Definition
Entity resolution compares records to decide whether they refer to the same real entity. In company research it also requires distinguishing a true duplicate from related but separate entities: a parent, subsidiary, branch, trading name, and acquired brand can share details without being interchangeable. The chosen unit of record determines which distinctions must be preserved.
Stable identifiers and corroborating evidence are stronger than name resemblance alone. Even identifiers require context: registrations can be jurisdiction-specific, and a shared website may serve an entire group. Keep matching evidence, relationship type, and unresolved contradictions visible. Combining records without preserving their source context can wrongly transfer capabilities, ownership assertions, or contact details from one business to another.
In origination work
Resolve records before counting market participants or launching repeated research. A branch can remain a useful location record while being linked to its operating company rather than counted as another acquisition target. Candidate matches should be reviewed before their evidence is merged. When a later acquisition changes the corporate structure, preserve the dated relationship history. The current map and the record of what was known previously answer different questions, and both can matter to an origination team.
The method in outline
Entity resolution is the process of determining which records refer to the same real-world entity, across one dataset or several. A typical pipeline normalises fields (case, punctuation, abbreviations, legal suffixes), generates candidate pairs by blocking on cheap keys such as postal code or the first letters of a name, scores each pair on several comparison fields, and classifies the pairs as match, non-match or needing review. Clusters of matched records are then merged into one entity.
The field has a long history. A 1959 paper in the journal Science on linking vital records is an early computerised example, and a 1969 statistical paper formalised a model in which each field's agreement or disagreement contributes weight towards a match decision. Government statistical agencies have published overviews of the methods and error rates, and modern open-source packages implement the same ideas at scale.
Errors and their costs
There are two errors. A false merge combines two different businesses, and its consequence is that facts about one are attributed to the other: a pending complaint, a licence status or a size figure lands on the wrong company. A false split leaves one business as two records, causing duplicate work and an inflated count of targets. Which error is costlier depends on the use; for outreach, a false merge can lead to contacting the wrong person, and for market counts, false splits overstate fragmentation.
Thresholds therefore should be chosen deliberately and tested on a labelled sample. Review a random sample of auto-matched pairs and a sample of auto-rejected pairs near the threshold, count the errors, and report them. A matching system that cannot state its error rate on a sample is making claims it cannot support.
Practices for research teams
Prefer authority-issued identifiers when they exist, and keep a crosswalk between identifier systems, for example from a state registration number to a federal contractor identifier to a legal entity identifier. Where identifiers are missing, use combinations of address, officers and dates, and mark the match as inferred.
Keep the audit trail: which rule fired, which fields agreed, who reviewed. Where an automated agent suggests a match, have a second pass check the evidence with the original sources, not just the agent's summary. The same reasoning that supports a verbatim quotation for every extracted fact applies to matches: show the evidence, not just the conclusion.
Related terms
- Identity resolution — Establishing which real business a candidate record, name, domain, or source actually refers to.
- Deduplication — Identifying and consolidating repeated records while preserving evidence, history, and genuinely distinct entities.
- Market map — A structured view of a defined market's participants, relationships, capabilities, and research coverage.
Further reading
- Resolve companies, brands, and locations
- Identity resolution
- Deduplication
- Provenance
- GLEIF Legal Entity Identifier records
- Article: identifying the same company across states
- Article: SAM.gov and USAspending contractor crosswalk
Sources and further reading
- A Theory for Record Linkage, Journal of the American Statistical Association (1969). The foundational probabilistic model for deciding whether two records describe the same entity.
- Automatic Linkage of Vital Records, Science (1959). The earliest computerised record-linkage study, still the origin of the name-matching problem.
- Overview of Record Linkage and Current Research Directions, U.S. Census Bureau (2006). A government statistician's survey of matching methods and error rates.
- Splink: probabilistic record linkage, UK Ministry of Justice Analytical Services. Open-source documentation that makes the Fellegi-Sunter approach concrete.
- Legal Entity Identifier (LEI) search, Global Legal Entity Identifier Foundation. Open, free entity identifiers with parent relationships. The nearest thing to a global primary key for companies.
To apply these distinctions to your own acquisition criteria, discuss your research brief or inspect the demo platform.