Notes · Catalog
AI catalog enrichment: sample before you scale.
Generated product data fails in predictable ways: invented specifications, drifting categories, and copy that sounds right but is wrong. A sample batch catches those failures before they reach thousands of SKUs.
Why catalog AI fails
Supplier data arrives as PDFs, titles, and spreadsheets with inconsistent fields. Ask a model to fill the gaps and it will, including the gaps it cannot know. The output is plausible, which makes it harder to catch.
Three failures matter most: invented attributes, taxonomy drift (the same product in two categories), and localisation that changes units or compatibility without changing the source.
Define the source of truth
Every generated field needs a source. Product title, specification sheet, supplier feed, or manual entry. If the source is missing, the field stays empty and routes to review.
- One attribute schema, shared across markets.
- Allowed values per field, not free text.
- Unit rules: metric and imperial, with a single canonical unit.
- Category tree with merge and dedupe rules.
- A rule that forbids generating safety, compatibility, or warranty claims without a source.
Run a sample batch first
Pick 200 to 500 SKUs across categories, suppliers, and data quality. Run the pipeline, then score accuracy field by field. Aggregate accuracy hides the fields that matter.
| Field | What to check | Acceptance |
|---|---|---|
| Attributes | Values match the source, including units. | Per-field target agreed before the run. |
| Taxonomy | Product sits in one correct category. | No duplicate categories. |
| Description | Claims trace to attributes. | No unsupported specs. |
| Localisation | Spelling, units, and market fit. | Source meaning preserved. |
Score a random sample by hand and include the worst-looking records. If the sample misses the target, change the schema or the source, not the prompt wording alone.
Confidence and review
Every generated record gets a confidence score and a link to its source. Low-confidence records go to a review queue with the source attached. The reviewer approves, edits, or rejects.
Publication is a separate step. A record that is generated but unreviewed stays out of the storefront.
Localise after the source is stable
Do not translate a catalog that is still changing. Stabilise attributes and taxonomy first, then generate market variants.
- US and Canada first if those are your live storefronts.
- Keep SKU identity and technical values unchanged across markets.
- Localise spelling, units, and regulatory copy separately.
- Re-run the sample evaluation per market.
Track cost per SKU
Add model tokens, retrieval, QA sampling, and review time per batch. Divide by published SKUs, not processed SKUs. Compare with the cost of manual enrichment and the revenue risk of bad data.
A pipeline that generates cheap data nobody trusts is more expensive than an empty field.
Sample-batch checklist
| Item | Pass condition |
|---|---|
| Schema | Fields, allowed values, and units documented. |
| Source rules | No source, no generated value. |
| Sample | 200+ SKUs across categories and quality tiers. |
| Accuracy | Per-field targets met on a blind sample. |
| Review | Low confidence routes to a queue, not the storefront. |
| Cost | Cost per published SKU below the manual baseline. |
Next
If the catalogue is the bottleneck, start with a schema and a sample batch before any bulk run.