Google puts WeatherNext 3 forecasts into Search, Maps and Gemini

Google has put a new weather model into its own products. WeatherNext 3 was announced on 3 September by Google DeepMind and Google Research. It now feeds forecasts in Search, Maps, the Gemini app and Google Cloud.

WeatherNext 3 forecasts every hour on a five-kilometre grid

The model reads live data from geostationary satellites instead of running the usual physics simulation. It works out temperature and moisture on a grid of about five kilometres, surface readings at ten kilometres and wind at twenty-five. The previous version worked on a twenty-five kilometre grid and produced a forecast every six hours. This one produces one every hour.

Google says the model carries roughly 2.4 times more parameters than WeatherNext 2. Ferran Alet, a research scientist manager at DeepMind, told TechCrunch that machine learning targets “approximate noisy physics from incomplete information”. That is the company’s own framing of what the model does.

Google says day-ahead rain forecasts improve by up to 50 percent

The headline figure from Google is up to 50 percent better precipitation forecasts a day ahead. Measured against NASA’s IMERG data the improvement is 60 percent, against the MRMS radar set 30 percent, and against rain gauges about 10 percent for short lead times. These are Google’s own test results, not an outside audit.

Weather questions now end inside Search rather than on a website

The part that matters for publishers is where the answer lands. A weather question used to send someone to a forecast site. Now Search, Maps and Gemini answer it from the same model. It is one more query type that stops before the click, alongside the shift Google has been building through AI Mode and its Flash models. Publishers with weather, travel or outdoor content should check their own numbers in the AI Overviews report in Search Console rather than assume the traffic is stable.

Ad triggers that read the forecast get a finer signal

There is a second, quieter use. Retailers already switch creative on the weather. In August Ace Hardware added weather triggers to its retail media network. Hourly forecasts on a five-kilometre grid make that kind of rule sharper, because a rain rule can follow a single city rather than a whole region.

Google also lists wind speed at 100 metres, cloud cover and sunlight in the model, which is aimed at energy planning. Marketers do not need that layer. They do need to know that the weather signal behind their campaign rules may now update far more often than their reporting does.