Matching company records across CRMs, directories, invoices, and registries requires more than comparing names. This checklist shows how to normalize legal names, identify stronger evidence such as domains and registration identifiers, separate subsidiaries from duplicates, and route uncertain matches for review without hiding important differences.
Overview
Company name matching is an entity resolution problem: the goal is to decide whether two records refer to the same legal or operational organization. A single organization may appear as “Northstar Technologies, Inc.,” “Northstar Tech,” “Northstar,” or under a trading name. Conversely, two records can share nearly identical names while representing different subsidiaries, branches, or unrelated companies.
Fuzzy matching and text similarity are useful for generating candidates, but they should not make the final decision by name alone. A reliable workflow separates three tasks:
- Normalize: create comparable forms of names, addresses, domains, and identifiers while preserving the original values.
- Compare: combine exact evidence, approximate string matching, and contextual fields to score candidate pairs.
- Decide: assign a match, non-match, related-entity, or manual-review outcome with an explanation.
Keep raw source values, normalized values, match features, decision thresholds, and reviewer notes. This audit trail makes later corrections safer and helps you measure whether a new source or parser changes results.
Checklist by scenario
1. Records differ only by legal suffix or formatting
- Trim whitespace, standardize case, and normalize punctuation and repeated separators.
- Extract a jurisdiction-aware legal suffix such as a local equivalent of corporation, limited, or limited liability company.
- Store the suffix separately rather than deleting it permanently.
- Compare the remaining name with token, character n-gram, or edit-distance methods.
- Check another field before automatically merging, such as address, domain, registration number, or phone number.
Do not use one global suffix list for every country. A word that looks like a corporate designator in one jurisdiction may be part of the actual name in another. Treat suffix removal as a parsing step with a confidence or provenance flag, not as an irreversible cleaning rule. A field-safe normalization pipeline is useful when names contain accents, punctuation, or inconsistent formatting.
2. Records use abbreviations, trading names, or brand names
- Capture separate fields for legal name, registered name, trading name, brand, and doing-business-as name when available.
- Expand only approved, unambiguous abbreviations; do not expand every short token heuristically.
- Compare the name against known aliases linked to a verified entity.
- Use domain ownership, address, registration data, or source-provided entity IDs to support the alias relationship.
- Keep the alias relationship distinct from the legal-name match.
A brand match can identify the same commercial group without proving that two records are the same legal entity. That distinction matters for invoices, contracts, compliance workflows, tax records, and account ownership.
3. Records may describe subsidiaries, parents, or branches
- Look for registration identifiers, country or state, legal address, and incorporation details.
- Compare domains cautiously: a shared domain may indicate a group relationship rather than identical legal ownership.
- Model parent, subsidiary, branch, franchise, and acquired-company relationships separately.
- Use “related entity” when evidence supports a corporate relationship but not an exact-entity match.
- Never collapse a subsidiary into its parent solely because the names share a distinctive token.
For example, “Oak Systems Holdings Ltd” and “Oak Systems UK Ltd” may be related but should not automatically receive the same entity ID. A deduplication system should preserve the distinction unless the business rule explicitly defines a group-level account.
4. Records come from multiple countries or languages
- Store the source language, jurisdiction, and script where available.
- Normalize accents and transliterations in a comparison field while retaining the original spelling.
- Apply language-aware tokenization and legal-suffix rules.
- Use phonetic matching only as candidate generation or supporting evidence, because pronunciation rules vary by language.
- Check local identifiers and addresses before accepting a high name-similarity score.
Transliteration can create collisions, and translation can change the meaning of a name. Review multilingual edge cases with examples from each source rather than assuming that one normalization strategy works everywhere. See the guide to multilingual entity matching for a broader candidate-generation approach.
Evidence hierarchy for match decisions
| Evidence | Typical strength | How to use it |
|---|---|---|
| Verified registration or company identifier | Very strong | Prefer exact comparison within the correct jurisdiction and entity type. |
| Source-confirmed canonical entity ID | Very strong | Use as a stable key, while checking that the source's scope and date are appropriate. |
| Exact domain or verified website relationship | Strong but contextual | Support a match; account for group domains, regional sites, and outdated domains. |
| Address plus normalized company name | Moderate to strong | Compare country, locality, and address components; shared offices may reduce certainty. |
| Phone, email domain, or tax-related field | Moderate | Use as supporting evidence, not as a universal identifier. |
| Name similarity alone | Weak to moderate | Generate candidates and rank them, but require additional evidence for automatic merges. |
What to double-check
Before assigning a final match, inspect the features that produced the decision:
- Name components: Are the distinctive tokens preserved after suffix removal? Did a parser remove a meaningful word?
- Identifiers: Are the identifiers from the same jurisdiction and referring to the same kind of entity? Is a leading zero or punctuation being lost?
- Domains: Is the domain current, regional, shared by a parent group, or merely a service provider's domain?
- Addresses: Is this a registered office, headquarters, branch, mailbox, or shared workspace?
- Time: Were the records active at the same time? A former name or acquired business may require a historical relationship rather than a duplicate label.
- Source quality: Was the value entered by a user, copied from an invoice, or supplied by a registry? Preserve provenance and avoid treating all sources as equally authoritative.
Use a labeled sample to tune thresholds. A practical model can combine name similarity, token overlap, domain agreement, address similarity, and identifier agreement, then define separate thresholds for automatic match, automatic non-match, and review. Thresholds should be tested against false merges and missed matches, not selected from a similarity score in isolation. A gold-standard entity-resolution dataset makes this evaluation repeatable.
Manual review form
For ambiguous organizations, record:
- Source record IDs and original names
- Candidate canonical entity and alternative candidates
- Jurisdiction, registration identifiers, domains, addresses, and effective dates
- Normalized name components and similarity features
- Decision: exact match, probable match, non-match, related entity, or unresolved
- Reason for the decision, reviewer, date, and evidence links or source references
Common mistakes
- Deleting legal suffixes without retaining them: This removes useful evidence and makes parser errors difficult to diagnose.
- Equating a shared brand with a shared legal entity: Brands, parent groups, and subsidiaries require relationship modeling.
- Trusting name similarity too much: Common tokens produce convincing false positives, especially in the same industry or city.
- Using domains as definitive proof: One domain may serve several entities, while a legitimate company may have multiple domains.
- Over-normalizing: Removing numbers, locations, or meaningful words can turn distinct entities into the same string.
- Auto-merging before measuring: Start with candidate generation and review, then automate only decisions supported by labeled examples.
- Ignoring reversibility: Keep original records, entity history, and merge events so a mistaken decision can be rolled back. The workflow in safe record deduplication applies the same principle.
When to revisit
Revisit company matching before each new source import, supplier or lead-enrichment refresh, seasonal planning cycle, organizational restructure, and audit of unresolved records. Also review it when workflows, registries, matching tools, or normalization rules change.
At each review, sample confirmed matches, non-matches, and manual decisions. Check whether a new jurisdiction introduces legal suffixes your parser does not recognize, whether a provider changes identifier formats, and whether subsidiaries or renamed organizations are being incorrectly merged. Recalculate precision and review workload on the labeled sample before raising automation thresholds.
End every refresh with three actions: document changed rules, queue uncertain records for review, and preserve a versioned decision log. If your system needs faster candidate generation, combine field-aware normalization with indexed approximate matching rather than applying one broad fuzzy search over every company name. This keeps entity resolution explainable while leaving room for careful human judgment where corporate structures are genuinely ambiguous.