Machine learning for product categorization: learned from your own catalog

There are three common ways to categorize a catalog, and they fail in different places.

A trained model needs no written definitions. It reads each product's name and description, and learns from your past decisions where your team draws each line. That matters most in the places rules and prompts handle badly: a deep tree where many categories sit close together.

StayCharted trained two models on the same 11,370 Wayfair products and then asked each one about 2,844 new products, using only the product name and the first 400 characters of the description, as they would arrive in a supplier feed. The 158 categories include Area Rugs, Bar Stools, Wall Sconces, Patio Tables and so on.

Always guessing the most common category would score 0.7%, which shows how many choices each product had.

How much a merchandiser still checks

Every answer carries a confidence. You pick a cutoff, and anything below it goes to your Review Queue.

At its default cutoff, the AI Classifier Model sends 15 products in 100 to a person, and it had three in four of those right anyway. The Dedicated AI Model sends fewer than 5 in 100. A higher cutoff means fewer mistakes go through, and more products go to a person. The trade-off is a setting you choose, and the report for your own model shows the same table on your own products.

What the mistakes look like

Almost all of the wrong answers, from both models, sit on a line Wayfair's own catalog draws two ways:

• Plant stands: some filed under Planter Accessories, others under Plant & Telephone Tables.
• Wall décor: some indoor, some outdoor.
• Holiday decorations: Ornaments & Tree Décor, or Holiday Accents & Decor.
• Tables: a patio table or a dining table.

Of the AI Classifier Model's 134 mistakes, the Dedicated AI Model got 83 right:

It also got 26 wrong that the AI Classifier Model had right, and both models got 51 wrong. These are the borderline calls in the catalog itself. Your team decides them once, fixes the past examples, retrains, and the model follows the decision.

• Many close categories. Vanities and Makeup Vanities, Wall Décor and Outdoor Wall Decor, Accent Chests and Armoires: a keyword list cannot separate these reliably, and a category name alone does not say where the line is. A trained model learns it from the examples on each side.
• Your own rules. Every catalog has conventions nobody wrote down, such as which department a corner plant stand belongs to. Past decisions carry those rules, so the model learns them without anyone writing a spec.
• A confidence you can act on. The cutoff table above is measured on products the model never saw, so you know before going live what share goes straight through and how often those are right.
• Consistency. A published model applies the same learned categories across imports; record its version and check performance as your catalog changes.
• Cost at volume. A trained model runs on each product as it is. You are not sending your full category list and instructions to a general AI model with every row of every feed.

This benchmark did not test a general AI model prompted with Wayfair's category names, so we make no claim here about how one would score on it. For that comparison on other public datasets, including the ones where the general model came out ahead, see assistant versus trained model.

The model learns whatever list your past products carry. That can be your own store tree, a marketplace taxonomy you map supplier products into, or a standard code set. For procurement catalogs, see UNSPSC classification. For retail and grocery product data, see GS1 GPC brick classification. For supplier spreadsheets headed into a catalog, see AI product categorization.

Deep trees also forgive some mistakes. When the Dedicated AI Model missed the exact category, it often still had the right department: 98.4% of products landed in the right department, against 97.3% in the exact category.

Training is not always worth it. Skip it when:

• You have a handful of broad categories that a supplier's own field or a short keyword list already settles.
• The stakes are low, and a wrong category costs little because someone browses and fixes it later.
• You have no categorized history yet. A model learns from past decisions, so start by categorizing a sample by hand.
• Your volume is small enough that one person can categorize the week's new products in an hour.

• AI Classifier Model. A ready-made AI language model reads each product's meaning and a private classifier learns your categories. It is on every plan, and Free includes one. On WANDS it scored 95.3%, and at its default cutoff 98.8% of the products it sent straight through were right.
• Dedicated AI Model. The AI language model itself is fine-tuned on your products, on a GPU. It is on the Business plan. On WANDS it scored 97.3%, and it was sure of more products.
• Classifier Model. Learns words and phrases, with no AI. It is on every plan. We did not measure it on this catalog.
• AI Image Classifier Model. For categories you can see better than read. On 40 categories of public fashion catalog photos it scored 91.1% (image model benchmark). It is part of the Vision add-on.

Plans and limits are on pricing, and the full feature list is on the product page.

1. Export products you've already categorized: name, description, and the category your team chose. The benchmark used up to 80 training products per category. About 20 well-chosen products per category is a reasonable starting point, with more for categories that look alike.
2. Upload them to StayCharted AI Model Trainer (AMT). AMT checks the data, and checks and masks personal information before training.
3. Train, then read the report. It is measured on products held back from training: accuracy overall and per category, the categories it confuses, and the cutoff table.
4. Fill new products. Use CSV, Excel or JSON files, the Google Sheets add-on, the API, or Snowflake and Databricks tables. Products below the cutoff go to the Review Queue, and your reviewers' answers train the next version.

Measured 23 October 2026 on StayCharted's production service. One run of each model, on one public catalog. Your catalog and categories will differ, and your own report measures them.

The Review Queue requires Essentials or Business. Retrain and publish a new version to use reviewer corrections in future answers.