Euclidean Feed Explained: Distance Math for Site Selection and Competitive Analytics

Pairwise-distance feeds quietly power retail, CRE, and CPG analytics: site selection, cannibalization, competitive catchment. Math is elementary; use-cases are decisive. This reference covers Euclidean vs drive-time, nearest-K structure, procurement, and joins to foot traffic. Pair POI quality in depth and retail site selection stack.

What a Euclidean Feed Actually Is

Precomputed straight-line distances between POIs: full N² impractical at 10M POIs; production feeds store nearest-K per source (K=50-500). Euclidean Feed ships meters, NAICS filters, IDs joining POI & Geofencing. Haversine on centroids: accurate within meters at retail scale.

Amortize compute offline → interactive portfolio queries (hundreds of candidates × thousands of competitors) vs 20-minute batch recompute per run.

Euclidean vs Drive-Time: When Each Is Right

Euclidean (direction-agnostic, network-independent, constant-time) site screens, competitive density, CPG-retailer mapping. Drive-time (road network, traffic) suits exact trade areas and delivery feasibility. Full pairwise drive-time at 10M POIs infeasible; nearest-50 drive-time expensive on subsets only. For portfolio questions, Euclidean and drive-time correlate >0.95 at retail distances, so cost-benefit favors Euclidean for strategic work per Google Distance Matrix cost reality.

Site Selection and Cannibalization

Rank own-brand stores nearest candidate site; overlay Global Mobility for cannibalization share. Rank competitors by distance + NAICS; overlay category share. Trade Area Explorer demonstrates lightweight pattern. Pre-register K and distance bands in pilot charter.

Competitive Catchment and Trade-Area Overlap

REITs and chains find nearest-K competitor brands, then O/D from mobility for shared draw. Euclidean indexes neighbors; mobility measures customers. NAIOP research cites catchment overlap as high-leverage leasing input.

Euclidean Feed Procurement Diagnostics

Checklist:

1. Configurable K per source. K=50 minimum retail; K=200+ dense urban.
2. Distances in meters, documented WGS84 haversine precision.
3. POI IDs join production catalog with NAICS + brand metadata.
4. Refresh cadence tracks POI open/close: lag poisons neighbor lists.
5. Drive-time companion for tactical lane or buyer-owned routing API.

Underperformance is usually join failure or stale POI: diagnosable with join-audit script in minutes. Score vendor with RFP matrix.

Barriers Are Where Straight-Line Distance Lies Most

Euclidean distance fails in a specific and predictable way: it ignores anything that stops a person moving in a straight line. Rivers, motorways, rail corridors, canyons, and coastlines all create pairs of locations that are close on paper and far in practice. A competitor eight hundred meters away across an uncrossable highway is not competing for the same lunchtime trade, and a straight-line model will insist otherwise every time it is run.

The useful discipline is not to abandon the straight-line feed but to know where it is unreliable. Compute the ratio of network distance to straight-line distance for a sample of pairs in each market you operate in. Where the ratio sits near one, the fast maths is a fine approximation and can be used freely. Where it runs high, the market has structural barriers and any straight-line result there needs checking before it drives a decision. That ratio is worth storing per market, because it turns a general caveat into a specific list of places to be careful.

Real Catchments Are Asymmetric, Distance Matrices Are Not

A distance matrix is symmetric by construction: the distance from A to B equals the distance from B to A. Human movement is not. Commuting flows run one direction in the morning and the other in the evening, one-way systems and turn restrictions make an approach easy and a return awkward, and a location on the homeward side of a road attracts trade that the identical site opposite does not. None of this appears in a distance feed.

Treat the matrix as a candidate generator rather than an answer. It is excellent at reducing millions of possible pairs to the few hundred worth examining, which is the expensive part of the problem and where its speed genuinely earns its place. The final read on any pair that matters commercially should come from observed behavior, ideally visit data showing where customers actually come from, because a trade area drawn from movement will regularly look nothing like the circle drawn from distance.

A Precomputed Matrix Ages Against a Moving Map

Precomputation is what makes distance feeds usable at scale, and it introduces the standard cost of any cache: the underlying POI file keeps changing while the matrix does not. Locations open, close, and move, and a matrix computed against last quarter’s file will confidently return neighbors that no longer trade and omit the competitor that opened last month. The error is quiet, because a stale distance looks exactly like a fresh one.

Tie the recompute cadence to the churn rate of the POI universe rather than to a convenient calendar, and check faster in categories that turn over quickly. Stamp every delivery with the POI file version it was computed from, and refuse to mix results computed from different versions in one analysis, since that is how a comparison acquires a difference that belongs to the input rather than the market. Where the matrix drives an automated decision, add a freshness assertion at the point of use so a stale input fails loudly instead of quietly.

Where Nearest-K Quietly Returns the Wrong Answer

Any count answered from a nearest-K list is censored at the K-th neighbor. Ask how many competing locations sit within a given radius and the honest answer is a floor whenever the K-th neighbor is closer than that radius, because the list ran out before the band did. The diagnostic is one comparison per source POI (K-th neighbor distance against your query radius), and it belongs in the pipeline as a truncation flag rather than in a one-time QA notebook. Never pool flagged and unflagged rows into a single density statistic. Truncation concentrates in dense markets, so pooling biases precisely the geographies the analysis exists to rank.

Co-tenancy eats the K budget before a real competitor appears. Malls, airports, strip centers, and office towers put many POIs on one parcel with near-identical centroids, so the top of the neighbor list is filled with tenants of the same building at trivial distances. Inside a large enclosed center the entire returned list can be internal to the building. Two fixes work together. Request K per category or per brand rather than K overall, and collapse to one record per parcel or building before ranking, keeping the co-tenant roster as a separate attribute instead of as neighbors.

Nearest-K is also directional, which breaks pair logic. Location B can appear in A's list while A is absent from B's, because B sits in a denser neighborhood where the cutoff bites sooner. Cannibalization and catchment-overlap models that assume a symmetric pair table drop those pairs without complaint. Union both directions before any pair analysis, and track the reciprocity rate as a health metric. A falling rate is evidence that K is too small for the density you are working in, not evidence that the market thinned out.