Product data enrichment: how to fix incomplete product catalogs

Product data enrichment is the process of filling in missing product fields and correcting wrong ones, such as GTIN, brand, size, color, material, category and images, using verified outside sources. Done well, every added value can be traced back to a source page, so a person can check it before it reaches your store, your marketplace listings or your shopping feeds.

The terms catalog enrichment, product information enrichment and PIM enrichment mean roughly the same thing. PIM enrichment usually refers to doing this work inside a product information management (PIM) system. Product attribute enrichment is the narrower case of filling structured attributes such as size, material and color.

Why do product catalogs end up incomplete?

Catalogs end up incomplete because product data arrives from many places in many shapes. Supplier spreadsheets use their own column names. ERP systems hold cost, stock and SKU, but rarely the fields shoppers filter on. Bulk imports often bring a title, a price and one image, and nothing else.

The gaps cause visible problems. Shoppers cannot find a product when the size or color filter has no value to match. Feeds get rejected when a required attribute is empty. Listings look thin next to competitors that show full specifications.

Which product fields usually go missing or wrong?

The table below lists the fields that most often need enrichment and why each one matters. The Google Shopping rules come from the Google Merchant Center product data specification, which is the reference for feed attributes.

Country lists and thresholds change over time, so check the Merchant Center page for the markets you sell in before you set required fields.

Where does enrichment data come from?

Enrichment data comes from five main kinds of sources. They differ in how much you can trust them and how much work they take.

GS1 describes a GTIN as a number a company uses to uniquely identify its trade items, meaning products or services that are priced, ordered or invoiced in the supply chain (GS1 GTIN standard). That is why a valid GTIN is the strongest single key for linking your record to outside data.

Why does product matching come first?

Product matching comes first because you can only copy an attribute from another listing once you are sure it describes the same product. If you match a 500 ml bottle to the 1 liter version of the same product, you copy the wrong size, the wrong GTIN and possibly the wrong images. The enrichment step then spreads the error across every channel you publish to.

A safe match uses identifiers before text. A matching GTIN, or brand plus manufacturer part number, is strong evidence. A similar title or image on its own is weak evidence and should go to a person. The product catalog enrichment page describes this rule as making uncertain matches visible: when identifiers do not agree, the match is held rather than assumed.

For a full guide to matching methods and their failure cases, see product matching.

How does a product data enrichment workflow work, step by step?

A reliable workflow has three stages: match, map and review. Each stage keeps the evidence for every value so that you can trace and undo changes.

1. Audit what is missing. Count empty or placeholder values per field and per category. Start with fields that block feeds or filters.
2. Pick sources per field. Decide which source wins for which field. For example, specs from the brand site, GTIN validity from GS1, category from your own taxonomy.
3. Match each SKU to its source record. Use identifiers first, never the title alone. Inspect variants, because a parent match does not mean every size or color matches.
4. Map source fields to your fields. Write the mapping down. For example, a Shopify store's vendor field maps to brand, product_type maps to a source category, and a variant's compare_at_price maps to a sale reference price. Keep the raw source value next to the normalized one, and treat a source category as the seller's taxonomy, not your final category.
5. Test on one record first. Run the mapping on a single product, compare the output to the live source page, then scale up.
6. Validate automatically. Check GTIN length and check digit, allowed values for fields like color and gender, unit formats, and required fields per category.
7. Send proposals, not overwrites, to review. Show each proposed value with its source URL and retrieval time. Do not overwrite a field your team owns without review. Show conflicting values side by side for the field owner.
8. Label inferred values separately. A value read from a source page and a value generated by a model are different kinds of data.
9. Leave gaps open when evidence is missing. If no source confirms a field, mark it unresolved instead of filling it with an assumption.
10. Publish and re-check. Push approved values to your PIM, store and feeds, and re-run the audit on a schedule.

Can you use AI and LLMs for product data enrichment?

Yes, large language models (LLMs) are useful for enrichment, but they need a check against the source page. They work well for pulling attributes out of long descriptions, normalizing values ("navy blue" to "blue"), suggesting a category from your taxonomy, and drafting descriptions from confirmed attributes. Many PIM tools now include these features. Akeneo, for example, describes AI features for generating, enriching and translating product content, with outputs that can be reviewed before publishing (Akeneo PIM).

The risk is hallucination: a model states a value that is not in the source. Researchers at OpenAI and Georgia Tech argue that this happens partly because training and evaluation reward guessing over admitting uncertainty, and they note that "guessing when uncertain improves test performance" (Kalai et al., 2025). In product data, that looks like a plausible GTIN, a material that is common for the category, or a size the product does not come in.

You can reduce the risk with a few rules:

• Give the model the source text and ask it to extract only what appears there, with an empty value when the field is not stated.
• Ask for the supporting quote or source location for each value, then check that the quote exists.
• Never let a model generate identifiers such as GTIN or MPN. Those come from the brand, the supplier or GS1.
• Store model output as inferred until a person or a rule confirms it.

Should you build or buy product data enrichment?

The right choice depends on catalog size, how many sources you use, and whether you have engineers to maintain a pipeline. Most teams combine options.

A common setup is a PIM as the system of record, supplier feeds and GS1 for identity, and a data API that returns structured product records from source pages for the fields that are still empty.

How can datawebot help with product data enrichment?

datawebot turns supported store URLs into structured product and variant records for people and AI agents. Today, store extraction works for Shopify stores. Each record keeps the source's own values, such as vendor, product type, variants, images and prices, together with the source URL and retrieval time, so you can map them into your catalog and review them as proposals. Missing values stay empty; datawebot does not generate missing facts.

There is an HTTP API with an OpenAPI spec and an llms.txt for agents, records export as CSV or JSON, and new accounts get 100 trial credits with no card required. See product catalog enrichment for how this fits into a review workflow.

Frequently asked questions

What is PIM enrichment?

PIM enrichment is product data enrichment done inside a product information management system. The PIM holds the master record, tracks completeness per channel, and routes new values through approval before they are published.

Can I copy attributes from a marketplace listing of the same product?

Only after you have confirmed it is the same product, ideally by GTIN or by brand plus manufacturer part number. Treat the listing as a seller's description and check key specs against the brand's own page.

Can AI generate a missing GTIN?

No. A GTIN is assigned to a company's product under GS1 standards, so it cannot be inferred from a title or image. Get it from the brand or supplier and check it with Verified by GS1. Google says to submit a GTIN only if you are sure it is correct.

Which fields should I enrich first?

Start with the fields that block feeds or search filters: brand, GTIN or MPN, color, size, variant grouping and category. Then move to specifications, extra images and compliance information.