Do learning paradigms in context-aware Earth embeddings suppress spatial variability? Geospatial embeddings (GE) offer user-friendly representations of large EO image archives with competitive performance across numerous predictive mapping applications. A wealth of GE are available, including Alpha Earth Foundations, TESSERA, or Major TOM. Importantly, architectures behind GE differ fundamentally - we can discriminate pixel-level against context-aware architectures: embeddings based on …