Google is expanding access to an experimental artificial intelligence system designed to analyse planetary data, identify public health risks and support disaster preparedness, after early trials with international organisations demonstrated how complex geospatial investigations could be completed in minutes rather than weeks.
The system, known as the Geospatial Reasoning agent, combines Google’s specialised geospatial foundation models, mapping tools and satellite-derived information with a planning process that breaks complex questions into multiple analytical steps. In a technical report published on 14 October 2026, Google Research said the approach performed better than general-purpose large language models on several geospatial reasoning benchmarks, particularly when tasks required the coordinated use of multiple datasets and tools.
The company is expanding experimental access through Ask Google Earth and an application programming interface that allows developers to connect the agent to their own systems. Access remains limited to selected early users as Google gathers feedback and evaluates the technology’s performance in real-world applications.
One of the most significant demonstrations involved the 2026 Ebola outbreak in the Democratic Republic of the Congo. The World Health Organization’s Regional Office for Africa, working through its Emergency Preparedness and Response Hub in Dakar, used the system to map remote mining corridors where human mobility and potential exposure risks were high. By combining satellite imagery showing forest loss with information about mining sites and population distribution, the team identified 48 exposed settlements containing more than 45,500 people considered at risk. Google said the analysis took minutes, compared with the weeks the process would normally have required.
The findings helped local responders deploy mobile laboratories and coordinate border surveillance. Separately, the Epidemic Modeling and Intelligence Unit at the Democratic Republic of the Congo’s National Institute for Biomedical Research used Google’s Planetary Prediction Engine to estimate the risk of Ebola spreading into areas without reported infections. The analysis combined mobility patterns, historical case trends and Earth AI datasets to generate weekly insights for planning.
“Mobile populations like miners move faster than traditional paper surveillance can track, creating dangerous blind spots,” said Dr Etien Koua, HIS Program Area Manager and Hub Coordinator for the WHO Africa emergency preparedness and response operation. “Earth AI fills those gaps with satellite and mobility signals—moving us from reactively chasing a virus to actively forecasting where it will strike next.”
Google said the system had also been tested by organisations working on disaster response, community development, marketing and infrastructure investment. People in Need used it to analyse 14 historical cyclone paths and identify villages with high exposure, reducing a process that would otherwise involve days of manual geographic information system work to minutes. Its GIS and Remote Sensing Coordinator, Alonso Gonzales, said the tool converted prepared complex data into immediate insights, allowing analysts to devote more time to anticipatory action.
GiveDirectly used the technology during Tropical Cyclone Maila to calculate population exposure across different wind-speed zones and to assess tornado damage in Mississippi. By uploading its own map boundaries, the organisation was able to conduct the analysis within a single workspace, cutting crisis-scoping work from days to minutes.
The Milken Institute applied the agent to infrastructure planning, combining lists of more than 1,500 candidate buildings with Google Maps Platform’s Solar API and census data. The aim was to screen potential locations for community resilience hubs using factors including rooftop solar potential, population mobility, nearby transport hubs, climate risks and eligibility for federal investment zones. Advertising group WPP used the system to combine Google’s Population Dynamics Insights with proprietary data for geospatial marketing campaigns, in a case study that Google said demonstrated improved campaign performance while preserving the privacy of the company’s information.
The underlying technology connects Google’s Gemini models to a range of specialised resources, including WeatherNext, Flood Forecasting, the Population Dynamics Foundation Model, AlphaEarth Foundations and Remote Sensing Foundations. It can also draw on Google Maps Platform, Google Earth Engine, Data Commons and Google Dataset Search. Rather than attempting to process enormous quantities of spatial information within a single model context, the agent constructs an analytical plan, selects relevant tools and datasets, and executes the workflow in a dedicated workspace.
Users can also supply private data through the agent’s API in the form of CSV spreadsheets, GeoJSON map boundaries or cloud tables. The system can then combine those inputs with Google’s global models and datasets. Its responses identify the analytical plan and sources used, enabling experts to inspect assumptions, assess attribution and substitute alternative datasets where necessary.
The Planetary Prediction Engine extends these capabilities beyond answering spatial questions to training predictive models. Given local outcome data, such as regional disease rates, crop yields or electricity-grid outages, the agent selects relevant population and environmental indicators and uses Google’s Population Dynamics Insights and annual Satellite Embedding dataset to train models for spatial prediction, finer-resolution mapping and forecasting. Google reported that, in its public health research, combining population and satellite signals improved predictions of health outcomes by 12 per cent compared with demographic baselines and enabled mapping at eight times finer spatial resolution.
Google evaluated four agent configurations using Gemini 3.8 Flash and Gemini 3.1 Pro, comparing a standalone language model with systems progressively equipped with basic search and mapping tools, specialised Earth AI capabilities, and a dedicated planning stage. The assessment covered four benchmarks: CBGB and UnivEARTH, developed externally, and two new benchmarks, GeoExpertQueries and GeoKTools, created by Google to test complex analytical workflows and the correct selection of geospatial tools.
The results indicated that specialised tools substantially improved performance on several quantitative spatial tasks, while upfront planning was particularly important for open-ended investigations requiring multiple analytical stages. However, Google acknowledged that GeoExpertQueries was scored by an automated evaluator comparing outputs with expert-designed workflows rather than checking whether the final answers were correct. The company therefore described the system’s output as a strong first draft for experts to review and develop, not a definitive answer.
The expanded experimental access includes an interactive deep-analysis preview in Ask Google Earth, an experimental agent-to-agent service for developers building customised workflows, and early-preview access to the Planetary Prediction Engine. Google said feedback from its initial partners had already informed improvements to road-network routing, support for larger uploaded datasets and the discovery of authoritative local data sources.
