Google Research Expands Geospatial Reasoning Agent Access

Executive Summary

• Google Research published a technical report demonstrating a Geospatial Reasoning agent that pairs domain-specific Earth AI foundation models with spatial tools and upfront planning, and states the combination enables complex spatial analysis that standalone LLMs cannot perform. Source
• The report releases two new benchmarks, GeoKTools and GeoExpertQueries, with development splits released in full to accelerate research on geospatial agents. Source
• Google Research expanded experimental access through a deep research update in Ask Google Earth and an Agent-to-Agent (A2A) service for developer teams building custom workflows. Source
• Early partners, including the WHO Regional Office for Africa, People in Need, GiveDirectly, the Milken Institute and WPP, tested the agent on disease outbreak, disaster preparedness, clean energy screening and marketing workflows. Source

Key Takeaways

• Google Research reports that domain-specific Earth AI skills produce a significant jump on GeoExpertQueries, GeoKTools and CBGB, while a basic-tools agent struggles on quantitative spatial tasks. UnivEARTH is the stated exception, where a standalone LLM scored 93.6%.
• The report attributes open-ended, real-world workflow performance to upfront planning, which keeps the agent on track across multi-step analyses and is described as necessary beyond narrower benchmark tasks.
• Partners could supply their own data. The API accepts CSV spreadsheets, custom map boundaries (GeoJSON) and cloud tables, allowing private locations to be joined with Earth AI models and catalogs.
• Access remains limited and experimental. Google Research frames the agent's output on GeoExpertQueries as a strong first draft for experts to review rather than a final answer, because scoring uses an autorater comparing process to expert workflows.

Google Research Technical Approach and Benchmark Design

The report describes an agent that connects Gemini to Earth AI foundation models, including WeatherNext, Flood Forecasting, the Population Dynamics Foundation Model (PDFM), AlphaEarth Foundations and Remote Sensing Foundations, alongside Google Maps Platform, Google Earth Engine, Data Commons and Google Dataset Search. Earth AI capabilities are packaged as modular skills the agent calls only when a question requires them. Before running complex operations, the agent builds a step-by-step analytical plan, selects fallback datasets when local coverage is sparse, and processes large map layers in a dedicated workspace outside the context window. Google Research states this avoids overwhelming agent context with massive spatial data while preserving accuracy and tracking data attribution, and that every response shows the plan and data sources used so experts can check assumptions or substitute datasets.

Evaluation compared four configurations across Gemini 3.8 Flash and Gemini 3.1 Pro: a standalone LLM; an LLM with basic web search and maps grounding tools; an LLM with Earth AI tools but no upfront planning; and the full Geospatial Reasoning agent that begins every query with a planning stage. The suite spans CBGB and UnivEARTH, both externally developed, plus the two new benchmarks. GeoExpertQueries uses complex real-world queries with expert-authored golden workflows in domains such as crisis resilience, site selection and public health. GeoKTools contains templated queries with verifiable answers, each requiring correct selection and use of one to four distinct geospatial tools.

Google Research Partner Deployments and Reported Outcomes

At the start of the 2026 Ebola outbreak in the Democratic Republic of the Congo, WHO AFRO's Emergency Preparedness and Response Hub in Dakar used the agent to map remote mining corridors where exposure risk and human mobility are high, cross-referencing forest loss from satellite images, mining sites from maps and web data, and population data. According to the report, the team pinpointed 48 exposed settlements and located more than 45,500 at-risk people in minutes, a process that normally would have taken weeks, allowing responders to deploy mobile laboratories and coordinate border surveillance. The Epidemic Modeling and Intelligence Unit at the DRC's National Institute for Biomedical Research used the Planetary Prediction Engine to estimate Ebola spread risk into uninfected zones by combining mobility flows with historical case trends and Earth AI models and datasets.

Dr Etien Koua, HIS Program Area Manager and Hub Coordinator for WHO AFRO Emergency Preparedness and Response, said mobile populations like miners move faster than traditional paper surveillance can track, creating dangerous blind spots, and that Earth AI fills those gaps with satellite and mobility signals. People in Need analyzed 14 historical cyclone paths to identify high-exposure villages in minutes rather than days of manual GIS work, according to Alonso Gonzales, GIS and Remote Sensing Coordinator. GiveDirectly uploaded custom map boundaries to calculate population exposure by wind-speed zones during Tropical Cyclone Maila and benchmark tornado damage in Mississippi. WPP connected Population Dynamics Insights with proprietary data while preserving privacy, and the Milken Institute combined more than 1,500 candidate buildings with the Solar API and census data to screen locations for community resilience hubs.

Google Research states that partner testing drove recent additions of road-network routing, support for larger user-uploaded datasets, and better discovery of authoritative local data sources.

Related: Report Finds Knowledge Gaps Stall Enterprise AI Agents

Google Research Planetary Prediction and Public Health Modeling

Beyond analytical questions, the agent powers the Planetary Prediction Engine. Given local outcome data such as regional disease rates, crop yields or grid outages, the agent 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 recent public health research cited by Google Research, combining population and satellite signals improved predictions of health outcomes by 12% over demographic baselines and enabled mapping at 8X finer spatial resolution. These figures are reported research results from a specific study, not general performance guarantees across other domains or datasets.

Google Research Access Paths and Developer Integration

Google Research is expanding early access to a limited group. Ask Google Earth offers a preview of expanded deep analysis capabilities inside an interactive chat experience where users ask complex spatial questions in natural language, analyze data including their own in Google Earth projects, and explore visual results on the map. For developer teams, an experimental headless Agent-to-Agent service allows early partners to connect the Geospatial Reasoning agent to custom agents and data pipelines, with step-by-step recipes for connecting to a Gemini for business environment. The Planetary Prediction Engine is also available for early preview through an interactive interface and via the Agent API. Access requires completing an interest form or joining the Early Access Pool.

For deeper context, see our Agentic AI analysis: "Can Composable Infrastructure Fix Enterprise AI Fragmentation".

Google Research Implementation Risks

The report names its own limitations. GeoExpertQueries scores come from an autorater that compares the agent's performance to expert workflows defining the steps an analysis should take, not the correct final answer, so Google Research characterizes the output as a strong first draft for experts to review and build on, not a final answer. Benchmark scores are averaged over three runs and each benchmark is scored differently, meaning comparisons should be made within a row rather than across rows. The agent's accuracy depends on dataset coverage, which is why fallback dataset selection exists when local coverage is sparse. Access is explicitly experimental and limited, and the reported 12% improvement and 8X resolution gains come from one public health study rather than a broad validation. Details on cost, latency, data residency and service-level commitments are not provided in the source.

Editorial independence disclosure: This article was produced independently from the underlying sources and reflects Business 2.0 News editorial judgment. Source note: all factual claims derive from the Google Research technical report announcement. Source

Additional coverage: IBM Brings AI-Powered US Open Fan Experience to Madison Square Park

What This Means for Practitioners

For developers and data teams evaluating geospatial agents, the report's central mechanical claim is that skills and planning matter more than model choice alone: a basic-tools agent struggles on quantitative spatial tasks, while the full agent with Earth AI tools and a dedicated planning stage performs substantially better across the reported benchmarks. Practically, that favors architecture decisions that expose tools as modular, inspectable skills and that move heavy map processing outside the model context. Because every response displays its plan and sources, procurement and analytics teams gain an audit trail for validating assumptions. Teams should treat access as experimental and plan for expert review rather than unattended decisions.