A latitude and longitude pair is one of the least informative facts a system can hold. It says where something sits in a reference frame. It says nothing about whether that location is inside a flood zone, five minutes from a hospital, on the wrong side of a river from the market it serves, or in the middle of a runway. Every question anyone actually asks about a place is a question about relationships, and coordinates carry no relationships at all.
Spatial intelligence is the layer that turns positions into those relationships, and then turns relationships into claims a decision can rest on. It is not a single algorithm. It is a stack: a geometry representation, a set of relational operations over that geometry, a semantic layer that says what the geometry represents, and a provenance layer that says how much any of it can be trusted.
The reason this is worth being precise about is that the consumer of spatial answers has changed. A human analyst looking at a map catches the obvious errors without noticing they caught them. They see the point in the ocean, they see the catchment that crosses a mountain range, they see the count that is three orders of magnitude off. A model consuming a JSON response sees none of that. It sees a number, and numbers are persuasive.
A coordinate is a label, not an understanding
Consider a single point: a warehouse, stored as two decimal numbers. On its own it supports exactly one class of query, which is retrieval by identity. Ask anything else and the point is insufficient.
Is the warehouse in the municipality that grants the permit? That needs a polygon and a containment test. Is it close to the highway? That needs the highway geometry, a distance measure, and a definition of close the question did not supply. Can a truck reach the port before curfew? That needs a road network, turn restrictions, a speed model, and a time. Is the surrounding area commercial or residential? That needs other features, a way of counting them, and a boundary within which to count.
In every one of those answers the coordinate contributed almost nothing. It served as an anchor, and everything of value was supplied by other data joined to it. Systems that treat coordinates as the valuable part, and joins as plumbing, end up with fast infrastructure that answers no useful questions.
The relations that carry the meaning
Nearly all spatial analysis reduces to a small set of primitive relations. Learning to see a business question as a composition of these is most of the skill.
- Containment. Is this geometry inside that one. It underpins jurisdiction, zoning, catchment membership, hazard exposure, and every administrative rollup. It is also the relation with the most brittle edges: a point exactly on a boundary, a polygon that partially overlaps, a building that straddles two districts.
- Proximity. How far is this from that, and is that near. Proximity is never meaningful without a distance metric and a threshold, and both are usually left implicit.
- Adjacency. Do these two areas share a border. Adjacency drives spread models, neighborhood statistics, and any analysis where influence passes across a boundary rather than through a network.
- Connectivity. Can you get from here to there, and at what cost. Connectivity is the relation that most often contradicts proximity, and it is the one most often missing from a dataset.
- Density. How much of something is in a unit of area. Density is the relation most vulnerable to the choice of unit, which is a large enough problem to deserve its own section below.
A question like "which of these five sites has the best access to labor" decomposes into connectivity from residential areas, population density within those areas, and containment within a commuting boundary somebody has to define. Written down, the arguable assumptions are visible. Left undecomposed, they are still there, buried in a score.
Straight lines lie
Straight-line distance, computed on a sphere or an ellipsoid, is cheap, has no dependencies, and is wrong for almost every human question.
The reason is that people and goods travel on networks, and networks are shaped by rivers, rail corridors, motorway access rules, one-way systems, gates, and land ownership. Two addresses six hundred meters apart with a limited access highway between them may be a twelve minute drive. Two shops on opposite banks of a river in a city with few bridges can be a kilometer apart and half an hour of travel. A point across an international border might be unreachable in any practical sense while sitting well inside any straight-line radius.
The error is not a constant factor that a multiplier can correct. It varies with the local network, which means it varies with exactly the geography the analysis is trying to distinguish between. A detour ratio calibrated on a dense grid city will flatter a fragmented delta city or a mountain valley town, and the analysis will conclude those places are better served than they are.
Straight-line distance still has legitimate uses. It is correct for physical phenomena that do not respect roads: noise, plume dispersion, radio coverage, the reach of a flood. It is also a valid first-stage filter, cheap enough to run over millions of candidates before spending real compute on routing the survivors. What it should not do is stand in for accessibility in the final answer, especially when that answer compares places, because that is where the varying error becomes a ranking error.
The answer depends on the boundaries somebody drew
Aggregate any spatial quantity and it must be aggregated into units. Change the units and the numbers change, sometimes enough to reverse the conclusion. This is the modifiable areal unit problem, and it has two distinct parts.
The scale part: the same underlying points aggregated into fine units and coarse units produce different densities, different variances, and different correlations. Coarser units average away variation, so relationships between variables usually look stronger at coarse resolution than they are at the level of individuals.
The zoning part: at the same resolution, different boundary placements produce different results. Shift a district line by a few hundred meters and a cluster of activity moves from one unit to another. Both maps are defensible. Both produce different rankings.
The consequences are constant. Administrative boundaries were drawn for governance, not analysis, and they often follow exactly the features that separate populations, so they cut through the phenomena being measured. Rankings of areas by rate are dominated by small denominators, so the top of any per capita table is mostly noise from the least populated units. Correlations between area aggregates say nothing reliable about individuals, and that inference is a well known fallacy committed in almost every quick analysis.
Nothing makes the problem disappear, but there are honest responses. Compute at the finest unit the data supports and aggregate upward for presentation only. Test whether the conclusion survives a different plausible boundary set, and report that it did or did not. Prefer boundaries drawn from the phenomenon, such as a catchment derived from travel time, whenever the question is about behavior rather than governance. State the unit next to the number, because a density with no stated unit is not a measurement.
What a human supplies without noticing
Give an experienced analyst a map with a point placed in the sea near a coastal city and they will not conclude that there is a warehouse in the sea. They will assume a sign error, a swapped coordinate order, or a datum problem, and they will check. This correction is instant and mostly unconscious, and it rests on a large body of assumed knowledge: that buildings sit on land, that a road ends somewhere, that a city has a rough size, that a hospital serves a population and therefore is not usually in a desert, that a river runs downhill.
A machine has none of this unless it is written down. Which means a spatial system has to carry, explicitly, several things that a human never has to articulate:
- Plausibility constraints. Which geometry types can legally intersect which others, what ranges a value can take, what a physically impossible result looks like.
- Scale awareness. That a difference in the fifth decimal place of a coordinate is meters and irrelevant to a regional question, while the same difference matters for a curb cut.
- Type expectations. That the thing called a school in one dataset may be a building footprint, a point at the entrance, a property parcel, or an administrative record with an address that was geocoded to a street centroid.
- Temporal expectations. That features appear and disappear, and that comparing two vintages is a different operation from querying one.
Encoding these is unglamorous work and it is where most of the reliability comes from. A validation rule that rejects a building centroid in open ocean prevents a class of downstream nonsense that no amount of model quality will catch later.
Semantics have to be explicit, including the doubt
The hardest part of spatial data is rarely the geometry. It is the meaning attached to it. Two datasets both containing a field called type, both containing the value hospital, may disagree on whether a clinic with eight beds qualifies, whether a veterinary facility counts, whether a closed facility is still listed, and whether the record refers to the institution or the building.
This is the classification layer, and it needs three things stated rather than assumed. The definition: what real world entities the label covers, in enough detail that two people would classify the same borderline case the same way. The source: an official register, a volunteer contributor, a vision model reading a storefront, or a text match on a name. The confidence: how likely the label is to be correct, which is a different quantity from the positional accuracy of the geometry.
That last point is routinely collapsed. A feature can be perfectly located and wrongly typed. A building footprint traced to sub meter accuracy from imagery is a precise geometry with no reliable class at all: the imagery shows a roof, not a function. When the class comes from a separate join, confidence in the class is the confidence of that join, not the accuracy of the outline. Systems that report one quality number usually mean positional accuracy, and consumers read it as overall trustworthiness.
Confidence also has to survive aggregation. Counting fifty features of which twenty are guesses yields a count of fifty with an unstated error bar that is neither symmetric nor small.
Three questions any spatial claim should answer
A useful discipline, applied to any figure before it leaves the system:
What exactly was measured. Not the label, the operational definition. "Twelve schools within two kilometers" needs to say two kilometers of what, measured from what part of the site to what part of the school, counting which categories of institution, from which source layer. Most disagreements between two analyses of the same place dissolve once both operational definitions are written out, because they turn out to be measuring different things.
Where it came from. Which dataset, which vintage, which license, and which transformations were applied between the raw source and this number. Provenance is not bureaucracy. It is the only way to know whether two figures can be compared, whether a result can be reproduced, and whether the data can legally be used the way it is about to be used.
How wrong it could be. Positional uncertainty, classification uncertainty, coverage gaps, and staleness, stated separately because they behave differently. A claim with no error characterization is not a stronger claim than one with error bars, it is a claim whose error is unmeasured.
A system that can answer these three for every figure it emits is doing spatial intelligence. A system that emits figures it cannot trace is doing spatial arithmetic and hoping.
Why the agent consumer changes the stakes
Every one of these failure modes existed long before language models, and for decades the mitigation was a human in the loop who looked at a map. That mitigation is being removed at speed.
An agent calling a spatial API does not render the result. It receives numbers, in a structure that looks the same whether the numbers are good or nonsense, and it integrates them into a chain of reasoning that will produce a confident summary either way. It has no map to glance at, no memory of the city in question, and no instinct that says the answer feels wrong. If the response says the nearest hospital is 1.2 kilometers away and that figure came from a straight line across a river with no bridge, the agent will report a twelve minute walk, and the summary will read exactly like a correct one.
The design response is to move the disambiguation into the data rather than hoping for it in the reasoning. Return the metric used, not just the distance. Return coverage status alongside an empty result, so absence of data cannot be read as absence of the thing. Return classification confidence next to the class. Make the units of a density explicit in the field name if necessary. Refuse to emit a comparison between figures computed under different definitions, or mark it clearly as one.
This costs response size and some elegance. It is still cheaper than a fluent, well structured, entirely wrong answer that nobody in the chain is positioned to catch.
The failure that actually ships
The one to watch for is not a crash. It is a plausible number produced by a pipeline where every stage was individually reasonable.
Somebody joins a point layer to a boundary layer to get counts per district. The points are addresses geocoded from text, and the geocoder fell back to street centroids where it could not match precisely, which happened more often in districts with less standardized addressing. Those centroids cluster along main roads, which are sometimes the district boundaries themselves. The containment test assigns them to one side. Density is then computed against administrative areas that vary in size by two orders of magnitude. A ranking is produced, a decision follows, and the result is an artifact of geocoding fallback behavior interacting with boundary placement.
No stage was wrong. The join was correct, the containment test was correct, the arithmetic was correct. What was missing was any record of how each value was produced and how much it could be off by, carried far enough forward that the final number could be questioned. That record is what separates a spatial system from a system that happens to store coordinates.