Can you ask a map questions in plain English?

Yes, if the layers are clean and the words are defined. Consumer maps now take plain questions about places. Asking a business's own map layers for analysis, such as which stands near the road were recently harvested, is harder for two reasons. The layers have to be clean: consistent projections and sound geometry. And the question's words have to mean something fixed. Near might be 200 metres or two kilometres, and recently might mean this year or the last five. An AI answering has to choose, and nothing in the map it returns shows which choice it made. Plain-language map questions work when a business has fixed both the layers and the words.

What the usual answer says

The pages that rank for this question are about consumer maps. Google's help page says users "can get recommendations and answers from Ask Maps". That is a different task: finding places, not measuring a business's own stands, roads and harvests. Descriptions of natural-language GIS for business make a similar promise: anyone can ask a map a question and get an answer instantly. They leave out what happens to the parts of the question the words don't pin down.

What the benchmarks show

GISAgentBench, a preprint from August 2026 by Abhinav Pothuri and colleagues, built its tasks from practitioners' real questions and scored agents against exact outputs. Its full text names the language problem directly: "practitioner requests are often underspecified, leaving the intended spatial predicate, measurement unit, or attribute schema to be inferred". Its example is a request for features inside a boundary, which "does not say whether features touching that boundary should be retained". The consequence is the dangerous part: such requests "produce syntactically valid but semantically wrong outputs rather than failed calls, so they cost accuracy without raising the error rate".

The same paper ranks the damage, and the data mattered more. Tasks involving geometry and topology errors or mismatched coordinate systems scored worst, and ambiguous wording and unit mismatches cost less. A vocabulary helps, and clean layers come first.

An earlier benchmark shows why valid-looking output reassures. GeoAnalystBench, published by Qianheng Zhang and colleagues in Transactions in GIS in 2025, set models 50 code-generation tasks from real geospatial problems. One commercial model produced valid workflows 95% of the time, while a small open model managed 48.5%. The hardest tasks for every model were those "requiring deeper spatial reasoning, such as spatial relationship detection or optimal site selection". A valid workflow is not a correct answer.

Where the words go wrong

Spatial words carry numbers and rules that plain English omits. Near needs a distance and a point to measure from. Adjacent needs a rule: sharing a boundary, or within some gap. Along a stream needs a buffer width and, often, a regulation behind it. Recently harvested needs a date range and a source, such as the harvest plan, the scale records or the satellite. Each choice changes the answer, and the map looks equally convincing whichever was made.

This is the spatial version of the problem that business definitions solve for ordinary data. A margin question needs a written definition of margin, and a map question needs a written definition of near.

A spatial vocabulary

Write down the spatial words the business actually uses, and define each one in a line: the distance, the layer measured from, the rule. Near a road might be within 300 metres of the road centreline layer. An adjacent stand might be one that shares a boundary. Along a stream might mean within the riparian buffer set by the applicable rules. Then require the AI to state, with every answer, which definitions it applied, so a forester can judge the answer quickly. In our experience the list is short, and it is worth having without AI, because it settles arguments between people who use the same word differently.

Quarri's own work on map questions is in design, not delivered, and this piece describes the general approach, not a product.

When it doesn't apply

Questions without spatial terms, such as the area of a named stand, need no spatial vocabulary. Exploratory questions, where a rough first map is enough to decide where to look, can accept the AI's defaults. And where definitions are set by regulation, such as buffer widths along watercourses, the vocabulary should quote the rule rather than restate it.

Quarri for forest management is built around how a forest operation runs, from the cruise to the settled account.

Quarri is an AI-native data platform for the timber supply chain. It connects buying, production, sales and inventory for forest management, sawmill, wood products and pulp, paper and packaging operators.