Google Research showed that a model learns far more about a place from where people actually go than from what the place says about itself. Their method, Mobility-Embedded Points of Interest, fuses a location's static text description with aggregated, anonymized visit patterns, and the gains are lopsided toward exactly the questions text cannot answer: predicting why someone visits a place improved by 81.9 percent and predicting its price level by 75.1 percent. The work is published as arXiv 2601.21149 and described in an August 21 post.
Key facts
- Five evaluation tasks: weekly opening hours, permanent closure detection, visit intent, weekly busyness, and price level.
- Largest gains: visit intent up 81.9 percent, price level up 75.1 percent; busyness error down 24.7 percent; opening hours up 16.2 percent; closure detection up 6.5 points of F1.
- Paper submitted January 29, 2026, last revised June 6, 2026 as version 3.
- Primary source: Google Research's blog post and the arXiv record.
The premise is easy to state and surprisingly deep. A listing says "cafe." That single word covers a place where people sit for three hours with laptops on weekday mornings, and a place where a line of people each spend ninety seconds before 8am and nobody comes back until tomorrow. Those are different businesses with the same label. The visit pattern separates them instantly; the text never will. Once you look for it, most of what you want to know about a place lives in behavior rather than description: whether it is expensive, whether it is a destination or a stop, whether it is open, whether it is still in business at all.
The method is more than bolting a visit count onto a text embedding. Each visit is encoded with its arrival and departure timing, so duration and time of day become part of the signal rather than being averaged away. Then contrastive learning aligns the visit-level representation with the place-level one, which is the standard trick for making two different kinds of data land in the same space where they can be compared. The pipeline also propagates visit patterns from nearby frequently visited places across several spatial scales, which handles the sparsity problem, since most locations have far too little traffic data on their own. An auxiliary objective keeps the result anchored to the text so the representation does not float free of what the place actually is.
The spatial propagation deserves a note because it is where the geographic reasoning lives. A quiet business next to a busy transit stop behaves differently from an identical business in an office park, and neither one has enough of its own data to learn that. Borrowing structure from well-observed neighbors at multiple radii is how the model gets a usable representation for the long tail, and it is the same instinct behind every good approximate nearest neighbor system: your neighbors tell you what you are.
The result to sit with is that mobility-only representations beat some text-only models on certain tasks. That is a strong claim about where information lives. It means that for those questions, knowing nothing about a place except the temporal shape of who comes and goes is more informative than reading its full description. Language models are trained on descriptions of the world. This is a concrete measurement of how much of the world the descriptions leave out, which is the same question underneath world models research from a completely different direction.
The uneven distribution of gains is itself the finding. Opening hours improved modestly, at 16.2 percent, because opening hours are frequently written down and text models can already read them. Visit intent improved by more than 80 percent because nobody writes down why people come. The size of each improvement is a rough measure of how much of that task's answer was never in the text to begin with, which makes this a useful diagnostic beyond the geospatial setting: wherever behavioral data produces a large jump, the text was never carrying the answer.
The honest caveat is that these are aggregated and anonymized mobility patterns from a company that holds an unusual amount of location data, and the method's quality depends directly on the density and representativeness of that data. Places and populations that are under-observed will get worse representations, and the spatial propagation step means a sparse location partly inherits its neighbors' character whether or not that is accurate. There is also no independent replication here; the gains are Google's own reported numbers on Google's own evaluation. The technique is well described enough to be reproduced by anyone with comparable data, which is a short list.
Originally published on Ground Truth, where every claim is checked against the primary source.
