Method

How SupplierProof builds a dossier

The pipeline is an agent with a budget, not an unbounded crawler. Free official APIs (EDGAR, GLEIF, Wikidata) always run when the geography fits. Sirene and ABN Lookup run when you add their keys. ImportYeti is US Customs ocean freight since 2015 — including overseas shippers — not EU/AU/air/truck. A miss is not proof the company is fake. Reddit (10 recent posts) and Trustpilot (company card + 10 recent reviews) run via Apify when a token is set. Trustpilot can fill a missing legal name; if it disagrees with the website, both names stay on the dossier. Tavily Search still supplies extra OSINT URLs on every claim.

1. 01

   Cheap facts first

   RDAP/WHOIS, DNS, TLS, IP geo, and Wayback run in parallel with no model. The crawl then reads the homepage, JSON-LD Organization, and real nav/footer links — not a fixed /privacy list.

2. 02

   Budgeted identity agent

   HTML is fetched directly (never by the model, and PDFs are skipped). Nav/scripts are stripped and the footer is placed first. JSON-LD and obvious contacts are cheap hints only.

3. 03

   Extract, then check coverage

   The model reads cleaned pages and decides brand vs legal entity (copyright, division/subsidiary, about copy). Regex is only a fallback if the model is down. IDs are never invented.

4. 04

   Official corroboration, routed

   UK Companies House, US SEC EDGAR, France INSEE Sirene, Australia ABN Lookup, GLEIF LEI, and Wikidata (official website → LEI). Each fires only on a matching identifier, TLD, or address. OpenCorporates is the last fallback. ImportYeti remains US ocean BOLs only.

5. 05

   Score, then intent guidance

   The numeric trust score is deterministic. The model only writes the summary and a three-column brief for your chip (Sourcing / Site buy / B2B): do now, ask them, watch for. Every chip and URL in the dossier is a source.