Definition
You measure catalog quality by defining a required attribute set for each category, then computing four rates against it across the catalog: coverage, validity, accuracy, and conflict. Coverage is how many required fields are populated; validity is how many populated values fall inside their allowed range or vocabulary; accuracy is how many values match the real product when checked against it; conflict is how often two sources disagree about the same fact. Structure and freshness, the other two dimensions of catalog quality, are checked separately — as schema conformance and as a value-age distribution — because neither reduces to a rate over a required attribute set. The remaining machinery of a measurement exercise — sampling, reporting, re-running — exists to make those four rates trustworthy and repeatable.
Key points
- Measurement starts with a written per-category requirement list; without it there is no denominator and coverage is meaningless.
- Automated checks give coverage and validity; accuracy needs a manually verified sample, because a machine cannot tell a plausible wrong value from a right one.
- Report rates by category and by source, not as one catalog-wide average.
- Record the run date and the requirement-list version, so the next run is a comparison rather than a fresh opinion.
What are the steps to run a catalog quality measurement?
Step one: pick the categories in scope and write down, per category, which attributes are required, recommended, and irrelevant. Step two: define the allowed form of each — unit, vocabulary, numeric range, identifier format. Step three: run automated checks across all products in scope and compute coverage and validity rates per attribute. Step four: pull a random sample of perhaps fifty products per category and verify values against the physical item or the supplier document, giving an accuracy rate the automated pass cannot produce. Step five: cross-reference duplicate records for the same product across sources to get a conflict rate. Step six: report all four rates — coverage, validity, accuracy, and conflict — split by category and by source, dated and tied to the requirement-list version you used.
Common pitfalls
- Measuring only what the existing tooling already reports, which quietly redefines quality as whatever is easy to count.
- Skipping the manual sample, so accuracy never gets measured at all and a catalog full of confidently wrong values scores well.
- Sampling non-randomly — usually from the best-maintained top sellers — and generalizing the result to the long tail.
- Re-running with a changed requirement list and reading the difference as improvement.
FAQ
How is measuring catalog quality different from catalog quality scoring?
Measurement is the procedure that produces raw observations — coverage, validity, accuracy, and conflict rates per category and source. Scoring is an optional layer above it that weights and compresses those observations into a single tracked number. You can measure thoroughly and never build a score.
How often should a catalog be measured?
Automated coverage, validity, and conflict checks are cheap enough to run on every ingest or on a nightly schedule. The manual accuracy sample is expensive, so most teams run it on a slower cadence and after any event likely to shift accuracy, such as onboarding a new supplier or a bulk import.
Source
The requirement list that step one depends on is usually anchored to a published specification rather than invented from scratch; Google's Merchant Center product data specification states which attributes are required, optional, or conditionally required per product type, and gives the allowed formats those validity checks test against.