Metadata quality is the degree to which the records describing your data are complete, accurate, consistent, conformant, current, traceable to a source, and confirmed by a named person. Most programs measure the first of those and stop. The metadata quality dashboard shows the share of described tables climbing each quarter, and nobody asks what the descriptions say. Since AI enrichment arrived, the answer is often a model's guess. Coverage went up. Quality did not move, and the dashboard cannot tell the difference.
This piece sets out what to measure, how to make each measure a check a machine runs, and what to do when one fails. The method rests on one fact the published frameworks leave out. Metadata held as structured data can be validated with SHACL, the W3C standard that validates data. A rule about metadata becomes a test, and the result becomes metadata too.
Metadata quality is measured by rules, scored by consequence, and raised by approval. The rule is a SHACL shape. The score covers the terms that decide answers. The fix is a named expert confirming what a model proposed.
Data quality asks whether the values are right. Metadata quality asks whether the description of those values is right, and whether anyone would stand behind it. The two fail in different places.
| Question | Data quality | Metadata quality |
|---|---|---|
| What is tested | The values in a column or a record | The record that describes the column: its definition, owner, mapping, scope, and dates |
| Example rule | Every order carries a customer ID that exists in the customer table | Every governed term carries exactly one approved definition and a steward who is a person |
| Who notices a failure | A report is wrong, a pipeline breaks, a customer complains | Nobody, for a long time. The description reads well and the data loads |
| What an agent does with a failure | Returns a wrong number | Returns a plausible answer built on a meaning nobody agreed |
The last row is why metadata quality became urgent. A person reading a stale description notices it. An agent reads it as fact. In our post What Is Metadata Management? we state that an inferred meaning is directionally right until being wrong carries a consequence. Metadata quality assessment is how you find the inferred, stale, and unowned records before an agent resolves against them. For a search result, a mostly right description is fine. For an agent approving a credit limit, 85% accurate isn't good enough.
The research literature settled on a short list. A 2016 review by McMahon and Denaxas in the AMIA Joint Summits proceedings found nine dimensions in prior work. They are accessibility, accuracy, completeness, conformance, interoperability, consistency, frequency, timeliness, and meta-metadata. The vendor guides narrow the list. The lakeFS guide, updated July 2026, lists completeness, accuracy, consistency, validity, uniqueness, and timeliness. The Fast.io guide, updated April 2026, lists completeness, accuracy, consistency, timeliness, and conformity. One dimension is missing from every list, and it is the one an AI system depends on most.
| Dimension | Question it answers | A check you can run | What a filled-field count misses |
|---|---|---|---|
| Completeness | Does every governed term carry the fields a consumer needs? | Definition, steward, scope, mapping, and review date all present | A description a model generated counts as present |
| Accuracy | Does the description match the data? | Profile the column against the rule in the definition; sample and confirm | Everything. A filled field counts, right or wrong |
| Consistency | Does one concept have one meaning everywhere? | Count definitions per term across catalogs, glossaries, and dashboards | Forty tidy descriptions of forty columns can hold five meanings |
| Conformance | Do values and identifiers follow the agreed form? | Date formats, code lists, identifier patterns, required vocabularies | A free-text date passes a presence check |
| Timeliness | Is the record current? | Review date in the future; no schema change since the last confirmation | A last-edited timestamp records who touched the text |
| Provenance | Where did this description come from? | Source of each field, crawler, model, or person, with the date | Model-written and steward-written text look identical |
| Attestation | Did an accountable person confirm this meaning, and until when? | Approver is a person, approval is dated, review is scheduled, expired records are flagged | All of it. No coverage metric has a column for it |
Attestation, or SME approval, is the dimension that separates a described record from a trusted one. Our post Ground Truth: SME-Approved vs Inferred Metadata covers why that distinction decides whether an agent's answer can be defended. If attestation is unmeasured, an enrichment agent can raise every other number and approve nothing.
A metadata quality score is a set of rules, run over a scope, with the pass rate reported. Three decisions matter more than the arithmetic.
There is a public example at scale. The European data portal runs a Metadata Quality Assessment over every dataset description it harvests. It scores five dimensions out of 405 points and assigns one of four ratings. The Excellent band starts at 351, as reported in an August 2026 arXiv study of the Nordic catalogues. The portal publishes a DCAT-AP SHACL validator alongside the dashboard, per its provider manual. The rules are SHACL shapes.
SHACL, the W3C Shapes Constraint Language, is usually introduced as a way to validate data. Our post What Is SHACL? covers the standard. Metadata held in RDF is data, so a shape applies to a metadata record the same way it applies to a customer record. A rule about what a governed term must carry is written once, runs on every record, and returns a report that names each failure.
| Quality rule | SHACL constraint | What the check catches |
|---|---|---|
| Every governed term has exactly one definition | sh:minCount 1 and sh:maxCount 1 on the definition property | Terms with no definition, and terms with two competing ones |
| The steward is a person, with a successor on file | sh:class on the steward property, sh:minCount 1 on the successor | Owner fields holding a team mailbox, a departed employee, or nothing |
| The definition is long enough to mean something | sh:minLength on the definition | A description that reads "customer table" |
| The review date is in the future | sh:sparql comparing the review date with today | Approved records that expired quietly |
| Each term maps to at least one physical column | sh:minCount 1 on the mapping, sh:node on the mapping shape | Business terms nobody has connected to data |
| Each mapping carries its approver and basis | sh:minCount 1 on prov:wasAttributedTo and on the approval date | Mappings a model proposed and nobody confirmed |
| Codes come from the approved list | sh:in or sh:class against the reference data set | Values that conform to no vocabulary |
Each constraint in that table is defined in the W3C SHACL specification. SHACL carries three severities: Violation, Warning, and Info. A missing steward on an authoritative term is a Violation that blocks publication. A short description on a long-tail asset is an Info that feeds the queue. The validation report is itself an RDF graph, so each result can be stored against the record it tested. The W3C Data Quality Vocabulary defines a quality measurement as "the evaluation of a given dataset (or dataset distribution) against a specific quality metric". The rule, the record, and the result are then data on one open standard. The dashboard reads the same graph the agent reads.
A presence check inside a catalog answers whether a field is filled. It cannot test that two systems hold one definition, that a term resolves to a column, or that an approval has an expiry. The record is text. A shape can test all three, because the record is data.
A failed check is a work item with a rule and a record attached. The action depends on the dimension that failed.
| Finding | Action | Who acts |
|---|---|---|
| Missing definitions, stewards, or mappings | Let a model draft them. Every proposal lands in the steward's queue with its evidence. AI proposes, your experts confirm | Automation drafts; a named steward approves |
| Five meanings of one term | Approve one definition at the concept, with its scope, and map the forty columns to it | The data owner for the domain |
| Expired approvals | Make review cadence a governed state. Consumers refuse an expired record and route it to the steward | The steward, on the schedule the record carries |
| Model-written text with no approver | Approval writes the approver, the date, the version, and the review interval onto the record | The steward |
| Catalog descriptions drifting from the record | Sync the catalog from the record. Your catalog stays, and it stops inventing meaning | The platform team |
Our post Active vs Passive Metadata: Why a Well-Organized Catalog Still Goes Unused covers why an approved record that never reaches the tools fails in a different way. Our post Metadata Management Framework: Four Stages and a Self-Assessment places these checks on a maturity path. Our post What Is Data Governance for AI? covers who the steward is. Our post Metadata Management Tools for Governance and AI Readiness covers the tooling.
TQ Data Foundation holds governed metadata as RDF, SKOS, OWL, and SHACL, so every rule in this piece runs as a shape on the record itself. Records carry when they were last confirmed and by whom, so staleness is visible. Proposed mappings arrive scored, and a steward accepts or rejects them. AI proposes, your experts confirm. The standards are open ones TopQuadrant helped author, so the rules and the results export in full. We maintain 300 reusable standards-based models. Your catalog stays. It syncs from the governed record. Our metadata management use case page shows how TQ Data Foundation does this in practice.
Measure the slice that decides answers first. Write the rules as shapes. Put a name on every record that passes. That is metadata quality an agent can rely on.