Record

Minimum publication package

case_id
CW-2026-02-BEAUTY-CATALOG
industry
Beauty / personal care
scope
Product detail pages, variant (shade/size) offers, brand disambiguation, and ingredient/claim passages used by retrieval and answer surfaces.
approved_by
NRLC editorial gate
approved_at
2026-09-12
release_id
unavailable — withheld
deployed_at
unavailable — withheld

Admission chain

Seven separate records

  1. 01

    Observed condition

    Multiple shade/size variants shared one Product URL identity with a single Offer, while other locales minted duplicate Product nodes under near-identical names. Ingredient lists and marketing claims in HTML did not match the properties attached to JSON-LD Product nodes. Brand markup sometimes named the parent conglomerate and sometimes the house line without disambiguatingDescription.

  2. 02

    Evidence

    unavailable — withheld: PDP HTML and JSON-LD pairs for representative SKUs unavailable — withheld: variant matrix export (shade × size × GTIN/SKU) unavailable — withheld: answer-engine prompt samples showing attribute mix-ups (screenshots)

  3. 03

    Decision

    Stabilize a product-group model: one ProductGroup (or clear isVariantOf relationships) with distinct Offers per purchasable variant; bind Brand once with disambiguation; force ingredient and regulated-claim passages to render from the same catalog fields that populate structured properties.

  4. 04

    Intervention

    Implemented catalog-driven PDP templates: variant Offers with SKU/GTIN, color/size properties, availability, and price; Brand node with stable name + disambiguatingDescription; ingredient and claim blocks generated from the same PIM fields as structured data; removed orphan duplicate Product graphs on locale mirrors.

  5. 05

    Verification

    For a frozen set of SKUs, extract Product/Offer/Brand JSON-LD and compare name, identifiers, shade/size, price, availability, and ingredient/claim text against the PIM export and rendered HTML.

    unavailable — withheld: post-change structured-data extracts unavailable — withheld: PIM vs HTML vs JSON-LD field reconciliation

  6. 06

    Result

    On the verified SKU set, variant Offers, Brand identity, and ingredient/claim passages matched the catalog source-of-truth. Duplicate Product graphs on the sampled locale mirrors were removed or redirected to the canonical PDP.

    Technical verification only — no sales lift, ranking, or AI citation-rate claims are published for this case.

  7. 07

    Limitations

    Client and product-line names are withheld; raw artifacts are not public. Retailer and marketplace listings were out of scope. Cosmetic claim compliance review was owned by the client’s regulatory team; this case covers consistency of published facts, not legal clearance of claims. No claim that third-party AI systems immediately corrected prior wrong answers.

Decision record

Alternatives considered

  • Only fix JSON-LD without touching PDP copy — rejected; visible text would remain the stronger retrieval span
  • Aggregate all shades under one Offer — rejected; availability and price drift by shade
  • Strip structured data sitewide — rejected; removes merchant eligibility without resolving entity confusion

Methods used

Inherited public methodology

  • Entity disambiguation and Product/Offer modeling
  • Catalog source-of-truth alignment
  • Decision traces for retrieval-visible passages
Decision TracesSeparates observation, evidence, interpretation, decision, action, and verification. Enterprise SchemaLarge-scale JSON-LD and source-of-truth records.