Product Catalog Matching: Models, Variants and Explainable Links
Separate models, variants and offers with identifier checks and reversible matching decisions.
KAIROXO
Product catalog matching asks whether two listings describe the same purchasable item. It is different from finding products with similar names. A headset sold alone, a headset bundled with an adapter and a refurbished headset can share most of their title while representing three different offers. Merging them can corrupt inventory analysis before any price comparison begins.
Start with a narrow category and a reviewable matching policy. An electronics catalog may rely on manufacturer identifiers and regional variants; apparel needs size, color and pack details. The goal is an explainable relationship between source listings and a canonical product, with unresolved cases kept visible rather than forced into a match.
Keep product, variant and offer separate
Use three conceptual layers. The product describes a model family, the variant describes a specific configuration, and the offer describes a seller's purchasable listing. A model family can contain several storage capacities and colors. The same variant can have multiple sellers, conditions and bundles. Store source URLs and identifiers at the offer layer so matching never removes access to the original evidence.
Write the matching target into the project contract. Matching model families can support assortment research, but fulfillment requires a stricter variant or offer match. A result labeled simply as same product is ambiguous unless the consumer knows which layer it means. Keep family-level links distinct from exact-variant links in exports and interfaces.
Normalize identifiers without manufacturing them
Collect GTIN, brand, manufacturer part number and source SKU when published. A source SKU is usually meaningful only inside that retailer, so do not treat identical SKU strings from two sites as global identity. Preserve raw values before normalizing whitespace or case, and record any identifier validation error rather than silently repairing a digit.
An identifier match is strong evidence only when its context agrees. Sellers sometimes copy a parent identifier across variants or publish malformed values. Check model, capacity and condition around the identifier. Missing identifiers should reduce confidence or trigger another path, not become empty strings that accidentally match every incomplete record in the dataset.
Describe attributes at the correct granularity
Map category-specific attributes into typed fields while retaining their source text. Normalize quantities with units: a two-pack and a single item cannot be compared by a bare number. Distinguish dimensions of the product from shipping package dimensions. Similarly, a compatible-with statement does not establish that an accessory is the device it supports.
Evidence | Helpful use | Typical mistake |
|---|---|---|
Manufacturer model | Identify a model family | Ignoring regional suffixes |
Capacity and color | Separate variants | Dropping a missing attribute |
Pack quantity | Identify the purchase unit | Treating a bundle as one item |
Condition | Separate new and refurbished | Inferring condition from price |
Generate candidates before deciding matches
Restrict candidate pairs by plausible brand and category before using title similarity. This reduces meaningless comparisons and makes review queues smaller. A text model can suggest candidates, but keep its score separate from the final decision. Scores from different models or categories are not automatically comparable probabilities.
Use explicit outcomes such as exact_variant, same_family, incompatible and needs_review. Record the evidence, rule version and reviewer for each accepted link. A deliberate unresolved result is preferable to a confident false merge. For a fictional 256 GB phone, a nearly identical 128 GB title should be rejected as an exact variant even if its similarity score is high.
Review hard negatives as well as easy matches
Build a labeled sample with confusing pairs: regional editions, compatible accessories, multipacks, refurbished items and products differing by one character. Measure precision on accepted links and recall on known matches separately. A system that matches very few items may look accurate while leaving most useful records unresolved.
Inspect performance by category and source, not only one overall average. Keep a holdout sample that was not used to tune rules. When an extraction change alters attributes, rerun matching against saved observations and compare decisions before replacing the catalog. This isolates a parsing regression from a genuine change in how products are represented.
Preserve reversible catalog decisions
Do not overwrite source records with canonical values. Store a relationship table or document linking observations to canonical products, with timestamps and versioned reasoning. If a reviewer discovers an incorrect merge, split the relationship and recalculate dependent reports without recollecting every page. This is particularly useful when a new retailer introduces identifiers with different semantics.
Keep access and matching responsibilities separate. A reliable HTTP response can contain a recommendation card, an unavailable offer or a consent page instead of the requested product. Validate page identity and required attributes before admitting a record to matching. KAIROXO can support the documented access stage; it does not establish catalog equivalence by virtue of delivering content.