeled datasets or extensive experience to explore this relationship, which can be time consuming and prohibitive. To tackle this issue, this study introduces a Local Principal Geodesic Analysis (PGA) pipeline for discovering and quantifying local variance in 3D point clouds. Specifically, a variant of PGA algorithm is proposed to project the large set of points in the point cloud data into a lower dimensional space. A Kruskal-Wallis H test was used to perform variance analysis across clusters to show significant differences between clusters (\(p < 0.05\)) and thus demonstrate feature specific variability. A half ball and a freeform structure with multiple scans were selected as case studies to validate the proposed algorithm. Metrics, such as adjusted Rand Index (ARI) and Normalised Mutual Information (NMI), are used to validate consistency in clustering against both objects, with mean values of 0.85 and 0.88 for the half ball, 0.88 and 0.91 for the freeform object, respectively. These confirm the robustness of the pipeline to both stability of cluster assignments across scans and local geometric variability.