How mobility gives language models a deeper understanding of place

Google says anonymized mobility patterns made place embeddings materially better at predicting how businesses actually operate.

Its ME-POIs framework blends text metadata with aggregate signals such as arrival times, stay duration, and nearby movement patterns. In tests on Los Angeles and Houston, it improved predictions for unseen places, including visit intent, price level, and busyness. The system also propagates regional mobility patterns to sparse locations, so smaller or newer businesses are not treated as having no activity. Google stresses the framework models places in aggregate and is not for individual personalization. Google Research's note