Generate Google product categories with supplemental feeds

Google's product taxonomy contains 5,595 product categories, and assigning the right one to every product in your feed is one of those tasks that matters more than most merchants realize. Product categories help Google understand what you are selling, which affects how and where your products appear in organic Shopping results. This page explains how Productrise's supplemental feed generation for product categories works: a two-step system that maps your product titles into a semantic network of all Google product categories and then uses an AI reranker to pick the best match.

The problem: 5,595 categories and messy product titles

The full Google product taxonomy file lists every valid product category along with its numeric ID. With well over five thousand entries, manually matching each product in your feed to the correct category is not realistic. Building a spreadsheet-based mapping system is not much better: it is slow, error-prone, and breaks the moment your catalog changes.

The obvious shortcuts do not work either.

Why word matching fails

A naive approach is to match words from the product title against words in the category names. In practice, product titles are messy. They contain brand names, materials, colors, dimensions, and marketing fluff, and they rarely use the exact wording of Google's taxonomy. A title like "leather strap for guitars" never contains the literal category name "Guitar Straps" in that sequence, and a typo like "guitar strapp" breaks exact matching entirely. Word matching heuristics produce too many false positives and misses to be reliable.

Why a single large LLM call fails

Another tempting approach is to send the entire taxonomy, all 5,595 categories, to a large language model in one prompt and ask it to pick the right category for each product. This fails for three reasons: the token size of the full taxonomy makes every call expensive, processing thousands of products this way becomes prohibitively costly, and models hallucinate frequently when asked to select from such an enormous list, inventing categories that do not exist or returning wrong IDs.

How the category mapper engine works

Productrise takes a different approach. Every one of the 5,595 product categories is mapped as a point in a semantic network, where categories with similar meanings sit close together. You can explore this network yourself with the public product category mapper, and read more about the methodology in our post on matching product titles to Google product categories.

When you enter a product title, the engine places that title inside the same network and finds the categories closest to it in meaning. Because the model works with meaning rather than exact words, it handles noisy titles, different word orders, and even typos. "Leather guitar strap", "leather strap for guitars", and "guitar strapp" all resolve to the same correct category.

One note on the visualization: the 2D map you see on the page is a simplified projection of a multidimensional model. Two points may look close on screen while the actual closest match in the full model is a different one. The distances that matter are computed in the underlying model, not on the flat map.

The two-step approach: retrieval plus AI reranking

The mapper engine alone is not always enough. For some products the top result is a near miss. A yoga mat with a long, noisy title may land closest to "Yoga Mat Bags and Straps", with the correct category, "Yoga and Pilates Mats", in second place by a very small margin.

This is why the supplemental feed generation feature uses a two-step process:

1. Retrieval: the match engine maps the product title into the category network and retrieves the ten most closely aligned product categories.
2. Reranking: those ten candidates are passed to an AI model, which picks the single best category from the shortlist.

Because the reranker only has to choose between ten options instead of 5,595, it works at low cost, and the chance of hallucination is minimal: it can only select from real, verified categories that the engine already retrieved. In the yoga mat example, the AI correctly picked the number two candidate over the number one, which is exactly the kind of judgment call the reranking step exists for.

Generating product categories for your feed

Inside Productrise, the product feed audit checks whether your products have a Google product category at all, and whether the assigned one is plausible. If your existing category does not appear in the top ten matches for that product title, the product is flagged. You can filter your catalog down to exactly those products, for example all products with a missing category, and generate categories only for them.

When you start a generation job, Productrise uses the best available title for each product: if a supplemental feed title already exists, that one is used, otherwise the original feed title. Each title goes through the retrieval and reranking steps described above, and the resulting category is written to your supplemental feed as a google_product_category attribute with the official numeric taxonomy ID.

You can generate a sample first to verify the results, then run the job for all selected products. The output is a live supplemental feed URL that you can add to Google Merchant Center on top of your main feed. Products are matched by ID, and the supplemental values overrule the main feed, so you never have to touch your original product data. For more on this workflow, see our guide on optimizing organic Shopping with supplemental feeds.

Language support

The category mapper engine and the generation process currently work best with English product titles. Support for other languages is on the roadmap, but for now, English input produces the most accurate category matches.

Try it yourself

You can test the underlying technology with the free product category mapper: enter any product title and see which Google product categories it maps to. To generate product categories for your entire feed, add your feed to Productrise and start a supplemental feed generation for product categories.