Google today put a new global weather model behind Search, the Gemini app, Maps, the Maps Platform Weather API and Earth Engine, moving forecast production from six-hour cycles on a 25-kilometre grid to hourly cycles resolved at five kilometres, with precipitation accuracy gains the company puts at up to 50% for forecasts a day or more ahead.

The model is called WeatherNext 3. According to Google, it was built jointly by Google DeepMind and Google Research, and the company describes it as the most advanced and accurate global weather model to date, citing independent live evaluations run by Brightband. That framing rests on a third party's leaderboard rather than on figures published in the announcement itself, which is a distinction worth holding onto: the post dated September 3, 2026 points readers to Brightband's rankings without reproducing them.

What the post does contain is a fairly detailed account of how the model differs from its predecessor, and the differences are structural rather than incremental.

Five kilometres, and a forecast every hour

Resolution is the headline. According to Google, WeatherNext 3 renders key surface variables such as temperature and moisture at a 5-kilometre resolution, other surface variables at 10 kilometres, and atmospheric variables including wind speed at 25 kilometres. The previous model, WeatherNext 2, produced forecasts on a 25-kilometre grid in 6-hour increments. Google characterises the net effect as a global weather picture roughly five times sharper.

The temporal change is the one with more operational consequence. A six-hour increment means a forecast that is, at worst, nearly six hours stale before it is replaced. An hourly cycle compresses that window to under sixty minutes. Google states the model generates a new forecast every hour, each grounded in the most recent satellite observations available, at up to 5-kilometre resolution.

Architecture is described in the post's first figure. The system ingests live one-hour geostationary satellite mosaics alongside traditional historical analysis, feeding both into a single Functional Generative Network mesh transformer. Outputs are not limited to a gridded field. According to the figure caption, the model produces dense gridded fields, discrete cyclone tracks, and station-level sparse coordinates natively, meaning the point forecasts are generated directly rather than interpolated out of a grid after the fact.

That last detail matters for anyone who has ever tried to reconcile a point forecast against a gridded one. Interpolation carries its own error. Predicting station coordinates natively removes a step.

What the model learns from

The more consequential change is upstream of the architecture, in the training data.

Most AI weather models, WeatherNext 2 among them, are trained on output from numerical weather prediction models. According to Google, those systems are supercomputer-driven physics simulations that carry a six-hour data lag, and that lag can produce biases for fast-changing variables such as rain and surface temperature. The model is learning from a simulation of the atmosphere rather than from the atmosphere.

WeatherNext 3 changes the input. Google says the model ingests a mosaic of live global geostationary satellite data, giving it a continuously updating view, and that it also trains directly on sparse weather station observation data. The stated purpose of the second source is regional fidelity: temperature and humidity can vary sharply over a few kilometres near coastlines, valleys and mountain ranges, and models trained on smoothed representations of the atmosphere miss those variations.

Google frames the geographic consequence explicitly. According to the company, the approach is particularly relevant for regions across Latin America, Africa and Asia-Pacific that have historically been underserved by high-resolution forecasting because of the supercomputing costs attached to traditional regional models. High-fidelity forecasting has been, until now, a function of national meteorological budgets. A single global model trained on satellite and station data changes the cost structure of who gets one.

There is a caveat Google itself supplies. The post carries a disclaimer stating that for official weather forecasts, severe weather warnings and public safety advisories, readers are directed to their local meteorological agency or national weather service. The model is not being positioned as a replacement for official warning infrastructure.

Precipitation, and the numbers behind the claim

Rain is where global models have traditionally performed worst, and it is where Google has put its most specific figures.

According to the announcement, precipitation is driven by fast-moving cloud processes on scales too small for physics-based simulation to capture cleanly, which is why AI forecasts often produce blurred estimates or miss the boundaries of severe storms. To address that, Google trained the model on two precipitation datasets: NASA's satellite-based Integrated Multi-satellite Retrievals for GPM, known as IMERG, and a Google global precipitation reanalysis built on satellite radar.

The reported result is expressed in Continuous Ranked Probability Score, a probabilistic accuracy metric. In medium-range global forecasts, Google reports CRPS improvements against baselines of up to 60% against IMERG, 30% for MRMS, and 10% against rain gauge measurements for early lead times. The post does not expand the MRMS acronym, does not name the baselines the comparison runs against, and does not state the lead times at which each figure was measured. Those are material omissions for anyone attempting to replicate or audit the claim.

A separate number covers the consumer surfaces. Google states that when planning a day or more ahead, people will see up to 50% more accurate precipitation forecasts, with the greatest improvements in regions where forecasts have historically been less reliable. The phrasing is a ceiling rather than an average, and no distribution is published.

One technical nuance sits in the third figure. The medium-range probability-of-precipitation comparison shows WeatherNext 2 at 25-kilometre resolution against WeatherNext 3 at 11-kilometre resolution, not at the 5-kilometre native resolution quoted elsewhere in the post. Google does not explain the difference between the two figures, which suggests the precipitation product and the surface-variable product run at different resolutions.

Wind at turbine height

Beyond resolution and cadence, the model adds a set of variables aimed squarely at energy markets.

According to Google, WeatherNext 3 forecasts 100-metre wind speeds, roughly at turbine height, for wind-energy output estimation, alongside high-resolution cloud cover and solar radiation levels intended to help solar farms estimate incident light at ground level. The company describes the data as crucial for clean energy planning, allowing grid operators and renewables developers to predict generation and match it against consumer demand.

This is a narrower audience than the consumer weather tile, but a commercially dense one. Day-ahead power markets clear on generation forecasts. An hourly-refreshed wind forecast at hub height is a tradeable input, not a convenience feature.

Where the model lands

Distribution is the part of the announcement with the most immediate reach.

According to Google, WeatherNext 3 will begin powering weather experiences within Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API and Google Earth Engine starting today. For developers and businesses, the company says global predictions updated hourly can be queried in BigQuery and Earth Engine or bulk-downloaded from Google Cloud Storage, with no model setup required.

The surfaces named are not marginal. Google's Search chief told I/O 2026 that AI Mode had passed one billion monthly users, and Gemini 3 became the default model behind AI Overviews globally in January 2026. Gemini began serving local results pulled directly from Google Maps in December 2025, and Ask Maps has since accumulated hotel information, local events and food ordering. As recently as September 2, AI Mode moved to Gemini 3.8 Flash. A weather model wired into that stack reaches a very large share of daily consumer queries by default rather than by request.

The Maps Platform Weather API is the piece most likely to change third-party products. Google Maps Platform pricing has been contested before, with a 2018 restructuring that required billing on all projects and a 2024 reduction of up to 70% for Indian developers. The announcement carries no pricing detail for the upgraded Weather API, and no statement on whether existing quotas or rate limits change.

Why an hourly forecast matters to media buyers

Weather has been a targeting signal in programmatic advertising for well over a decade, and the constraint has rarely been the creative logic. It has been the granularity and freshness of the underlying feed.

The commercial case is easy to state. On September 2, Ace Hardware's retail media network RedVest Media set out weather-triggered programmatic advertising that adjusts creative messaging automatically as rainfall or snowfall conditions change, alongside control-group measurement, a DoorDash storefront integration and a creator partnership. Demand for snow shovels, generators, dehumidifiers and roof sealant tracks conditions a forecast can anticipate several days out. The trigger is only as useful as the forecast underneath it.

A six-hour refresh on a 25-kilometre grid supports a campaign that switches creative by region and by day. An hourly refresh on a 5-kilometre grid supports something narrower: a campaign that switches by neighbourhood and by hour. Whether buyers can act on that depends on their own decisioning latency, not on Google's, and most programmatic stacks are not built to re-key creative rules hourly. The signal improving does not automatically mean the activation improves with it.

The format most likely to absorb the change first is digital out-of-home, where a screen's location is fixed and known, and where conditions at that specific coordinate determine whether an umbrella advertisement or a sunscreen advertisement is the right play. Programmatic DOOH has spent two years building the plumbing for exactly this kind of conditional delivery. Location precision at five kilometres is close to the resolution at which an individual panel's catchment area becomes meaningful.

Data-side vendors have already been treating weather as a confounder rather than a trigger. Adsquare's Control Condition methodology, used in its attribution products and rolled into Amazon DSP and The Trade Desk integrations during 2026, simulates behaviour for unexposed audiences while accounting for seasonality, public holidays and weather conditions. Adjusting for weather requires knowing the weather. Better inputs improve control models as readily as they improve targeting.

The measurement problem the signal creates

There is a reason weather-triggered advertising has an awkward relationship with attribution, and improving the forecast makes it sharper rather than softer.

When the same condition that fires the advertisement also drives the purchase, conventional attribution credits the campaign with sales the storm would have produced anyway. This is a textbook case of endogeneity: the exposure is correlated with the unmeasured driver of the outcome. The remedy is experimental. The IAB's Guidelines for Incremental Measurement in Commerce Media, published on November 3, 2025, rank randomised control tests, holdouts and matched markets at the top of the causal hierarchy and platform-reported metrics at the bottom.

The same trap catches footfall attribution in weather-sensitive categories. A garden centre sees more visits on the first warm Saturday in March whether or not it advertised. Matched-visit reporting establishes sequence, not cause. A finer-grained weather signal makes the targeting tighter and the naive attribution number larger at the same time, which is precisely why incrementality testing becomes more rather than less necessary as the trigger improves.

Elsewhere in the category, The Weather Company has spent roughly a year working with the neuroanalytics firm Neuro-Insight on what it terms emotional targeting, testing which neural pathways activate when advertising is processed alongside particular forecast data, with a commercial test run alongside S.C. Johnson. No sample size or effect size has been published. AccuWeather and The Weather Company were both among the 23 launch member organisations of the Ad Context Protocol in October 2025, which places weather data firmly inside the agentic advertising standards conversation rather than outside it.

What Google did not publish

Several gaps sit in the announcement, and they are the sort that determine how much weight the claims can carry.

The accuracy superlative is attributed to Brightband's independent live leaderboards, but no scores, margins or competing models are named in the post. The CRPS figures are given as ceilings without lead-time breakdowns or named baselines. No pricing, quota or rate-limit information accompanies the Maps Platform Weather API upgrade. No adoption or query-volume figures are given for any of the consumer surfaces. And while the post links to a research paper, the announcement text itself does not summarise the evaluation protocol.

The company is candid about the limit of the exercise. According to Google, the atmosphere will always retain a degree of unpredictability, and the argument made for WeatherNext 3 is narrower than a claim of solved forecasting: training on real-world observations rather than on simulation output moves the forecast closer to what is happening on the ground.

For advertisers, that is the practical read. A weather signal that refreshes hourly at five kilometres, distributed free through consumer surfaces and priced through a cloud API for everyone else, changes what a weather trigger can plausibly resolve. It does not change the measurement discipline required to prove the trigger worked.

Timeline

Summary

Who - Google DeepMind and Google Research, publishing jointly under the WeatherNext team byline. The affected parties include Google Search, Gemini app, Google Maps and Earth Engine users worldwide; developers building on the Google Maps Platform Weather API, BigQuery, Earth Engine and Google Cloud Storage; grid operators and renewable energy developers; and advertisers running weather-conditional creative in programmatic and out-of-home channels.

What - WeatherNext 3, a global AI weather model generating hourly forecasts at up to 5-kilometre resolution for surface variables such as temperature and moisture, 10 kilometres for other surface variables and 25 kilometres for atmospheric variables including wind speed. It trains on live geostationary satellite mosaics and sparse weather station observations rather than on numerical weather prediction output, uses a Functional Generative Network mesh transformer, and outputs gridded fields, cyclone tracks and native station-level point forecasts. Precipitation training draws on NASA's IMERG dataset and a Google satellite radar reanalysis, with reported CRPS improvements of up to 60% against IMERG, 30% for MRMS and 10% against rain gauges at early lead times. New variables include 100-metre wind speeds and high-resolution cloud cover and solar radiation.

When - Announced on September 3, 2026, with integration across Google Search, the Gemini app, Google Maps, the Google Maps Platform Weather API and Google Earth Engine beginning the same day.

Where - Global. Google specifically identifies Latin America, Africa and Asia-Pacific as regions historically underserved by high-resolution forecasting because of the supercomputing cost of traditional regional models.

Why - Numerical weather prediction models carry a six-hour data lag that biases fast-changing variables such as rain and surface temperature, and their coarse grids smooth out local variation near coastlines, valleys and mountain ranges. Training on live satellite and station observations removes the intermediate simulation step. For advertising specifically, the change moves the weather trigger from a daily regional signal to an hourly neighbourhood-scale one, which sharpens conditional creative and simultaneously widens the gap between attributed and incremental results.