The PDQ Framework — 9 Dimensions of Product Data Quality

The classical data-quality dimensions — completeness, accuracy, consistency — were defined for human-readable systems of record. PDQ extends them with the dimensions that matter when the reader is an AI agent: structured-data presence, identifier integrity, extractability, and the copy-versus-spec accuracy penalty. The result is a measurement that maps directly onto the gaps an agent will hit when it crawls your storefront tomorrow.

The nine dimensions

Each dimension is scored 0–100 independently. The composite PDQ score is the weighted sum, minus the Accuracy penalty.

1. ACR — AI Catalog Readiness · 25%

Can an agent reliably extract title, identifiers, brand, price, availability, image, and spec from the product page? This is the gate. If the fundamentals aren't parseable, everything downstream — recommendations, comparisons, transactions — is invisible.

2. Completeness · 25%

Are the attributes Google Merchant (and the broader category vocabulary) expects for this product category present? Agents lean on category-specific attributes to filter and compare. A pair of running shoes without size, color, gender, and material is essentially un-recommendable.

3. Conformity · 10%

Do values follow expected formats — units, identifier checksums, codes, enumerations? A weight that's "12oz" in one product and "350g" in another and "Twelve ounces" in a third looks like three different things to an agent. Rule packs are category-aware: Vehicles, Books, Beauty, Electronics, Apparel, Food, Toys.

4. Consistency · 10%

Are units and vocabulary consistent across the catalog? Agents trust catalogs whose internal vocabulary is stable. Mixed units, synonym sprawl, and casing drift all signal data quality problems and reduce confidence.

5. Uniqueness · 10%

Are SKUs, titles, identifiers actually unique? Duplicate SKUs and identifiers break linking, comparison, and recommendation logic, and signal a catalog with cleanup debt.

6. Validity · 10%

Are critical and recommended fields present and non-empty? A field that's present in the schema but empty in the payload is worse than missing — it asserts a value and then provides none.

7. Referential Integrity · 5%

Do image URLs resolve? Are images on a CDN? Do SKU and category links work? Broken images and dead links signal a stale catalog; agents downgrade trust accordingly.

8. Accessibility · 5%

Is the spec data extractable as clean key:value pairs? Specs hidden inside images, infographics, or styled prose can't be reliably extracted.

9. Accuracy — penalty

Do the title and description claims match the spec table values? When copy contradicts the spec, conservative agents refuse to quote the product — or worse, propagate the wrong fact. This is a penalty, not a positive: catalogs lose points for contradictions.

Bands

• Excellent (85–100) — agent-ready. Re-audit quarterly.
• Good (70–84) — mostly there. Ship the top 3 gaps.
• Fair (50–69) — material gaps. Allocate a sprint.
• Poor (0–49) — foundational issues. Scope a project.

Framework provenance: v2 vs v1

PDQ v1 had eight sub-dimensions focused on data quality for human-facing systems. PDQ v2 expands to nine top-level dimensions, with the original eight rolled into the new ACR dimension. The change reflects the shift from optimizing for human readability to optimizing for agent parseability. ACR carries the heaviest weight because it's the gate everything else depends on.

Related

Agentic commerce readiness audit · GEO audit · AEO audit · Checklist