Google publishes benchmarks for its mapping AI and opens limited early access

On October 14, 2026, Google Research published a technical report, released two benchmark development sets and widened limited early access to its experimental Geospatial Reasoning agent.

The system connects Gemini to specialized models and data for questions involving places, populations and environmental conditions. Its tools include weather and flood models, satellite-image models, Google Maps Platform information and Google Earth Engine. Rather than asking a language model to handle the whole analysis alone, it gives the agent access to those specialized capabilities.

Before running a complex task, the agent lays out an analytical plan and selects alternatives when local data is sparse. It processes large map layers in a separate workspace, outside the model’s limited working context. Google says this keeps bulky spatial data from overwhelming that context while preserving accuracy and tracking where the data came from.

Responses show the plan and data sources so experts can examine assumptions or substitute datasets. Results combine text, tables and maps. Organizations can also supply their own information through the developer interface: CSV spreadsheets, GeoJSON files describing geographic boundaries, or cloud tables. The agent can then combine those locations with Earth AI’s models and catalogs.

Google evaluated four configurations across four benchmarks using Gemini 3.8 Flash and Gemini 3.1 Pro. The comparison progressively adds tools and planning, allowing the researchers to assess their contributions rather than treating the complete agent as a single unexplained package.

• A standalone language model.
• The model with basic web-search and map-grounding tools.
• The model with Earth AI tools, but no upfront planning.
• The full agent, which plans its retrieval and processing steps before execution.

Two tests are new. GeoKTools uses templated questions with verifiable answers that require selecting and using one to four spatial tools. GeoExpertQueries instead tests open-ended work, including crisis resilience, site selection and public health, against expert-written workflows. Google released the development portions of both benchmarks for other researchers to use.

Google says specialized Earth AI tools improved results on GeoKTools, GeoExpertQueries and the external code-generation benchmark CBGB. Upfront planning was especially important for open-ended GeoExpertQueries tasks, helping prevent shortcuts such as substituting web searches for spatial analysis. The company’s conclusion is that tool access alone is not enough for these longer workflows.

The fourth benchmark, UnivEARTH, was an exception: a standalone model scored 93.6%. Google suggests that its source NASA articles and answers being publicly available online may explain that result. Meanwhile, GeoExpertQueries uses an automated evaluator to compare the agent’s steps with expert workflows, not to establish a correct final answer.

We therefore consider the agent’s output as a strong first draft for experts to review and build on, not a final answer.

Shravya Shetty and Luke Barrington, Google Research

Access remains limited to early users. Ask Google Earth is offering an experimental deep-analysis update for people who want to ask questions in natural language and explore results on a map. Developer teams can seek access to an experimental agent-to-agent API—a software interface for connecting the agent to their own agents and data pipelines.

Partner testing has already shaped the system. Google says feedback prompted additions including road-network routing, support for larger uploaded datasets and better discovery of authoritative local sources. In one reported test, People in Need used the agent to analyze 14 historical cyclone paths and identify highly exposed villages in minutes rather than days of manual geographic analysis.