Google announced an expanded experimental Earth AI preview on October 14, 2026, inviting aid organizations, local planners and infrastructure teams to explore tools for disaster analysis and site planning.
The system works with geospatial data—information tied to places. Google’s technical report describes an agent that turns a plain-language question into a plan, retrieves relevant data and performs calculations across sources such as maps, satellite observations and population records.
From cyclone exposure to candidate buildings
Google reports two examples of early partner use: People in Need analyzed 14 historical cyclone paths to identify highly exposed villages. Milken Institute teams screened more than 1,500 candidate buildings for community resilience hubs, combining their own lists with information including rooftop solar potential, census data and nearby transport.
These are partner accounts published by Google, not independently verified findings about accuracy, time saved or better outcomes.
For the separate question of access to storm forecasts, see RohitAI’s coverage of WeatherNext 2.
What the evaluation actually measures
In Google’s published evaluation using Gemini 3.8 Flash, the full agent scored 62.7 on a 0–100 scale on GeoExpertQueries, an assessment of complex spatial analysis. The same model with specialized geospatial tools but without the full planning and execution guidance scored 52.2. Both figures are averages over three runs.
Crucially, the scoring protocol uses an AI judge to compare the agent’s recorded analytical work with expert reference workflows. It does not independently establish the correctness of the final answers. A score of 62.7 is therefore not a claim that 62.7% of planning decisions would be correct.
The practical implication is to assess the reasoning and inputs, not just the finished map. For a site-screening pilot, define a distance threshold and a geographic boundary, then have a domain expert check the datasets and conclusions. This is a suggested evaluation approach, not a tested result. Existing Ask Google Earth guidance says vague terms such as “near” can be interpreted arbitrarily.
How to apply
Google describes three ways to explore the preview:
Route | Intended task |
|---|---|
Ask Google Earth | Ask spatial questions and explore results on a map. |
Geospatial Reasoning Agent API | Connect the agent to an organization’s own software and data workflows. |
Planetary Prediction Engine | Build predictive models from local outcome data through an interface or the API. |
Apply through the interest form linked from Google’s announcement. Access remains limited and experimental; applying does not guarantee entry. The announcement and research access page reviewed for this article give no public preview price or general-availability date.
Before choosing a pilot, confirm regional eligibility, export options and data-use terms for that specific preview route. The existing Ask Google Earth documentation describes restrictions and data handling for that interface, but predates this expansion. It should not be treated as a complete specification for the new preview or API.
Reporting note: This AI-assisted brief draws on Google’s announcement, technical report, evaluation protocol and documentation. No hands-on testing was performed.