Unifying 400,000 records across sources with no shared key

Combo had 124,000 restaurant listings from Google Maps and 277,000 from French business registries. No shared identifier between them. A Google listing for "Chez Marcel" might match "SAS Groupe Restauration Marcel et Fils" in the registry—or it might not. Without linking the two, neither dataset was useful for prospecting.

• No shared identifier. Restaurant names almost never match legal entity names.
• Ownership structures hide behind holding companies and brands. One owner, five legal entities, three trade names.
• Naive matching = 34 billion pairs. Not happening.

Blocking by postal code + 500m radius cut 34 billion pairs down to 410,000 (99.999% reduction).

Scoring: 70% geographic distance, 20% address similarity, 10% name similarity (trigram Jaccard). Bidirectional best-match ranking to avoid one-to-many collisions.

Vector search (text-embedding-005, BigQuery IVF index) for semantic queries across the unified data. Daily refresh with hash-based change detection.

Why it worked

• Geography is the strongest signal when names diverge. 70% weight on distance.
• Conservative thresholds: average accepted score 0.83, average match distance 50m. Ambiguous matches flagged, not forced.
• Registry provider choice mattered. One had 99% coordinate coverage, another 40%. That gap made geographic matching viable.
• Incremental updates with change detection. No full rebuilds, daily refresh.