Models trade text descriptions for real human traffic patterns

A text profile rarely tells the whole truth

Language models typically recognize points of interest (POI) solely through text metadata, such as a name, an address, and a broad category. This approach works for large businesses with a rich online presence but struggles with smaller local shops. Google has now blended text descriptions with aggregated and anonymized mobility data, specifically arrival times, stay durations, and movement density in the immediate surroundings.

Nearby activity rescues isolated places

The biggest drop in reliability for older models occurs in the so-called long tail: hidden cafes or newly opened stores with few records. ME-POIs addresses this gap by looking beyond the business itself to analyze the rhythm of the entire street or neighborhood. The dynamics of the surrounding area can transfer missing activity data even to data-poor locations.

When numbers beat text labels

According to Google's results, mobility data alone outperformed text-only models in certain tasks. This was evident, for instance, when estimating a business's price level. The resulting representation serves to improve aggregate estimates without requiring personalization or profiling at the individual visitor level.

Current traffic outpaces delayed online profiles

Beyond pricing, the new framework was tested on detecting permanent business closures and forecasting busyness. A drop in visit rhythm can signal a company's end before the internet even notices. The research relies on public benchmarks from two cities, leaving deployment in everyday conditions outside these controlled areas as an open theoretical step.

Lilith's verdict

You can make up a marketing label for a business, but you cannot fake when people show up and how long they stay. Google simply showed that for AI, actions are more interesting than words.