Google Product Category: How to Map Product Descriptions to a Category Taxonomy

Google's product category documentation is blunt: every product is automatically assigned a category, based on your title, description, brand and GTIN. The attribute overrides that guess, and Google accepts the override in three cases: categories with required attributes (apparel, mobile phones, software), campaigns organized by category, and alcohol.

So the usual advice to fill google_product_category on every SKU is half right. On a clean catalog, Google's classification is usually fine, and a hand-mapped category one level too broad is worse than a correct automatic one. We set it where it changes eligibility, policy or campaign structure. The rest of the effort goes into product_type.

`product_type` is your own breadcrumb, up to 750 characters, such as Home > Women > Dresses > Maxi Dresses. You can send up to five values. Only the first one organizes bidding and reporting, so send one detailed path.

The country detail matters. Outside those 11 markets (the US, the UK, Italy and the Netherlands among them), only product type can organize product groups. Google Ads lets you subdivide by product type up to five times, so the first five levels of your breadcrumb should be the ones you actually bid on. The Google Merchant Center guide covers the other identity attributes, including when you need a GTIN and what to do without one.

Map the tree, then the leftovers. A catalog with 400 leaf categories needs 400 decisions plus a few rules. That's a week, where classifying 20,000 SKUs one at a time is a quarter.

1. 1

   Export your category tree

   Pull every leaf node with its product count from your platform or PIM. Skip collections like New In.

2. 2

   Map each leaf to one Google ID

   A two-column lookup table: your node, Google's ID. Pick the deepest match and note the judgment calls.

3. 3

   Split the mixed nodes

   An Accessories node holds belts and bags alike. Add rules on title keywords or attributes (if title contains "belt", use 169).

4. 4

   Clean up product_type

   Replace "Sale" and "New In" with the real breadcrumb. In OpenAI's feed, product_type takes priority over Google's category.

5. 5

   Classify the leftovers from descriptions

   Products with no usable node go to an AI step that picks from a shortlist (below).

6. 6

   Push and compare

   Send both attributes through a supplemental data source keyed on id. Review only the products where Google's category disagrees with yours.

Mapping from the product description when the tree is no help

This is where generic AI tools go wrong. Ask a model for the Google product category and it returns a plausible path that doesn't exist in the file. We've seen feeds with hundreds of those.

What works is the setup researchers use. A dual-expert approach presented at an ACL workshop has a domain model propose the top candidate categories, then a general LLM compares them and picks one. In practice:

1. Build the shortlist from keywords in the title and description, matched against the taxonomy (10 to 20 candidates).
2. Give the model the product's text, its attributes and the shortlist with IDs. Ask for one ID from the list, plus a confidence level.
3. Reject any answer that isn't on the list. Send low-confidence products to a person.

DataWeave reported 92% accuracy across 1,240 fashion classes with a translation-style model. Every method still fails on a description that never names the product. "Our softest piece yet" can't be classified by anyone.

What a fixed row looks like

The Google category barely mattered here, because Shoes is a leaf. The title and product_type did the work. See product feed optimization for supplemental sources and rules.

Keep your mapping table in your PIM or feed tool. We fix the product data it depends on. The Optimizer rewrites titles and descriptions so they name the product type, and extracts attributes like material, fit or waterproofing that both classifiers and agents read. The free AI Readiness Audit scores Shopping Metadata alongside three other dimensions. Through integrations, AndromedAI imports from Shopify, Akeneo, SAP or XML feeds and publishes to Google Merchant Center, Shopify and Shopware, among others.