AI product catalogue enrichment.

What catalogue enrichment actually does

AI product catalogue enrichment turns thin supplier data into a complete product record: structured attributes (fibre, size range, fit, care, colour family), a title in your naming convention, a description, filter tags and alt text. The model reads the supplier’s spreadsheet row, any spec sheet PDF and the product photos, and proposes values for each field with the source it used.

Your platform may already do part of this. As checked on 19 October 2026, Shopify’s Sidekick assistant can generate product description text in the admin, and Cin7 lists AI-generated product descriptions among its features. Those are fine for a handful of products. A dedicated workflow is worth it when a buyer drops 600 new SKUs into a shared drive two weeks before a season launches and every one needs consistent attributes, not just prose. That is a document processing job with a merchandiser at the end.

How the workflow runs

Inputs

Supplier line sheets (XLSX or CSV), spec sheets and swing-tag artwork as PDFs, flat-lay and on-model photos, your attribute dictionary with allowed values, and a brand style guide for titles and copy.

Systems

Shopify (product CSV or Admin API, including product metafields), Cin7 Core or Omni as the inventory master, a shared drive for supplier files, and a model API on terms that exclude training on your data.

Steps

Match each row to a style and its variants; extract attributes from documents; derive visual attributes from photos only; map free text to allowed values; draft title, description and tags; run claim checks; send to the review queue.

Human review

A merchandiser approves every record before import. Anything carrying a regulated or performance claim (fibre content, country of origin, waterproof, UPF, “organic”) needs the source document open beside it.

Outputs

An import file or API payload with only the approved, changed fields, plus a log of which source supported each attribute.

Measures

Records approved without edits, fields left blank for lack of evidence, claim flags raised, hours from supplier file to live product, and “not as described” returns.

Risks

Invented attributes, claims without grounds, wrong variant mapping, overwritten good data, and copy that contradicts the physical label.

Illustrative example: a Christchurch outdoor clothing retailer

Illustrative example, synthetic numbers: a retailer with one store and a Shopify site loads 180 new styles (about 600 SKUs across colours and sizes) each season, twice a year. Assumptions: writing attributes and copy by hand takes 12 minutes a style; reviewing a model draft against the source takes 4 minutes; 15% of styles need a supplier query and 10 extra minutes.

The fee line assumes 6,000 input tokens per style (three photos, the supplier row, the attribute dictionary and instructions) and 900 output tokens, on OpenAI’s gpt-6.1-sol at US$2 per million input tokens and US$10 per million output tokens as listed on 19 October 2026. The saving is in review hours and in catching supplier gaps before launch, not in model cost.

The claims rule: no grounds, no claim

Under the Fair Trading Act 1986, section 13 prohibits false or misleading representations that goods are of a particular composition or quality, and about their place of origin. Section 12A goes further: a claim is unsubstantiated if you lack reasonable grounds when you make it, even if it turns out to be true. The Commerce Commission notes that substantiating a claim afterwards may not be enough.

So the workflow carries a hard rule: descriptions may only restate facts present in the source documents. A claims list (fibre percentages, “merino”, “recycled”, “waterproof”, “Made in New Zealand”, ratings and certifications) triggers a check that the claim appears in a named supplier document. If it does not, the field stays blank and the record is flagged. The Commerce Commission’s consumer information standards also require fibre content and care labelling on clothing and textiles, and country of origin labels on clothing and footwear, so the online record must never contradict the physical label.

What stays with a person

The merchandiser decides what goes live, chooses the hero attributes and approves any claim. Buyers resolve supplier queries. Nobody lets the model set prices, publish directly, or merge variants it was unsure about.

Where catalogue enrichment goes wrong

• Blank cells wipe good data. Shopify’s help centre says that when you import with “Overwrite products with matching handles”, blank non-required cells overwrite existing values with blanks (Shopify CSV guide). Export, change only approved columns, and re-import that subset, or write through the API.
• Variant metafields are not in CSV. The same guide says product metafields import by CSV but variant metafields do not. Plan size-level attributes for the API instead.
• Photo inference creep. A model shown a shiny jacket will happily call it waterproof. Limit photo-derived fields to what a photo can prove.
• Template sameness. Hundreds of descriptions with the same opening read as filler. Vary structure from real product differences, and check with a sample read before import.
• Unchecked confidence. Validate every field against allowed values; our guide to structured outputs and validation shows how.

Catalogue enrichment pairs naturally with customer question answering, which depends on complete records, and with inventory exception reporting. The retail sector guide shows where each fits, and our Shopify integration notes cover the API side.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.