Data linkage rates: how to measure whether business records really connect

A link should let someone verify the relationship

An email mentioning a project name is a clue. A record with an explicit project identifier is a different kind of evidence. Treating both as equally reliable can make a connectivity score look stronger than the underlying dataset.

Start by naming the relationship you need: a ticket to its email thread, a file to its project, a comment to a file version, or a meeting attendee to a consistent employee identity. A useful assessment distinguishes explicit identifiers, reviewed matches, and unresolved candidates.

That distinction is central to our AI-Ready Data framework and to a useful connection manifest.

Define the denominator before calculating the percentage

For a record-level cross-system linkage rate, one possible definition is: eligible records with at least one validated relationship to another system, divided by all eligible records. This is a proposed assessment definition, not a universal industry standard.

Do not silently exclude records because they failed to match. Explain legitimate exclusions—such as out-of-scope dates—and retain the original counts. An empty denominator is “not measured” or “not applicable,” not an automatic 100%.

Separate connectivity from correctness

It is possible to connect every record incorrectly. A linkage report should therefore distinguish coverage from the correctness of reviewed matches.

Sample-based correctness checks should state how links were chosen for review. Reviewing only obvious matches will not tell you much about ambiguous ones.

Show the joins a buyer can reproduce

Publish the permitted join keys, system pairs, transformation notes, and exceptions in the assessment package. Include a few reviewed examples with record references. Keep any protected identity lookup separate from the buyer-facing files.

W3C’s PROV framework describes provenance in terms of entities, activities, and the people or systems involved in producing data. Our application is to record how a claimed relationship was established, rather than presenting an unexplained edge in a graph.

Linkage also needs a time dimension. A project identifier reused after migration may not mean the same thing in every export. Document namespaces, version changes, and the period in which a join rule applies.

Check the links again after de-identification

Cross-system linkage in an original export does not prove the same connections exist in the shareable version. Compare validated, in-scope original links with the transformed records using an authorized baseline.

Record losses caused by removed identifiers, missing attachments, inconsistent replacements, or excluded records. Our de-ID and identity guide explains why a stable pseudonym helps connectivity but does not itself prove anonymity.

Turn the calculation into a useful next step

Choose a workflow whose connections matter to the intended buyer, then ask what remains untraceable. A single missing approval can matter more than hundreds of linked routine notifications. Use the sample preparation guide to bound that review.

Discuss quality ranking with Clear Harness, Ohio Training Data’s partner. If you are also exploring whether your company has a licensing opportunity, start the separate Troveo assessment (referral link; we may be compensated).