Methodology

Sources

Every row traces back to one of three source kinds: brands' own store locators (extracted directly, not scraped from a third-party directory), government registries (FDIC-insured bank branches, CMS-certified hospitals, NCES schools, FAA airports, and others — 15sources total), and open geospatial data (Foursquare Open Source Places and Overture Maps) used to fill category and geographic coverage the first two don't reach. No device location data, no carrier or app SDK panels, no location-history modeling. If a location isn't in a brand's own locator, a government registry, or an open dataset, it isn't in the catalog.

Extraction cadence

Most brand store-locators are refreshed weekly — 194brands this week — so an individual brand's data can be more current than the catalog's own publish cycle. The full catalog is reassembled, conflated, and published once a month (the current release: 2026-07-02); that monthly release is what every product page, price, and sample on this site reads from, so the whole catalog is point-in-time consistent even though individual sources refresh on their own schedules.

Conflation & dedup

The same physical location often shows up in more than one source — a brand's own locator and an open dataset, or two overlapping government registries. Records are matched on name, address, and geographic proximity and conflated into a single canonical business location when they represent the same place, with the surviving record's fields chosen by source priority (a brand's own locator outranks an open-data supplement for that brand's own locations). Every canonical business location carries a stable, content-derived id and, where one exists, the original source's own identifier, so a row can always be traced back to what produced it.

Confidence scoring

Field-level confidence is informed by source agreement: when a name, category, or address is corroborated across independent sources, confidence is high; single-source fields or fields with conflicting values across sources are scored lower and are more likely to be the ones showing up in a brand's field fill-rate table on its product page. Confidence scoring drives internal QA sampling and category assignment — it doesn't hide or drop rows; the per-field fill rates on every brand page and in the data dictionary are the customer-facing view of the same underlying signal.

QA gates

Every monthly release passes roughly 1,800 automated tests and 28 release checks before it publishes — schema contract tests, per-tier row-count sanity checks, address- and category-field validity, id-format compliance, and duplicate detection among them. A release that fails a gate doesn't ship; the storefront always reflects the last release that passed all of them.

Provenance tiers

Every business location carries one of four provenance tiers, shown as a tier chip everywhere the row appears — on sample tables, product pages, and in the schema itself. The tier tells you exactly where a row came from and how it was verified, not just that it's "in the dataset."

Brand-extractedrows come directly from a brand's own store locator. Gov-registryrows come from a government source and carry that source's own identifier. FSQ-derived rows fill in branded, unbranded, and sub-location coverage from Foursquare Open Source Places. Overturerows are the open-data supplement used only where the first three tiers don't reach. The Ground Truth US product is the first two tiers only; the Full US master is all four.

Known limitations

Brand coverage grows every month but isn't exhaustive yet — smaller and regional chains are still being onboarded. Independent, non-chain businesses are covered only where a government registry or open dataset includes them; the catalog is strongest on chains and registered institutions. There is a detection lag between a real-world opening or closing and its appearance in the change feed, bounded by that brand's extraction cadence above. FSQ-derived and Overture rows carry lighter per-field verification than the brand-extracted and government-registry tiers — this is exactly what the tier system above is for: it tells you which guarantee applies to which row, rather than asserting one blanket confidence level over the whole file.