Automated product categorization for marketplace catalogs

Automated product categorization maps seller listings into the categories your marketplace actually uses. A useful system returns a valid category ID, supports its choice with relevant attributes, and leaves ambiguous listings for review.

It should not invent a new category every time a seller uses unfamiliar wording.

Start with your taxonomy

Suppose a seller describes an item as “small wooden bedside thing.” Your marketplace might classify it under Furniture → Bedroom → Bedside tables. The system needs the allowed categories and their definitions to make that choice.

Store category IDs separately from display names. Names can change without breaking historical mappings. Include exclusions and examples: a wall shelf placed beside a bed is not necessarily a bedside table.

Choose whether a listing can have one primary category, several category relationships, or separate browsing and reporting classifications. Do not leave that decision to the model on each request.

Use rules for known mappings

If an approved supplier consistently sends “nightstand” for bedside tables, a mapping rule may be enough. Reserve model inference for unfamiliar descriptions, mixed catalogs, or cases where several attributes must be considered.

Keep seller-specific mappings when the same label means different things across suppliers. An industrial supplier's “accessories” can refer to a different product family from a fashion seller's “accessories.”

Record which rule or model version produced the proposal. This makes a new error pattern traceable when a taxonomy or prompt changes.

Classify from evidence, not attractive wording

Use the product's function, form, and known specifications. Separate these from promotional descriptions. “Perfect for travel” does not turn a full-size dining chair into travel equipment.

For each proposal, keep the original category, proposed category ID, relevant input, and any unresolved ambiguity. If two categories remain plausible, show the alternatives to a reviewer rather than publishing a generic confidence label.

Keep external channel categories separate

Google has its own product taxonomy and a google_product_category attribute. That is a channel-specific classification. Your marketplace may need a more detailed or differently organized browsing hierarchy.

Maintain an explicit mapping between internal categories and external destinations where needed. Changing your internal category should trigger validation of the export mapping, not silently send the same string to every channel.

See B2B catalog onboarding for a case where industry-specific classifications and unit-of-sale requirements need additional care.

Evaluate the categories that matter

Build a checked sample with frequent categories, rare categories, and confusing boundaries. Include items that should not be classified because they are out of scope.

Measure precision by category and inspect the confusion pairs. A high average can hide a systematic error between replacement parts and complete products. Review results whenever categories are added, combined, or renamed.

Do not treat a numerical confidence score as a guarantee. Choose automatic-publication thresholds using checked examples and the consequences of a mistake.

Update the rest of the listing

A new category can change required attributes, search filters, and display rules. Validate the listing against the target category before publishing the change. Otherwise, categorization can create a new incomplete record.

Goodfolk's catalog worker is intended to prepare these corrections against your approved data model. Start with one taxonomy branch and a review queue. Use the data quality audit to check whether the changes make the catalog more useful rather than merely more consistently labeled.