Google ME-POIs Boosts AI Understanding of Physical Places

Google Research has unveiled Mobility-Embedded POIs (ME-POIs), a novel framework designed to help artificial intelligence models better comprehend the physical world. While large language models like Gemini excel at processing static text metadata such as addresses and business categories, they often lack a sense of real-world functional rhythms. ME-POIs addresses this by blending static text representations with aggregated, anonymized mobility patterns—such as arrival times, stay durations, and surrounding movement patterns—sourced from public benchmark datasets.

The integration of ME-POIs with advanced text models yielded dramatic performance improvements during testing in Los Angeles and Houston. When evaluated on unseen places, the framework achieved an 81.9% relative gain in predicting visit intent, a 75.1% improvement in price-level classification, and a 24.7% increase in busyness estimation accuracy. Researchers tested the system across five distinct tasks: predicting opening and closing hours, classifying price levels, detecting permanent closures, classifying visit intent, and forecasting busyness.

To build these representations, the ME-POIs pipeline uses a three-step process: visit alignment, spatial multiscale visit propagation, and text-mobility synergy. The system maps temporal sequences into a dense vector space to establish a functional signature over a one-year cycle. To solve the data sparsity problem for small businesses with few recorded visits, the spatial propagation mechanism transfers aggregated visit patterns from busy, data-rich neighbors to nearby quiet places. Finally, the framework aligns high-level language embeddings from models like Gemini with the mobility vectors by maximizing their cosine similarity.

By comparing ME-POIs against standard text-only baselines and trajectory-based models like TrajGPT, the researchers demonstrated that collective physical actions can be more descriptive than formal text labels. Google noted that the framework focuses entirely on aggregate, anonymized data rather than individual personalization. This work is part of the broader Google Earth AI initiative to convert planetary data into actionable intelligence.