Product data enrichment checklist

Product data decays. Suppliers change specs, new SKUs arrive half-filled, categories get reorganized, and marketplaces tighten their rules. Enrichment is the work of filling and fixing that data so each listing is complete, consistent and true to its source. This checklist covers what to define first, what to check on every listing, and how to know the work is right.

Before you touch a listing: write the spec

Enrichment without a written target produces confident inconsistency. For each category, write down:

• [ ] Required attributes, and which are optional
• [ ] Allowed values for each attribute, as a controlled list where possible
• [ ] Units and formats: cm or in, decimal places, how ranges are written
• [ ] Title formula, for example product name + product type + capacity + color
• [ ] Trusted sources, in order: manufacturer sheet, supplier feed, physical sample, approved website
• [ ] What "unknown" means: leave blank and flag, never guess
• [ ] Channel rules for each marketplace you sell on. They change, so check each channel's current requirements instead of relying on last year's notes.

The listing checklist

Identity

• [ ] SKU is present and unique
• [ ] GTIN, UPC or EAN passes its check digit (see below)
• [ ] Brand and manufacturer part number match the source
• [ ] No second listing exists for the same product

Title

• [ ] Follows the title formula for the category
• [ ] No supplier codes, all-caps words or keyword stuffing
• [ ] Agrees with the attributes. A "black" title on a "navy" product is an error.

Category

• [ ] Mapped to the deepest correct node of your category tree
• [ ] Consistent with similar products already in the catalog

Attributes

• [ ] Every required attribute filled from a trusted source, or left blank and flagged
• [ ] Values come from the allowed list only
• [ ] Units converted and rounded the same way across the catalog

Variants

• [ ] Parent and child relationships are correct
• [ ] Variant attributes such as size and color differ between children and match each child

Images

• [ ] Image count and main-image rules meet each channel's current requirements
• [ ] Each image shows the variant it is attached to
• [ ] Alt text describes the product

Description

• [ ] Facts in the description match the attributes
• [ ] AI-written text is checked against the source before it goes live
• [ ] No claims the source does not support

Provenance

• [ ] Each filled value records where it came from, so anyone can re-check it later

How to check a GTIN check digit

GTINs, UPCs and EANs end in a check digit. To verify one, take the digits before the check digit. Starting from the rightmost of them, multiply alternately by 3 and 1, and add the results. The check digit is the smallest number that brings that total to a multiple of 10, so it is 0 if the total already is one.

Illustrative example. For the 12-digit code 012345678905, the first 11 digits weighted from the right give a total of 85. Adding 5 reaches 90, so the check digit should be 5, and it is. A mismatch means the code was mistyped or invented. A spreadsheet formula can run this check over the whole catalog before anyone enriches a single field.

Worked example

Illustrative example. Invented data, not from a customer.

Current listing: title "SS bottle 750 blu" · category "Kitchen" · capacity (blank) · material (blank) · color (blank) · dishwasher safe (blank) · weight (blank)

Supplier sheet: Summit Insulated Bottle, 750 ml, stainless steel, blue, hand wash only

Title rule: product name + product type + capacity + color

After enrichment:

• Title: Summit Insulated Water Bottle, 750 ml, Blue
• Category: Kitchen > Drinkware > Water Bottles
• Capacity: 750 ml (25.4 fl oz, if your catalog shows both)
• Material: Stainless steel
• Color: Blue
• Dishwasher safe: No (the supplier sheet says hand wash only)
• Weight: left blank and flagged, because the supplier sheet does not list it

The flag goes on a list for the buyer to request from the supplier. Nobody guessed a weight.

Where to start when the catalog is large

You rarely have the hours to fix everything at once. Rank the work so the first weeks pay off:

1. Listings that sell. Start with the SKUs that bring in most revenue or traffic. A missing size on a best seller costs more than one on a product nobody views.
2. Listings that block. Fix anything a marketplace has rejected or suppressed for missing data.
3. Attributes that drive filters. Size, color, material and compatibility decide whether a product appears when shoppers filter. Fill those before descriptive extras.
4. New SKUs at intake. Enrich new products as they arrive, so the backlog stops growing while you work through it.
5. The long tail. Work through the rest by category, so one person learns one set of rules at a time.

Track progress as listings meeting the spec per category, not as edits made.

Checking the work

1. Gold listings. Before live work, have your category owner fill 30 to 50 listings and keep the answers. Use them to train and test anyone doing the work.
2. QA sampling. Each week, re-check a sample of finished listings against the source and the spec. Sample across every person and every category, not only the busy ones.
3. Error types. Log each error as one of: wrong category, missing attribute, wrong value, unit error, style break, variant error, image mismatch. The pattern tells you whether to fix training, the spec or the source data.
4. Decision log. When someone asks whether teal is Blue or Green, answer once, date it, and add it to the spec.

Pitfalls

• Guessing. A plausible wrong value is harder to find than a blank.
• Copying from other retailers' listings. Their data may be wrong, and their text may be protected. Use manufacturer and supplier sources.
• Bulk edits without a preview. One bad find-and-replace can overwrite thousands of good values. Export first.
• Unit drift. Rounding 25.36 fl oz to 25 in one place and 25.4 in another creates false duplicates in filters.
• Taxonomy changes mid-project. Version the category tree and record which version each listing used.
• Unreviewed AI descriptions. Generated text can invent features. Check it against the attributes, the same way you would review any model output.

A short version to pin above your desk

• [ ] Spec written per category
• [ ] Identifiers validated before enrichment starts
• [ ] Title formula applied
• [ ] Deepest correct category
• [ ] Required attributes filled from trusted sources, or flagged
• [ ] Variants consistent
• [ ] Images match the variant and channel rules
• [ ] Descriptions match the attributes
• [ ] Source recorded for every value
• [ ] Weekly QA sample with error types

If you want a trained team to run this checklist in your store admin or PIM, see product data enrichment. For AI-written descriptions, the LLM output review playbook covers how to check them, and how we check quality explains gold sets and sampling in more detail.