The brief
Open a GIS dataset and you get polygons and attributes. Before any useful analysis can start, a planner has to mentally reconstruct what the place actually is from a list of numbers. Ask a language model to do the same thing and it has no way to perform that reconstruction at all.
A coordinate array says nothing about whether a street is vibrant or desolate, safe or hostile, historically loaded or finished last year. Attributes help a little, and standard formats offer them as a flat bag of properties with no hierarchy, no relationships and no indication of what any of it means to the people using the space.
This is not a failure of GeoJSON or WKT. Those formats were built for geometry and they are excellent at it. The gap is that spatial workflows now run through language models, and nothing exists that is optimised for reasoning about place rather than rendering it.
The problem arrived from the other direction first. Building WalkGrid meant scoring 51 environmental features across a hexagonal grid, and every one of those scores had to live in a column because there was nowhere else to put it. A format that could carry the meaning alongside the geometry would have made that system simpler, and no such format existed.
The insight
The move is to treat a place as a semantic object rather than as a shape with labels attached.
GEON takes its cue from TOON, which showed that describing objects by purpose and behaviour rather than by physical properties unlocks substantially better model reasoning. The spatial equivalent is a format where every field name carries meaning, every value is interpretable, and relationships are named rather than implied by coordinate arithmetic.
What that buys is semantic density: reasoning-relevant information per token. A model asked which public spaces near a station suit outdoor dining can actually reason over PURPOSE, EXPERIENCE and ADJACENCIES, instead of attempting geometric computation it was never built for. The token comparison in Fig. 1 is the same argument from the other direction, since a context window spent on coordinate precision is a context window unavailable for thinking.
The second consequence took longer to notice. Because the format is legible to a model in both directions, it supports generative spatial reasoning. Give a model a GEON description of an underused street and it can propose interventions and express them back in GEON, which closes a loop between analysis and design that previously required a human at every handoff.
What it does not do
Experiential fields are observer-dependent, and honestly so. One field visit in February does not capture a place, two researchers will not agree on visual complexity, and no amount of controlled vocabulary makes a subjective judgement objective. The fields record an observation with provenance, which is the strongest claim available and weaker than a measurement.
Human curation is simultaneously the point and the bottleneck. Rich documents require someone who knows the place to sit down and write one, and that does not scale to a city, let alone a country. Generation from OpenStreetMap gets a baseline in place but cannot populate the fields that justify the format.
It is version 0.1.0 with community conventions still forming. There is no governance process yet for vocabulary extensions or domain-specific fields, no empirical evaluation of whether GEON representation measurably improves downstream model performance on spatial reasoning benchmarks, and no temporal, 3D or real-time variants beyond sketches. Calling it a standard would be premature. It is a proposal, versioned so it can be argued with.
