Google has advanced its experimental Geospatial Reasoning agent — a system that links Gemini with Earth AI foundation models and planet-scale datasets to perform multi-step spatial analysis and to autonomously train predictive models. The company published a technical report with systematic evaluations, introduced two new benchmarks for the research community, shared insights from partner field tests, and is expanding limited early access to developers and invited users.
Technical report and released benchmarks
The technical report evaluates agent performance across four benchmarks. In addition to existing CBGB and UnivEARTH datasets, Google released two new benchmarks: GeoExpertQueries and GeoKTools. GeoExpertQueries contains complex, real-world queries paired with expert-authored golden workflows for domains such as crisis resilience, site selection, and public health. GeoKTools provides templated queries with verifiable answers that require correct selection and use of 1–4 distinct geospatial tools.
The report compares four agent configurations on Gemini 3.8 Flash and Gemini 3.1 Pro: (1) a standalone LLM, (2) an LLM augmented with basic web search and maps grounding, (3) an LLM with Earth AI tools (e.g., Remote Sensing Foundations, WeatherNext 3) but without upfront planning, and (4) the full Geospatial Reasoning agent that begins each query with a dedicated planning stage to determine capabilities, fallback data, and the stages of retrieval and processing required.
Two main takeaways emerge: domain-specific tools materially improve quantitative spatial performance, and upfront planning is necessary to tackle open-ended, multi-step workflows like those in GeoExpertQueries. The report also notes limitations: GeoExpertQueries scoring uses an autorater that compares agent steps to expert workflows, so outputs should be treated as strong first drafts for expert review rather than final answers. UnivEARTH is an exception where even a standalone LLM scored 93.6%, likely because source NASA articles and answers are publicly available.
How the agent is connected and how it works
The agent connects Gemini to Earth AI foundation models and broad data sources, including WeatherNext, Flood Forecasting, the Population Dynamics Foundation Model (PDFM), AlphaEarth Foundations, Remote Sensing Foundations, plus Google Maps Platform (place information, routing, air quality, rooftop solar potential), Google Earth Engine, Data Commons, and Google Dataset Search.
Earth AI capabilities are exposed as modular skills that the agent invokes only when a user question requires them. Before executing complex operations, the agent builds a step-by-step analytical plan, selects fallback datasets when local coverage is limited, and processes large map layers inside a dedicated spatial workspace. Each response displays the agent’s plan and the data sources used so experts can validate assumptions or substitute datasets.
The system also supports user-provided data: organizations can upload CSVs, GeoJSON boundaries, or cloud tables via the Geospatial Reasoning agent API and ask the agent to join their private locations with Earth AI’s global models and catalogs.
Planetary Prediction Engine and predictive gains
Beyond analysis, the agent powers the Planetary Prediction Engine. Given local outcome data (for example, regional disease rates, crop yields, or grid outages), the agent automatically selects relevant population and environmental signals, fuses Population Dynamics Insights and Custom Satellite Embeddings, and trains validated machine learning models for spatial prediction, super-resolution downscaling, and forecasting. In cited public health research, combining population and satellite signals improved predictions of health outcomes by 12% over demographic baselines and enabled mapping at 8X finer spatial resolution.
Evaluation results and key insights
The evaluation shows that equipping agents with Earth AI skills produces a significant improvement across GeoExpertQueries, GeoKTools, and CBGB benchmarks compared with basic-tool or standalone LLM configurations. However, for open-ended expert workflows (GeoExpertQueries), upfront planning is critical to keep the agent on track across multi-step analyses and to prevent unreliable assumptions or shortcuts (such as substituting web search for spatial tools).
The technical report provides more detail on methodology and the benchmark development splits for GeoExpertQueries and GeoKTools have been released to accelerate research.
Early partner testing and real-world use cases
Google gave early access to a small group of social-impact and commercial partners to test the agent on challenging, real-world workflows. Key examples include:
-
World Health Organization (WHO) Regional Office for Africa (WHO AFRO) & Democratic Republic of the Congo’s National Institute for Biomedical Research (INRB): At the start of the 2026 Ebola outbreak in the DRC, WHO AFRO’s Emergency Preparedness and Response Hub in Dakar used the Geospatial Reasoning agent to map remote mining corridors by cross-referencing satellite-derived forest loss, mining site data, and population datasets. In minutes they identified 48 exposed settlements and located more than 45,500 at-risk people, a task that would normally take weeks. Simultaneously, the INRB’s Epidemic Modeling and Intelligence Unit (UMIE) used the Planetary Prediction Engine to estimate risk of spread into uninfected zones, combining mobility flows, historical case trends, and Earth AI models to provide weekly insights for operational planning.
-
People in Need: the agent analyzed 14 historical cyclone tracks to rapidly identify high-exposure villages, reducing days of manual GIS work to minutes.
-
GiveDirectly: teams uploaded custom map boundaries to quickly compute population exposure by wind-speed zones during Tropical Cyclone Maila and benchmark tornado damage in Mississippi, compressing crisis scoping from days to minutes inside a single workspace.
-
WPP: used geospatial reasoning to combine Population Dynamics Insights with proprietary data to improve geospatial marketing campaign performance while preserving data privacy.
-
Milken Institute: pre-development teams combined lists of 1,500+ candidate buildings with Google Maps Platform’s Solar API and census data to screen locations for community resilience hubs, accounting for rooftop solar potential, mobility, nearby transit, climate risks, and federal investment zone eligibility.
Partner testing guided improvements such as road-network routing, support for larger user-uploaded datasets, and better discovery of authoritative local data sources.
Expanded early access and how to apply
Google is expanding experimental access to a limited set of early users for several components:
-
Ask Google Earth: a preview of expanded deep analysis capabilities directly inside Ask Google Earth, offering an interactive chat experience for complex spatial questions, analysis of user data in Google Earth projects, and map-based visual exploration.
-
Geospatial Reasoning Agent API: an experimental headless Agent-to-Agent (A2A) service for developer teams who want to power custom workflows and connect the Geospatial Reasoning agent to their own agents and data pipelines. Google provides developer recipes for integrating this service with Gemini for business.
-
Planetary Prediction Engine: available for early preview both through an interactive, plain-language modeling interface and via the Agent API for custom workflows.
Interested organizations can complete an interest form or sign up for the Early Access Pool to request participation.
Why it matters
The Geospatial Reasoning agent aims to bridge large planetary datasets and practical decision support. The technical evaluation and partner case studies suggest that domain-specific tools plus explicit multi-step planning substantially improve reliability and speed for spatial analysis — particularly in emergency response, public health, and infrastructure planning. Google emphasizes that outputs should be reviewed by domain experts given current limitations and the autorating approach used in some benchmarks.
