How Near works
It searches backwards
A map answers “what is near this point?”. Near answers “which points satisfy all of these at once?” — then ranks them worst-requirement-first, so a spot two minutes from everything beats one next to a park but at the limit for groceries.
Results are spread at least 400 m apart, so ten answers are ten neighbourhoods rather than ten doors on one street.
Walking times are straight-line
Distance divided by 80 m/min — about 4.8 km/h. It is not a routed isochrone. In cities cut by rivers, motorways or rail this overstates how far you can really get, and we would rather say so than imply an accuracy the number does not have.
What the numbers mean
Every search reports how many candidate spots were examined and how many passed. When only 16% of Berlin satisfies five requirements at five minutes, that figure is the answer — it says something true about the city that a list of pins cannot.
The data
A full-planet OpenStreetMap import: 372,883,008 places, last refreshed 2026-03-02. Coverage is global but uneven, and strongest in Europe and North America. Nothing here is a review or a rating; it is what contributors have mapped.
Where the model is, and is not
A language model reads your sentence when plain keyword matching cannot, and it may only choose from the fixed category list this API publishes. It never supplies a place name, a distance, a count or a score. Every fact on the page comes from the spatial query. If the model were ever in a position to invent a supermarket, the architecture would be wrong.