Google’s experimental Geospatial Reasoning agent performs better on an open-ended expert benchmark when it writes a plan before it begins calling mapping and Earth-observation tools. With Gemini 3.8 Flash, the full agent scored 62.7 on GeoExpertQueries, compared with 52.2 for the same specialist tools and skills without upfront planning. A basic-tool configuration scored 3.3.[1][2]
The result is not a licence to let an AI make unreviewed disaster, health or land-use decisions. GeoExpertQueries grades whether the system’s analytical workflow resembles a consensus method written by geospatial experts. It does not directly certify that every final map, count or recommendation is correct. Planning improved methodological discipline under this test; expert verification remains part of the product claim.[2][3]
The plan decides which geography problem is actually being solved
A question such as “Which settlements are most exposed to an outbreak near mining routes?” contains several hidden choices: the outbreak boundary, settlement dataset, route definition, population vintage, distance method and treatment of missing areas. A generic language model may answer with plausible prose or search for an existing article. The Earth agent instead loads domain skills, selects tools, states a multi-stage workflow and keeps intermediate datasets outside the conversation window.[1][2]
Four configurations in the evaluation
| Configuration | What it can use | What the comparison tests |
|---|---|---|
| Standalone LLM | Language-model knowledge and ordinary reasoning | Whether the answer can be recovered without geospatial computationSource 2 |
| LLM plus basic tools | General web and map-style capabilities | Whether generic retrieval is enoughSource 2 |
| LLM plus GRA tools | Specialist Earth AI skills and 62 geospatial tools, without upfront planning | Value of the domain toolset by itselfSource 2 |
| Full GRA | Specialist tools, modular skills, execution guidance and an upfront plan | Added value of explicitly decomposing the analysisSource 2 |
The benchmark separates exact computations from defensible workflows
Two new evaluation sets
| Measure | Value | Context |
|---|---|---|
| GeoKTools development split | 85 queries from 19 templates | The tasks require one to four tool families and have programmatically verified answers.Source 3 |
| GeoKTools held-out split | 210 queries from 72 private templates | Private templates limit direct contamination and test compositional tool use.Source 3 |
| GeoExpertQueries development split | 25 prompts | Ten are solvable with the current toolset and fifteen intentionally exceed it.Source 3 |
| GeoExpertQueries held-out split | 76 prompts | The full set contains 101 audited prompts authored and cross-verified by three experts.Source 3 |
| GeoExpert score | 0, 0.5 or 1 per judged dimension | A model judge compares the execution trace with expert workflows across repeated runs.Source 3 |
On the exact-answer GeoKTools benchmark, adding specialist tools produced the dominant gain. With Gemini 3.8 Flash, the specialist-tools configuration scored 86.0 and the planned full agent 85.9—effectively level within the reported variation. On the open-ended GeoExpertQueries benchmark, planning mattered more: 62.7 versus 52.2. The design lesson is not “planning always wins”; it is that the value of planning rises when the task requires defensible dataset and method choices rather than one numerical lookup.[2]
Selected reported scores with Gemini 3.8 Flash
| Benchmark | Standalone LLM | Basic tools | GRA tools, no plan | Full GRA |
|---|---|---|---|---|
| GeoExpertQueries | 0.0 | 3.3 | 52.2 | 62.7Source 2 |
| GeoKTools | 40.5 | 40.0 | 86.0 | 85.9Source 2 |
| UnivEARTH | 93.6 | 98.3 | 98.8 | 98.1Source 2 |
A real outbreak example shows the promise—and the provenance problem
Google reports that work with the World Health Organization’s Africa office and the Democratic Republic of the Congo’s national biomedical institute identified 48 exposed settlements and more than 45,500 people at risk around an Ebola outbreak in minutes, a workflow the partners said would normally take weeks. That is a consequential partner-reported result, not an independently audited clinical trial or a general performance guarantee.[1]
The operational advantage is the ability to combine population, routes, satellite signals and user-supplied boundaries while preserving an explicit sequence of steps. The risk is equally clear: each result inherits the coverage, date, licence and error structure of every dataset and model in the chain. A polished map can conceal an old census, a missing road or a proxy variable that was reasonable only in another country.
Planning can become an audit record
The most valuable product feature may therefore be inspectability. A plan can record why a dataset was chosen, how a buffer distance was defined, which population year was used and where uncertainty enters. Reviewers can challenge those decisions before an attractive final visualization becomes institutional fact.
Google is expanding experimental access through Ask Google Earth and an agent-to-agent service, but the system remains an early-access research capability. Its benchmark result is strongest as evidence about architecture: specialist tools do most of the work, and an explicit plan can improve difficult open-ended analysis. The map is the output. The reasoning trail is what makes it reviewable.[1]
Sources & further reading
- Putting Earth AI to work: Geospatial agents for our planet’s biggest challenges Google Research · 14 Oct 2026 · Accessed 16 Oct 2026
- Google Earth AI Geospatial Reasoning Agent: Capabilities and a Systematic Evaluation Google Research / arXiv · 11 Oct 2026 · Accessed 16 Oct 2026
- Geospatial Reasoning Benchmarks: GeoKTools and GeoExpertQueries Google Research / GitHub · Accessed 16 Oct 2026