The most complete record of European public buying in private hands.
Collected source by source since March 2026 from 72 national portals and TED, back to 2002 where the archives reach. Every number below is measured on the corpus itself — methods are published with every figure we release — the full method lives here.
Why this is hard to rebuild
The corpus is not a scrape; it is 72 independently fought integrations, each with its own failure modes conquered one by one:
- Portals that cap results at 2,000 per query — solved by partitioning day×category, discovered only after the naive corpus silently lost 60% of records.
- Portals that block datacenter IPs — routed through residential egress.
- Session-stateful pagination, ISO-8859-1 encodings, APIs whose sort order lies, feeds that re-publish on every status change.
- Archives reaching to 2002 — parts of which are no longer served by the sources themselves. That history cannot be re-collected at any price.
Replication is not a budget question alone; it is 18 months of sequential discovery, because each portal teaches you its traps only in production.
What it runs on
The entire collection, normalization, translation, verification and publication pipeline runs autonomously — 50+ scheduled services with failure capture, self-rebuilding read models, a machine-written and machine-fact-checked signals desk, and a public live feed of its own operation. Infrastructure cost: under €100 per month, all-in. Marginal cost of the next country: an integration, not a team.
Structure, not just storage
Buyers normalized across 18.6M notices; awards linked to registry-identified companies (NIF, SIREN, CUI, EIK, Y-tunnus…); CPV taxonomy throughout; selection criteria structured for 87% of open above-threshold tenders; English title layer; per-buyer activity cycles ready for prediction.
Live coverage by country — darker = more open tenders. 18 countries read below the EU threshold at the national source.