Skip to main content
Extracting product attributes with AI in WISEPIM Attributes are the structured specs that power filters, comparison tables, feeds, and faceted search, material, dimensions, color, weight, compatibility, and more. The problem is they are usually buried inside description text or missing entirely. Attribute enrichment uses AI to read each product’s existing content and extract those specs into clean, structured attribute fields. What it enables: make your catalog filterable and feed-ready. Structured attributes drive on-site faceted search, marketplace requirements, and the comparison data shoppers use to choose, all extracted automatically from text you already have.
Attribute enrichment costs 1 credit per product. Extracted values are written to your product’s attribute fields, mapped to the attributes defined for that product’s family.

​
When to use it

  • Specs live in description text but not in structured fields, so shoppers can’t filter on them.
  • A marketplace or feed requires specific attributes (Google Shopping needs GTIN, color, size, material; Amazon has category-specific required fields).
  • Filters return incomplete results because many products are missing the attribute being filtered on.
  • As part of Auto-Fill, which extracts missing attributes using your default attribute prompt.

​
How to run it

1

Select products

Check products individually, or filter (for example, by a missing key attribute) and select all matching results.
2

Open the Enrich modal

Click Enrich with AI in the toolbar.
3

Choose Attributes

Pick Attributes as the enrichment type.
4

Select a prompt

Choose an attribute prompt from your Prompt Library. The prompt guides which attributes to look for and how to format values.
5

Start Enrichment

Click Start Enrichment and track it in the Process Tracker.

​
How extraction works

The AI reads each product’s title, description, and any existing data, then extracts values for the attributes defined on that product’s family. It only fills attributes that exist in your structure, it does not invent new attribute types. Values are normalized to match your conventions where the prompt asks for it (for example, “Red” rather than “red”, or units like “cm”).
Attribute extraction is only as good as the source text. If descriptions are thin, run Web Research first to pull in specs from external sources, then extract attributes from the enriched data.

​
What good looks like

  • Consistent value formatting. “Stainless Steel” everywhere, not a mix of “stainless”, “S/S”, and “inox”. Ask the prompt to normalize.
  • Units included and consistent. Dimensions and weights carry their unit, in the same system across the catalog.
  • No invented values. A blank is better than a guess. Tell the prompt to leave an attribute empty if the value isn’t present in the source.
  • High coverage on filterable attributes. The attributes shoppers filter on (size, color, material) should be populated on nearly every product.

​
Act on what you find

Low coverage usually means the source text doesn’t contain the specs. Run Web Research to fetch specs from external sources first, then re-extract. Outcome: attributes populated even where your own data was thin.
Add normalization rules to the prompt (“Capitalize color names; express all dimensions in cm”). Re-run. Outcome: clean, uniform values that group correctly in filters and feeds.
Confirm the attribute exists on the product family, then enrich. If the value genuinely isn’t determinable from content, Web Research or a supplier import may be needed. Outcome: feed-ready products that pass marketplace validation.
Review extracted attributes in the Data Quality tab, which scores attribute coverage and flags suspicious values. Outcome: confidence that extracted specs are accurate before they go live.

Web Research

Fetch specs from external sources when your own data is thin.

Data Quality

Track attribute coverage and catch suspicious values.

Enriching Products

Overview of all 14 enrichment types.

Prompt Library

Build and organize your attribute prompts.