Google’s Gemini-powered Geospatial Reasoning agent scored 62.7 on a test of open-ended geospatial workflows, compared with 3.3 for the same Gemini 3.8 Flash model equipped with general search, Maps and code tools, a Google-authored preprint reports. The scores use a benchmark-specific 0-to-100 workflow scale, not percentages of correct answers.
Google Research announced on October 14 that it is expanding experimental access to the agent through an update to Ask Google Earth and an agent-to-agent service for developers. The technical report linked from the announcement was submitted to arXiv three days earlier. It tests whether specialized Earth data and a planned sequence of analysis steps help a language model answer questions that require more than looking up a fact.
What the agent adds to Gemini
The Geospatial Reasoning agent, or GRA, connects Gemini to Google Earth AI models and geospatial services, including satellite-image analysis, weather and flood forecasts, Google Earth Engine, Maps data and demographic information. These sources describe different parts of a place, so answering a question about a flood, a disease risk or a potential site can require selecting and combining several datasets.
GRA organizes capabilities into modular “skills” that it loads when a question needs them. Before retrieving data, it writes a multistep plan; it then runs the analysis in a separate workspace, keeping large map layers and tables outside the model’s text context. The final response includes the plan and data sources, which lets a specialist check the assumptions or substitute different data.
Geospatial tools raised Flash’s score; planning effects varied by model
The researchers compared four configurations using the same underlying Gemini model: the model alone; the model with general web, Maps and code tools; the model with GRA’s geospatial tools but without its upfront planning and execution guidance; and the full GRA. They ran each comparison with Gemini 3.8 Flash and Gemini 3.1 Pro.
The main open-ended test, GeoExpertQueries, contains 101 problems based on practitioner tasks, with 76 held out for testing. Each problem has a reference workflow built from analyses by three independent geospatial experts. An automated scorer compares answers with the evidence and analytical steps in those workflows; the authors say they validated the scorer against human judgments.
On GeoExpertQueries, Gemini 3.8 Flash with general tools scored 3.3 ± 0.8. Adding GRA’s geospatial tools and skills raised the score to 52.2 ± 5.4, while the full agent scored 62.7 ± 2.5. The “±” values are standard deviations across three runs of each configuration’s aggregate score. In this Flash result, adding geospatial tools accounts for most of the increase; upfront planning and execution guidance add another 10.5 points.
That planning gain did not appear with every model. Gemini 3.1 Pro scored 50.8 with GRA’s tools but no planning layer and 47.6 with the full agent. The results suggest that geospatial tools helped on this test, while the added planning layer’s effect depended on the model.
The scores measure workflows, not field accuracy
GeoExpertQueries scores how closely an answer follows expert workflows, not whether a real-world decision was correct. In another test, UnivEARTH, every tested configuration scored between 86.9 and 98.8, including models with no Earth Engine access. The authors describe this benchmark as saturated because its questions were derived from published NASA Earth Observatory articles that contain the answers; a high score therefore need not show that a model performed geospatial analysis.
The paper is by Google researchers and includes a coauthor affiliated with Spatial Bytes; the team developed the new benchmarks used in its evaluation. The report gives results averaged across three runs but provides no independent external rerun. Google describes access to GRA as experimental early access.
Sources
- Google Research: “Putting Earth AI to work: Geospatial agents for our planet’s biggest challenges” (Oct. 14, 2026)
- Google Earth AI Geospatial Reasoning Agent: Capabilities and a Systematic Evaluation (arXiv:2610.15750)
- Illustrative map: Blue Marble Earth Equal Projection, adapted from NASA imagery by Tom Patterson, CC0 1.0.
