Google brings MetNet precipitation nowcasting to Search across Africa

What Google actually shipped

Google announced that short-term precipitation forecasts are now available across Africa directly in Search. This is not a new research paper or a standalone product — it is an existing capability surfaced through the query box people already use.

The forecasts come from MetNet, Google Research's nowcasting model. According to the announcement, it predicts global precipitation within a five-kilometer radius every 15 minutes for the next 12 hours, and produces those forecasts in under one minute.

The model relies on satellite data and ground observations, which Google specifically frames as producing state-of-the-art results in data-sparse regions — the technical detail that makes an Africa rollout meaningful rather than routine.

The interface is Search, and that is the design decision

The frontend choice here is quiet but deliberate. Google did not build a weather app, ask users to install anything, or publish an API for developers. The forecast appears inside Search results.

For a region where dedicated weather infrastructure and specialized apps may be uneven, embedding the output in Search means the delivery surface is one people already have and already trust for answers. The model's accuracy is only useful if it reaches someone before the rain does; Search is the shortest path to that person.

This reframes the achievement. The 'under one minute' generation time matters precisely because it fits a latency budget for on-demand queries — you can ask, and the answer resolves in the time a search result takes to load.

Why the data-sparse claim carries weight

Most weather forecasting quality tracks the density of ground sensors and radar. Africa has historically had sparser coverage than North America or Europe, which is exactly where conventional numerical models degrade.

MetNet ... uses satellite data and ground observations to produce state-of-the-art precipitation forecasts in data-sparse regions of the world.Montana Labs

The claim Google is making is that a satellite-and-observation model can partially substitute for the dense instrumentation that isn't there. That is the specific reason this expansion is more than a geographic checkbox — it targets the conditions where traditional methods struggle most.

The implication: a research model earns its keep at the point of delivery

For applied teams, the instructive part of this announcement is not the model — MetNet already existed. It is that the value was realized by routing an existing model's output into a high-traffic, zero-install surface.

A precipitation nowcast that lives in a paper reaches researchers; the same nowcast rendered in Search reaches a farmer or a commuter deciding whether to travel. The engineering that closes that gap — fitting the forecast into a sub-minute, query-driven response inside an interface people already open — is what turns a capability into a service.