According to this other SO question:

The current implementation encodes geographic hash codes atop standard MongoDB B-trees. Results of $near queries are exact. One limitation with this encoding, while fast, is that prefix lookups don't give exact results, especially around bit flip areas. MongoDB solves this by doing a grid-neighbor search after the initial prefix scan to pick up any straggler points. This generally ensures that performance remains very high while providing correct results.