Search for data quality dimensions and you get six words: completeness, uniqueness, timeliness, validity, accuracy, consistency. That list is real, it has a proper source, and it was written for customer and transaction records rather than for catalogues. Applied to a product record without translation it produces scores that look precise and change nothing. This page does the translation.
Where the data quality dimensions come from
Worth knowing, because most articles on this topic present a list with no attribution at all.
The academic root is Wang and Strong, “Beyond Accuracy: What Data Quality Means to Data Consumers”. It appeared in the Journal of Management Information Systems in 1996, volume 12, issue 4. They identified fifteen dimensions in four groups: intrinsic, contextual, representational and accessibility.
The six-word list most people quote comes from a DAMA UK Working Group paper of 2013. It is called “The Six Primary Dimensions for Data Quality Assessment”. Completeness, uniqueness, timeliness, validity, accuracy, consistency. It is short, it is usable, and it is the reason those six words appear on every data quality page on the internet.
The standards bodies went further. ISO/IEC 25012:2008 defines fifteen data quality characteristics. Five are inherent in the data itself: accuracy, completeness, consistency, credibility and currentness. Three depend on the system holding it: availability, portability and recoverability. Seven sit in both camps: accessibility, compliance, confidentiality, efficiency, precision, traceability and understandability.
ISO 8000 is the series that matters most for product data, because it was built around master data exchanged between organisations. ISO 8000-110:2021 covers syntax, semantic encoding and conformance to a data specification. ISO 8000-120:2016 covers provenance, ISO 8000-130:2016 accuracy, ISO 8000-140:2016 completeness. ISO 8000-61:2016 sets out a data quality management process model.
There is no single agreed list, and anyone telling you otherwise is selling something. The DAMA NL research paper by Van Nederpelt and Black (2020) collected sixty-five candidate dimensions. It reduced them to sixty preferred definitions. A 2025 review in Big Data and Cognitive Computing compared the major frameworks. Miller, Chan, Whelan and Gregório found they cover between four and fifteen dimensions each, with inconsistent naming.
One caution, because it catches people out. ISO/TS 8000-311 is titled “product data quality for shape”. That part is about CAD geometry, not catalogue attributes. Do not cite it in a catalogue context.
Why the standard six do not survive contact with a catalogue
The generic framing assumes a customer table. One row per person. A fixed set of columns. An internal system of record, and an owner who can be asked what the true value is. Product data has none of those properties.
The schema changes per category. A hex bolt needs eight attributes. A variable speed drive needs sixty. There is no fixed column set, so “completeness” has no denominator until someone defines one per category.
The truth lives outside your business. You do not know the tensile grade of a fastener. The manufacturer does, and they told you in a PDF. Accuracy is therefore a question about your supplier’s data, not about your data entry.
Sparseness is correct, not a defect. Ninety per cent of the fields on a catalogue-wide export are legitimately blank for any given product, because they belong to other categories. A tool measuring null rates across a flat export will report a catastrophe that does not exist.
We see the consequence on nearly every engagement that starts with a governance-led scoring exercise. The chart is produced, the number is low, and nobody can name a single record to fix. Our product data services work usually begins by throwing that chart away and rebuilding the measures per category.
The six data quality dimensions, rewritten for product catalogues
Completeness
Generic definition. The proportion of data stored against the potential for all of it to be stored.
Catalogue version. The proportion of category-mandatory attributes populated, measured per category, against a written definition of mandatory for that category.
The denominator is the whole argument. Define three tiers per category: mandatory (cannot publish without it), commercially important (drives filtering and comparison), and optional. Then score against tier one only. A catalogue at 96 per cent on tier one and 40 per cent on tier two is in decent shape. A single blended number hides both facts.
What failure looks like. A category where the mandatory list was never written down, so completeness is reported against every column that exists in the PIM. The score is meaningless and everybody knows it.
Uniqueness
Generic definition. No product should be recorded more than once.
Catalogue version. One record per real-world sellable item, at the agreed level, with identifiers that resolve to exactly one product.
Duplication in catalogues is rarely a straight copy. It is the same physical part bought from two suppliers under two part numbers. It is the item that came across in an acquisition and now exists twice with different descriptions. It is a GTIN reused on a successor product after a discontinuation.
How to measure it. Group by manufacturer part number plus brand, then look at the groups with more than one record. Then repeat on GTIN. Then, for the hard cases, group by a normalised signature of the key attributes. The third test finds duplicates the first two cannot.
What failure looks like. Two SKUs, two stock holdings, two sets of enriched content, one physical product in the racking.
Validity
Generic definition. Data conforms to the syntax of its definition.
Catalogue version. Every value conforms to the type, unit, format and permitted list defined for that attribute in that category.
This is the dimension that machines can test properly, and it is where you get the fastest return. Numeric attributes containing text. Units missing entirely. Values outside the permitted list on a controlled vocabulary. Dimensions expressed as “approx 1.5m”. Voltage recorded as “240v/110v” in a single-value field.
How to measure it. A rule per attribute, run over the whole catalogue, expressed as a pass rate. That rule library sits alongside our taxonomy and attribution work. The rules cannot be written until the schema is settled.
What failure looks like. A filter on the website returning nothing, because the value it needs is stored as free text with a stray unit suffix.
Accuracy
Generic definition. The degree to which the data correctly describes the real-world object.
Catalogue version. The value matches the physical product, checked against the manufacturer’s current published source.
Accuracy is the only dimension you cannot measure from inside your own database. It requires going back to the source document and comparing. That makes it expensive, so it is sampled rather than tested exhaustively.
How to measure it. Take a stratified sample, perhaps 100 SKUs across your top categories, weighted towards high revenue and high return rates. For each one, open the current manufacturer datasheet and check the mandatory attributes value by value. Record the error rate per attribute, not per product. That tells you which fields are unreliable, which is actionable.
What failure looks like. A specification that was right in 2019, a manufacturer revision in 2022, and no mechanism that connects the two.
Consistency
Generic definition. The same object is represented the same way everywhere.
Catalogue version. The same attribute carries the same meaning, unit and format across every product in a category. The same product carries the same values across every channel.
Two separate tests hide in here and they need separating. Internal consistency is whether “Length” means the same thing for every product in the category. Cross-channel consistency is whether the website, the trade portal, the marketplace listing and the printed catalogue agree with each other.
How to measure it. For internal consistency, profile the distinct values of each attribute within a category and read the list. Human eyes on 40 distinct values find the problem in a minute. For cross-channel, pick 50 SKUs and diff the published values across channels.
What failure looks like. “Colour” holding “Blue”, “blue”, “RAL 5010”, “Navy” and “Dark Blue / Black trim” in the same category. This is the most common finding in every catalogue we have ever profiled, without exception.
Timeliness
Generic definition. The degree to which data represents reality at the required point in time.
Catalogue version. Two clocks, not one. How old is the record against the manufacturer’s current revision, and how long does a new or changed product take to reach a sellable state.
The second clock is the one your business feels. Most distributors can quote new line to live on the website to within a week. The number is usually far worse than they would like. It also predicts revenue loss more directly than any completeness score.
How to measure it. Record the date each attribute set was last verified against source. Separately, measure elapsed days from supplier notification to published product, sampled over a quarter.
What failure looks like. Products live on the site whose data has not been checked against a manufacturer source since the original migration.
A worked example on one SKU
Take a fastener line item as it arrived from a supplier feed. Hex head set screw, M10 x 50, grade 8.8, zinc plated, sold in boxes of 100.
| Dimension | Test applied. | Finding |
|---|---|---|
| Completeness | 8 mandatory attributes for the Fasteners category. | 5 of 8 present. Missing head type, drive type, standard reference |
| Uniqueness | Group by brand plus manufacturer part number. | 2 records, one from each of two distributors, same physical part |
| Validity | Thread size against controlled list, length numeric with unit. | Length arrived as “50mm” in a numeric field. Fails type |
| Accuracy | Compare against manufacturer datasheet. | Grade recorded as 8.8, datasheet says 8.8. Passes. Coating recorded as “zinc”, datasheet says zinc flake. Fails |
| Consistency | Distinct values of Finish across the category. | “Zinc”, “ZP”, “Zinc Plated”, “BZP” all present. Fails |
| Timeliness | Last verified date against source. | Blank. No verification date held at all |
Six tests, one SKU, five failures. None of them is a data entry mistake. Every one is a structural gap: no mandatory list, no deduplication key, no type enforcement, no source-verification loop, no controlled vocabulary on Finish.
That is the point of running data quality dimensions properly. They convert a vague sense that the catalogue is poor into a list of specific mechanisms that are missing. The same six tests applied across the technical specifications in an industrial catalogue produce the same shape of answer.
Three data quality dimensions the standard list misses
Granularity
Is the record at the right level? A product, a variant and a sellable SKU are three different things, and catalogues routinely collapse them.
Size and colour held as free text on a single record rather than as structured variants is a granularity failure. So is one record per pack size where the pack size is not modelled as an attribute. Nothing in the standard six catches this. It is also more expensive to fix than everything else on this page combined. It changes the shape of the data model, not the values in it.
Comparability
Can two products in the same category be compared on the attributes that matter to a buyer?
This is stricter than validity and stricter than consistency. Both products can hold valid, consistently formatted values and still be incomparable, because one records bore diameter and the other records nominal size. Comparability is the dimension that decides whether your faceted search works, and it is the one your customers experience directly. Get the attribute schema right and comparability follows. Get it wrong and no amount of enrichment rescues it.
Provenance
Where did this value come from, and when?
ISO 8000-120:2016 exists precisely because master data exchanged between organisations needs to carry its origin. In practice almost no catalogue holds it. So when a customer disputes a specification, nobody can say where the value came from. A 2018 PDF, a supplier spreadsheet, a competitor’s website or somebody’s best guess. Holding source and verification date per attribute set costs very little at the point of enrichment and is nearly impossible to reconstruct later.
Syntactic, semantic and pragmatic: the more useful cut
ISO 8000-8:2015 splits quality three ways. For catalogues it is often a better starting frame than the six dimensions.
Syntactic quality asks whether the data conforms to its own specification. Right type, right unit, right permitted value. A machine can test all of it, cheaply, over the whole catalogue.
Semantic quality asks whether the data corresponds to the real thing it describes. Only a source document or a physical product can answer that. It has to be sampled.
Pragmatic quality asks whether the data is fit for the use it is put to. The same record can be pragmatically excellent for a trade counter and useless for a marketplace listing with a 200-character title limit.
The reason this cut is useful is that it explains why data quality tools disappoint. They test syntactic quality well, semantic quality not at all, and pragmatic quality only if you tell them what the channel needs. A catalogue can pass every automated rule in the platform and still be wrong about the product and unsuitable for the channel.
Turning dimensions into measures you can run
Data quality dimensions are a vocabulary, not a method. Four steps make them operational.
- Pick one category with real revenue behind it. Not the whole catalogue.
- Write the mandatory attribute list for that category on one page, and get a merchandiser to sign it.
- Write one testable rule per dimension per attribute. Validity rules first, because they are cheap and they find the most.
- Run the rules, sample fifty records by hand for accuracy, and publish both numbers together.
That last point matters. An automated pass rate published without a hand-checked accuracy sample is misleading. Once a board has seen a green number, it is very hard to reopen. The same discipline works on a builders merchant catalogue and on a half-million-line industrial parts file.
Key takeaways
- The six dimensions come from the DAMA UK Working Group (2013), with older roots in Wang and Strong (1996). ISO/IEC 25012 lists fifteen characteristics and ISO 8000 covers master data specifically.
- Completeness has no meaning until a mandatory attribute list exists per category. Everything else follows from that.
- Accuracy cannot be measured from inside your own database. It has to be sampled against manufacturer sources.
- Add granularity, comparability and provenance to the standard six. Catalogues fail on all three and the generic list does not test them.
- ISO 8000-8’s syntactic, semantic and pragmatic split explains why tooling alone never gets you there.
- Publish an automated pass rate and a hand-checked accuracy sample together, or the number will mislead.
Most catalogues do not need a scoring framework. They need to know which mechanisms are missing and what fixing them costs. A thirty minute call on your own data is usually enough to tell. Talk to us. You can also see our wider product data services, and the classification work that has to come first.