Google researchers have created a new framework that teaches artificial intelligence models to understand physical places not just by reading descriptions, but by analyzing how people actually move through and use them. The breakthrough, called Mobility-Embedded Points of Interest (ME-POIs), combines traditional text-based information with real-world movement patterns to give AI a much deeper grasp of what makes a location tick.

Traditional language models like Gemini excel at processing text, but they struggle with understanding the dynamic reality of physical spaces. A restaurant's category and address tell you what it is on paper, but they don't capture whether it's a lunch hotspot or a late-night destination. ME-POIs bridges this gap by layering anonymized mobility data, such as arrival times, stay durations, and surrounding movement patterns, directly onto text-based place representations.

How Does Mobility Data Transform AI's Understanding of Places?

The framework works through a three-step process that transforms raw geographic and movement data into a mathematical signature that captures both a place's identity and its actual function. First, the system analyzes aggregate visits to specific locations, mapping temporal arrival windows, departure trends, and typical stay durations into what researchers call a "functional centroid." This creates a unique multidimensional signature that represents the anonymized mobility patterns associated with that place over a full year and across different days of the week.

The second step tackles a persistent problem in geospatial AI: the "long tail" problem. While famous landmarks and popular chains generate abundant visit data, the vast majority of small local businesses, neighborhood boutiques, and newly opened cafes suffer from severe data sparsity. ME-POIs solves this through spatial multiscale visit propagation, which recognizes that visits are usually regionally constrained. A small boutique on a high-end shopping street shares behavioral traits with its neighbors, so the framework looks at adjacent places across multiple spatial scales: the immediate street, the block, and the wider neighborhood. It then statistically transfers the aggregated visit patterns of busy, data-rich neighbors to nearby sparse places.

The third step enriches textual descriptions by aligning language embeddings with newly generated mobility vectors. Rather than discarding text, the framework preserves structural semantics, such as knowing that a place sells food, while absorbing operational context, such as whether it functions as a lunch spot or a late-night diner.

What Real-World Improvements Did the Research Show?

Google tested ME-POIs across five distinct prediction tasks in two large, culturally distinct metropolitan areas: Los Angeles and Houston. The results were substantial. The framework delivered a 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 across unseen places. These improvements matter because they demonstrate that the model develops genuine understanding rather than simply memorizing local patterns.

The five tasks researchers evaluated included:

• Opening and Closing Hours Prediction: The framework can infer the exact schedule of a business based on its mobility signature and text description.
• Price-Level Classification: ME-POIs can distinguish between a luxury high-end boutique and a thrift store based solely on mobility context and semantic information.
• Permanent Closure Detection: Businesses often close long before owners update online profiles or crowd-sourced reports are filed; the embeddings help flag such closed businesses.
• Visit Intent Classification: The framework estimates aggregate search and navigation interest for a place, serving as a proxy for overall popularity.
• Busyness Forecasting: ME-POIs predicts future crowd densities and peak hour dynamics of businesses.

To prove the model develops general intelligence, researchers trained it on observed places and then asked it to predict attributes on entirely unseen locations. This approach ensures the framework isn't simply memorizing local patterns but learning transferable principles about how places function.

How to Leverage Mobility Data for Better AI Place Understanding

• Combine Static and Dynamic Data: Integrate text metadata with real-world mobility patterns to create comprehensive place representations that capture both identity and function.
• Address Data Sparsity Through Regional Learning: Use spatial multiscale propagation to transfer behavioral patterns from data-rich areas to sparse locations, enabling predictions even for newly opened or rarely visited businesses.
• Align Semantic and Mobility Signals: Maximize the connection between language embeddings and mobility vectors to preserve what text describes while absorbing operational context that text alone cannot capture.

The implications extend beyond academic interest. Businesses, urban planners, and mapping services could use these insights to better understand neighborhood dynamics, predict foot traffic, and optimize operations. For AI systems powering navigation apps, local search, and business intelligence tools, this framework offers a path to more accurate and contextually aware recommendations.

The research highlights a broader shift in AI development: moving beyond pure text understanding toward models that grasp the physical world's temporal and spatial complexity. By treating places not as frozen sets of words but as dynamic entities with measurable functional rhythms, Google's framework demonstrates that the next generation of AI systems will need to think like urban observers, not just language processors.