How mobility gives language models a deeper understanding of place

Google Research is pushing language models past the tidy world of addresses and category labels. Its new Mobility-Embedded POIs, or ME-POIs, framework adds aggregated, anonymized movement patterns to place descriptions so models can pick up how a place actually behaves over time.

The basic idea is simple enough: a café is not just a café because of what the text says about it. It also has a rhythm. People arrive at certain times, stay for certain lengths, and move through the surrounding area in patterns that repeat across days and seasons. ME-POIs turns that activity into an embedding, a numerical signature that combines the place’s identity with its functional behavior.

To build that signature, the system runs through three steps: visit alignment, spatial multiscale visit propagation, and text-mobility synergy. It uses aggregate arrival windows, departure trends, and stay durations, then spreads signal from busy neighboring places to nearby sparse ones across the street, block, and neighborhood. That matters because the long tail is a real problem here. Famous landmarks and huge shopping centers generate plenty of data. Small repair shops and newly opened cafés usually do not.

Google Research tested the framework in Los Angeles and Houston on five tasks: opening and closing hours, price level classification, permanent closure detection, visit intent classification, and busyness forecasting. The model was trained on observed places and then asked to predict attributes for places it had never seen before. Against text-only embeddings, trajectory-based geospatial models, and hybrid baselines, ME-POIs improved performance across the board.

The headline numbers are hard to ignore. With advanced text models, ME-POIs produced up to an 81.9% relative gain in visit intent prediction, a 75.1% improvement in price level classification, and a 24.7% increase in busyness estimation accuracy on unseen places. In some cases, mobility-only models even beat text-only models, which is a neat reminder that what people do somewhere can say more than the polished words attached to it.

Google is careful about scope here. The framework is about aggregate behavior across broad populations and time frames, not individuals, and it is presented as part of the company’s broader Google Earth AI effort. The direction is clear: for some kinds of place understanding, the street traffic may be more revealing than the listing page.