Product matching: keep source records behind each link
Retain original product records and the evidence for each proposed match. A Mars Petcare industry reference with a practical dataset example.
Harmonya’s Mars Petcare case describes gaps in connecting retailer-specific identifiers to UPC-level product information across online retailers. The provider reports linking previously disconnected IDs to UPCs and enriching scattered category and manufacturer details. The page includes a named Mars Petcare data-enablement manager’s commentary on improved coverage. It does not publish pair-level match evidence, thresholds or a duplicate-deletion policy. The data-delivery proposal below applies those distinctions to a separate example.
Define the object that the match should identify
For a marketplace dataset, PandaBuyNow proposes agreeing whether a relationship identifies the same product family, a selected variant or a specific selling unit. Keep the source platform, listing identifier, variant identifier where available, source URL and observation time with each original record. Give any internal row identifier its own field.
Deliver a separate relationship table containing the two source-record references, proposed match level, supporting attributes, conflicting values and review status. An accepted link can connect records while preserving the different listings and offers they describe. Retain a rejected candidate with its reason when that history is useful to the recipient.
Test duplicate-looking rows with a fictional example
Consider invented records A-11 and B-22. Both name MUG-C, but A-11 describes one mug and B-22 a pack of two. For a task requiring the same selling unit, preserve both records and mark the proposed link as rejected for a pack-count difference. A shared title alone does not resolve that distinction.
A third record, C-33, describes one mug but leaves its colour unspecified. Retain it as a candidate with colour evidence unresolved. These invented identifiers and details describe no Mars Petcare or PandaBuyNow product records.
Make the matching decision readable in the export
Agree the status vocabulary and the evidence required for each state. A CSV relationship row or a JSON object should let the recipient find both source records and understand why a candidate was accepted, rejected or left open. Keep observed values separate from normalized comparison values, with the transformation rule where relevant.
Before expanding the dataset, review representative data rows containing a supported match, a selling-unit mismatch and an unresolved attribute. For a PandaBuyNow discussion, provide product URLs, available identifiers, the matching level your application needs and its expected JSON or CSV structure. Agree feasible fields, source references and data-quality notes in the delivery scope.