ME-POIs Adds Mobility Signals to Place Embeddings

Google Research introduced Mobility-Embedded POIs, or ME-POIs, on August 21, 2026. The framework combines static place descriptions with anonymized, aggregated mobility signals to model both what a location is and how people use it over time.

For developers building mapping, search, recommendation, or location-aware AI systems, the research moves mobility data into the place representation itself. Visit behavior becomes an embedding feature rather than only a prediction target such as a user’s next destination.

From Place Identity to Place Function

Text and multimodal models, including Gemini, can represent a point of interest through its name, address, category, and online description. Those features establish a location’s identity, but they provide limited information about its functional rhythm. A restaurant’s metadata may not indicate its busiest hours, a store’s price tier, or whether a business has effectively stopped operating before its listing is updated.

ME-POIs encodes temporal movement patterns around a place and aligns them with language representations. The resulting embedding carries two forms of information: semantic identity from text and observed population-level function from movement.

This is closely related to domain-specific embedding models, where the goal is to represent signals that general-purpose language embeddings do not capture reliably.

Three-Part Embedding Pipeline

ME-POIs processes geographic points and temporal visit patterns through three stages.

Visit alignment aggregates arrival times, departure trends, and stay durations across a weekly and annual cycle. A temporal encoder uses contrastive learning to form a functional centroid, a multidimensional signature describing how a place is used at different times.

Spatial multiscale visit propagation addresses sparsity at infrequently visited locations. Temporal patterns from data-rich anchor POIs are transferred across multiple spatial scales, including the immediate street, block, and neighborhood. This allows the framework to produce useful representations for long-tail places with limited direct observations.

Text-mobility synergy aligns mobility vectors with high-level text embeddings, including Gemini text embeddings, by maximizing cosine similarity. Mobility information is therefore layered onto the language representation instead of remaining in a separate geospatial model.

The design resembles geospatial representation work such as custom geospatial embeddings, while focusing specifically on temporal human use of individual places.

Results Across Los Angeles and Houston

Google Research evaluated ME-POIs on unseen places in Los Angeles and Houston across five map-enrichment tasks. The reported improvements were measured against text-only and trajectory-based baselines, including Gemini, OpenAI text embeddings, TrajGPT, CTLE, and Skip-Gram.

Task	Reported improvement
Visit intent classification	Up to 81.9% relative gain
Price-level classification	Up to 75.1% improvement
Busyness forecasting	Up to 24.7% lower MAE
Opening and closing hours prediction	Up to 16.2% improvement
Permanent closure detection	Up to 6.5% higher F1

Visit intent classification used Google Maps direction queries as a proxy for popularity. Busyness forecasting measured future crowd density and peak-hour dynamics, while closure detection targeted businesses that had become permanently inactive before online metadata reflected the change.

ME-POIs representations trained only on mobility data also outperformed strong text-only models on selected tasks, including price-level classification. This result shows that movement patterns can contain useful commercial and behavioral information even when a place’s description says little about it.

Engineering and Privacy Implications

If you build a location search or recommendation system, a text-only index can miss time-dependent relevance. Combining semantic retrieval with mobility-aware vectors could support queries and rankings that account for peak periods, likely price positioning, current activity, or stale business listings.

The framework also introduces operational tradeoffs. Mobility signals are population-level and aggregated, so they describe patterns around places rather than the intent or identity of individual visitors. Google Research emphasizes that ME-POIs cannot be used for personal tracking or individual customization.

For production systems, treat mobility as a time-sensitive feature. Embeddings may need refresh schedules, regional validation, and task-specific evaluation rather than a single permanent vector store. Teams working on multimodal retrieval can also review cross-modal RAG pipelines when deciding how to combine text, spatial, and behavioral representations.

The practical next step is to evaluate whether your place-ranking or map-enrichment workload depends on functional signals that text omits, then measure those signals on unseen locations before adding mobility features to a production retrieval pipeline.