POI data schema
The POI data schema defines the record structure for the two location-data classes the location intelligence substrate consumes: retail chain locations ingested from OpenStreetMap, and institutional anchor locations ingested from the Overture Maps Foundation. Both are normalised into a unified flat-file JSONL schema before cluster analysis runs.
Why it matters: no proprietary data purchase is required. Every record derives from a publicly licensed source and is version-controllable in the same ledger as the rest of the platform, so a customer's location dataset is auditable end to end.
Record types
Service-business records represent individual retail chain locations — hardware stores, warehouse clubs, hypermarkets, food anchors. Each is identified by a chain_id key linking it to a chain configuration file, and by a brand_wikidata field holding the Wikidata QID for the brand.
Service-places records represent institutional anchors — hospitals, universities, airports — ingested from Overture Maps via the taxonomy.primary category field. These use a category_id key (hospital, university, airport) in place of chain_id.
Core fields
Both record classes share the following:
| Field | Type | Notes |
|---|---|---|
location_name |
string | Display name; COALESCE of brand name and category fallback |
brand_wikidata |
string or null | Wikidata QID (e.g. Q13556979); null for civic places with no brand identity |
street_address |
string or null | Freeform address from OSM addr:housenumber + addr:street, or Overture addresses |
city |
string or null | Locality from addr:city, addr:town, or addr:municipality |
region |
string or null | Province, state, or NUTS-3 region |
iso_country_code |
string | ISO 3166-1 alpha-2 |
latitude |
float | WGS 84, 7 decimal places |
longitude |
float | WGS 84, 7 decimal places |
naics_code |
string | NAICS industry classification |
top_category |
string | NAICS top-level category description |
sub_category |
string | NAICS sub-category description |
source |
string | osm or overture |
confidence |
float | Confidence score (OSM: fixed 0.85; Overture: from dataset) |
Why it matters: one shared field set across both classes means downstream analysis code does not branch on record type for anything except the chain-versus-category key.
Chain identification and the Wikidata QID
The brand_wikidata field holds the Wikidata QID for the retail brand. QIDs are persistent, language-independent, and community-maintained, which makes them the preferred chain identifier across both commercial and open POI datasets — they are brand-level rather than name-level, so two stores spelled differently but sharing a QID belong to the same chain.
The OpenStreetMap community tags retail locations with brand:wikidata=<QID>, and the ingest uses this tag as its primary query filter; a location tagged with the correct QID is captured regardless of local name spelling. Overture Maps exposes the same identity via brand.wikidata in its Places schema, extracted at ingest for service-places records.
Why it matters: chain identity survives translation, rebranding of a local storefront name, and inconsistent data entry — which is what makes cross-border comparison possible at all.
Overture taxonomy schema
Overture Maps deprecated the categories struct in November 2025 and removed it in the June 2026 release. The replacement taxonomy struct exposes taxonomy.primary (equivalent to the old categories.primary) and taxonomy.alternate, an array of secondary category associations with optional attribute structs.
Category identifiers are unchanged across the migration: a query that previously read categories.primary = 'hospital' becomes taxonomy.primary = 'hospital' with no change to the filter values.
Spatial deduplication
OSM data for large-format retailers sometimes includes both a node and a way for the same physical location — the building footprint as a way, the entrance as a node. The ingest deduplicates by rounding each location's coordinates to four decimal places (roughly 11 metres) and treating records that round to the same pair as the same building, retaining the record with the most complete address fields.
Records sharing coordinates at this resolution but carrying different chain_id values under the same brand_wikidata QID are treated as sub-format or co-branded stores — a fuel station sharing the parent retailer's QID, for instance — and are candidates for the parent-child model below.
Why it matters: without this pass, a single store can appear two or three times and inflate every count computed downstream of it.
Parent-child sub-location model
Large-format retailers frequently operate ancillary services at the same address: pharmacies, fuel stations, optical centres, garden centres. In raw OSM data these appear as separate POI elements, each with a distinct name and sometimes a distinct chain_id.
A configuration-driven parent mapping resolves this: each sub-entity chain_id known to be an ancillary service maps to its canonical parent chain. Sub-entities are excluded from cluster scoring and surfaced only on the parent's info card; the map shows one marker per parent location. This follows the industry-standard parent-child POI pattern, in which the parent record holds the canonical address and coordinates and sub-entities share that anchor while carrying their own service classification.
The Placekey standard — a globally unique location identifier with a What@Where structure — expresses the same relationship via a shared Where component: two POIs at one address share the geocell suffix while their brand-hash prefix differs. A Placekey-based spatial-matching approach is a planned future mechanism rather than the current one; the schema retains a placekey field for it, not yet populated during ingest.
Why it matters: counting a hypermarket's fuel station as an independent anchor would inflate a cluster's apparent brand diversity, which is the signal the tier system is built on.
Address completeness
Address coverage varies by country. OSM coverage of addr:housenumber and addr:street is strong in Western Europe and Canada, moderate in the United States, and sparse in some Nordic and Southern European markets. A planned enhancement will spatial-join POI records against the Overture Addresses theme within a 15-metre radius to back-fill missing street-level addresses; that theme provides structured records for over two billion global addresses derived from authoritative national registries.
Data update cadence
Service-business records are re-ingested per chain on demand — typically when a new chain is added to the configuration, or when quarterly coverage audits flag anomalies. Service-places records are re-ingested against new Overture quarterly releases; the ingest script's Overture object-store path must be updated to reference each new release.
See also
- Location intelligence substrate — the flat-file GIS architecture and storage layer
- Business clustering — the clustering service that consumes these records
- Places filtering — civic-anchor filtering downstream of ingest
- Regional name resolution architecture — how cluster coordinates become regional names
- Location intelligence platform — the application surface these records are served through
OpenStreetMap data © OpenStreetMap contributors, licensed under ODbL. Overture Maps Foundation data under CDLA Permissive 2.0.
Cite this record: /wiki/poi-data-schema — revision a2fa91b3, last updated 26 August 2026.