Earth AI’s geospatial agents need a better question than “map this”

More than a chatbot with a globe

On October 14, Google Research described advances to its experimental Geospatial Reasoning agent, which combines specialist foundation models, spatial tools and planning. The update includes a technical report and expanded early access through Ask Google Earth and an agent-to-agent service.

Google reports partner testing in areas such as disaster preparedness and clean-energy screening. This remains experimental access, not a general guarantee that any map question can be answered or that a generated analysis is suitable for a real-world decision. We have not run the agent or independently verified its reported results.

Source: Google Research: Putting Earth AI to work — Geospatial agents for our planet’s biggest challenges ↗

Write a map brief before a map prompt

Our practical suggestion is a short question contract. Define the area, the date range, the thing being measured and the kind of answer you need. Add the question you do not want the system to answer.

For a harmless classroom example, imagine comparing publicly mapped tree cover around two campuses. Do you mean the property boundary or a surrounding walking area? Are you comparing the same year? Do you want total tree-covered area or the share of each campus covered? Each version asks something different, even if all four maps look appealing.

Then name the limits. Tree cover alone should not become a claim about student well-being, access to parks or the quality of a neighborhood. Those conclusions would require other evidence and a separate analysis. A boundary in the question prevents an attractive map from silently answering a much bigger question.

Ask what each layer contributes

Before accepting a result, request a plain-language account of the inputs and transformations. Which dataset was used? What period does it describe? Was information averaged across an area? Did a model estimate something that was not directly observed? Keep unresolved answers visible alongside the map.

A useful comparison changes one decision at a time. In the fictional campus exercise, first use the same boundary rule for both places. Then try a second boundary rule and see whether your conclusion changes. If it does, report that sensitivity instead of selecting whichever version produces the most dramatic contrast.

This is a proposed reasoning exercise, not a claim about the behavior of Google’s system. You can practice it with an existing public map and a notebook; access to an experimental agent is optional. For emergency, infrastructure or health decisions, rely on the responsible authorities and qualified analysis, not a classroom exercise or an AI-generated map.

The promising idea is that an agent can coordinate tools needed for a spatial question. The reader’s job is still to understand what question those tools actually answered. A map becomes more useful when its boundaries are clear, even when its colors become less exciting.