Set up catalog collections and AI filters
Review product data, turn useful metafields into filters, organize collections, and test product discovery end to end.
Catalog collections and AI filters help the Agent narrow a product set in ways that match shopper language. The configuration is only useful when the underlying products and metafields are complete, current, and meaningful for the active store or Shopify market.
This guide covers the current Catalog Products, Filters, and Collections areas, including AI-assisted filter suggestions, manual review, activation, and the test panel. It avoids treating a generated suggestion as automatically correct.
How this fits into Humind
Catalog data is the factual base for product discovery and recommendation. Integration state, product visibility, metafield filters, collections, promotions, markets, and storefront deployment all influence what a shopper can see and what the Agent can explain.
Verify source data before compensating with Agent guidance. If a price, variant, collection, or attribute is wrong in the catalog, correct or resync that source first, then retest the exact product journey in Humind.
Before you start
Access: Catalog read access is required to review products, filters, and collections. Catalog write access is required to change visibility or filter configuration.
- Confirm the correct company, catalog source, and Shopify market if applicable.
- Wait for the relevant catalog sync to complete and review a representative product sample.
- Choose shopper questions that should narrow by attribute, collection, or promotion.
Step-by-step workflow
Verify product records
Open AI Agent, Catalog, then Products. Search representative products and inspect the fields the Agent will need: title, description, variants, price, availability, collections, and useful metafields. Product visibility determines whether a record should participate in the customer experience.
Correct source data or resync before building filters on top of an inconsistent field. A filter cannot compensate for blank, mixed-format, or incorrectly scoped metafield values.
Review suggested filters
Open Catalog, then Filters. Humind can analyze catalog metafields and propose filter candidates. Review each candidate's display name, type, data type, unit, description, examples, synonyms, and buckets where those fields apply.
Activate only attributes shoppers understand and products use consistently. The interface keeps active filters visible and allows reanalysis, but reanalysis should follow a meaningful catalog change rather than become a routine reset.
- Use shopper language in display names.
- Reject internal or technical fields that should not guide recommendations.
- Check units and value normalization across products.
Organize product collections
Open Catalog, then Collections. Use collections for product groups that have a clear merchandising or discovery purpose. Confirm membership and naming against the source catalog before relying on the collection in tests or widgets.
Keep collections and filters conceptually separate: a collection groups products, while a filter narrows by an attribute. A product can belong to a collection and still need several filters for shopper preferences.
Test discovery behavior
Use the Filters test panel or Test product to ask realistic questions. Include one query that should use the new filter, one that combines two criteria, one with a synonym, and one that should return no matching product rather than invent a result.
Then test the broader journey in Playground. Review the selected products and explanation together, not only whether the filter name appears in text.
Permissions and important caveats
- AI suggestions are drafts for merchant review, not automatic approval of every metafield.
- Saving filter changes can trigger background materialization; allow the update to finish before judging results.
- Changing product visibility affects whether the Agent can use that product.
- Collections and promotions can be market-specific through the active company context.
Verify the result
Use this checklist before considering the work complete:
- Representative products contain consistent values for every active filter.
- Filter names, types, synonyms, units, and buckets match shopper language and source data.
- Collection membership is correct for the active catalog or market.
- Positive, combined, synonym, and no-result product tests behave honestly.
Troubleshooting
A useful field is not suggested
Confirm the metafield exists on enough current products and has consistent values. Reanalyze only after the source catalog is ready, or review the field through the available filter editor.
A filter is active but test results ignore it
Wait for the save and reindex work to complete, confirm the selected products have the expected values, and retest in a fresh session. Disable the filter if the source values are unreliable.
A collection contains unexpected products
Check the collection in the source catalog and the active company or market. Correct membership at the source when it is synced rather than repeatedly patching downstream behavior.