AI Catalog Enrichment for a Multi-Brand Retailer
A pipeline that turns messy supplier feeds into complete, consistent product pages: attributes extracted, titles normalised, descriptions written in the brand voice, and every change reviewed before it goes live.
Supplier data arrived in 60 different formats. A third of products were missing key filter attributes, so they never appeared in filtered search, and writing descriptions by hand took about 20 minutes per product.
The approach.
- 01Defined the attribute schema per category with the merchandising team.
- 02Built extraction from supplier feeds, spec sheets and images into that schema, with a confidence score per field.
- 03Generated titles and descriptions in the brand voice from a style guide and approved examples.
- 04Put a review queue in front of publishing: low-confidence fields and new categories always go to a person.
- 05Synced approved records back to the PIM and storefront nightly.
Claude extracts attributes from text and images, writes titles and descriptions to the style guide, and flags fields it cannot support from the source.
How we measure it.
Time per product is measured in the review tool. Attribute coverage is counted per category against the agreed schema. Conversion is compared for filtered-search sessions against the same weeks before launch.
Before you ask.
Not at first. Beyond Technologies routes every record through review until a category's approval rate is consistently high, then only low-confidence fields go to a person.
Yes, for visible attributes such as colour, material and style. Technical specifications that are not visible are flagged for a person rather than guessed.
Want this measured on your own data? Start with a pilot.
Let's put AI to work in your business.
A 30-minute call. You bring the workflow or the roadmap — we'll tell you what's feasible, what it costs, and what we'd build first.