Google Search VP Nick Fox: WeatherNext 3 Integrates into Search and Maps
Nick Fox stated that weather is one of Google's top daily searches. The company has integrated its advanced weather model WeatherNext 3 directly into search and maps, achieving up to a 50% improvement in long-term precipitation forecast accuracy, particularly in areas that historically had poorer forecasts.
On the same day, Google DeepMind and Google Research published that the model connects to real-time geostationary satellite mosaics, initializing every hour. Compared to WeatherNext 2's 25 km, 6-hour steps, surface temperature and dew point can reach about 5 km, while other surface variables are about 10 km, and three-dimensional atmosphere is about 25 km. For medium-term precipitation, the relative improvement over the numerical forecast baseline is about 60% for IMERG's CRPS, about 30% for the U.S. MRMS radar, and about 10% for early gauge timing. Training also utilized NASA IMERG and proprietary satellite radar reanalysis. Outputs include 100 m wind speed, cloud cover, and solar radiation components, aimed at wind power and photovoltaics. A 64-member ensemble is used. The 6-hour cycle forecast extends to 15 days, with other hourly initialization forecasts up to 48 hours. Brightband real-time evaluation ranks it among the most accurate global models at the time. Data is made available to enterprises via BigQuery, Earth Engine, and cloud storage Zarr. The product is integrated into search, Gemini, maps, map platform weather interfaces, and Earth Engine. GraphCast and GenCast forecast datasets have been deprecated.
Who is buying and who is selling: The event is driven by integrating global forecasts into the default search box, not by meteorological agency shutdowns. Buyers include users and enterprise interfaces for weekend travel and logistics scheduling, while sellers are platforms that refresh frequency through direct satellite connections. Beneficiaries include search and map dwell time and weather API billing, while those under pressure are applications and national models that still update every 6 hours. Funds are flowing from advertising and Maps Platform into ensemble inference, with reduced precipitation errors directly increasing the click-through reliability of "should I take an umbrella" queries.
Source: Public Information
ABAB AI Insight
Weather queries are one of the most frequent, localized, and easily verifiable intents in search. Users can immediately verify if a forecast is wrong, so improvements in accuracy will quickly translate into default engine stickiness. Fox's announcement of the model as a search feature indicates that DeepMind's weather capabilities are no longer confined to papers and Earth Engine but aim to dominate decision-making moments.
The capital path starts with GraphCast proving that AI can outperform traditional numerical models, followed by WeatherNext productizing resolution and hourly refresh, and finally recouping data and computational costs through distribution in search and maps. The focus is on selling the ability to "plan whether it will rain in the next day or two." Renewable energy variables allow the same output to be sold to power station dispatch, creating a shared backbone for consumer and industrial use.
Comparative objects include the European Centre for Medium-Range Weather Forecasts operational systems, national radar and rain gauge networks, and mobile weather applications that still rely on national models. The industry's position is shifting from research-oriented AI forecasting to expanding into default consumer entry points, with control resting on who integrates the latest fields into billions of queries.
The structural change is a transfer of pricing power. The mechanism is that forecast errors shift from being part of professional products from public meteorological agencies to becoming part of platform experiences; whoever reduces precipitation Brier scores will enhance the conversion of local life queries. Hourly satellite initialization shortens the traditional six-hour lag of analysis fields, preventing convective rain bands from being smoothed out by 25 km grids.
ABAB News · Cognitive Laws
The most searched questions daily deserve a dedicated model training.
As the forecast grid tightens, localized showers emerge from the averages.
Direct satellite input into the model is closer to what is seen when looking up than waiting for analysis fields.