## TL;DR
Company names are not unique; domains are close to it. Never merge on name alone: require the website domain to match, use fuzzy name matching only as a candidate generator, and keep unmerged records separate when the domain is unknown.

```text
enrichment agent matched companies by name only and merged two different acme corps
```

## Use this when
- Two different companies got merged into one record.
- Enrichment shows the wrong website or employees for a company.
- A 'merge duplicates' job keeps producing false merges.

## Not for this skill when
- Contact deduping, which should key on email.
- Same company with name variants, which is a normalization problem.
- Intentional merges of known subsidiaries.

## Steps
1. Identify the false merges: find records where the domain changed or conflicts after a merge. Verify: you have concrete examples of wrong merges.
2. Split the false merges back into separate records. Verify: each record carries its own correct domain.
3. Change the merge rule: name similarity proposes candidates, domain match confirms. Verify: the rule is documented and the code enforces it.
4. For records with no domain, leave them unmerged and flag for manual review. Verify: the review queue exists and has a decision SLA.
5. Re-run enrichment on the split records and spot-check the results. Verify: a sample of re-enriched records shows correct companies.

## Variant phrasings
### company dedupe name only false merge
Common phrasing of the problem.
### entity resolution companies domain match
The identity-angle search.
### enrichment merged wrong companies
The symptom phrasing.

Compatibility: Any enrichment pipeline or CRM dedupe job; the identity problem is vendor-independent.

## Why it happens
Thousands of companies share names like 'Acme Corp' across industries and countries. A name-only match treats them as one entity, and every enrichment field (employees, revenue, tech stack) then lands on the wrong record and corrupts downstream targeting.

## Edge cases / pitfalls
- Domains can change with rebrands; keep a history of past domains per record.
- Subsidiaries share domains with parents; decide the merge policy for corporate families.
- Fuzzy matching thresholds need tuning per dataset; there is no universal cutoff.
- Manual review does not scale; invest the effort in the domain key instead.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_nJ9ZAj9-m4Lm3xrQM227Hw