STAYCHARTED AMT / RETAIL AND E-COMMERCE

Ecommerce product categorization: your categories, learned from your own catalog

Ecommerce product categorization means putting each product into the right place in your store’s category tree, so it shows up in the right aisle, the right filters and the right feeds, such as Google Shopping. AI does it by learning from the products you’ve already categorized: each new product gets a suggested category and a confidence, and the unsure ones go to a person.

FROM YOUR EXAMPLES

Insulated stainless steel bottle, 750 ml, screw-top lid.
Drinkware

Measured on a real store catalog: StayCharted’s AI Classifier Model was right on 95.3% of 2,844 held-out Wayfair products across 158 categories. At its default cutoff, 85.2% qualified with 98.8% accuracy. Its Dedicated AI Model reached 97.3%; 95.5% qualified with 99.0% accuracy. Answers below each cutoff go to review. See how →

Where this fits in your work

Apply catalog categories

Suggest the approved category for new product descriptions.

Sort product photos

Use Vision to classify images from category folders or picture links.

Prepare supplier imports

Review consistent category suggestions before importing a supplier file into your catalog.

Where does a door mat go?

Suppose a new supplier row says “Natural coir door mat, rubber backing, 60 × 40 cm, indoor or covered outdoor use.” Your store has both “Home → Rugs & Mats → Door Mats” and “Garden → Outdoor Mats.”

Category to review
Home → Rugs & Mats → Door Mats
Confidence to check
If the model returns 78% and your review cutoff is 85%, the answer needs review.
Why a person checks
The description fits two merchandising categories. A reviewer checks the intended use and your store’s labeling rules, confirms the category and keeps that decision for the next training version.

Your catalog software applies the reviewed category; the model supplies a suggestion.

Measured on a real store catalog

We trained two models on the same 11,370 products: a StayCharted AI Classifier Model (Essentials and up) and a StayCharted Dedicated AI Model (Business and up), from Wayfair's public catalog (the WANDS dataset), every product already in one of 158 categories: Area Rugs, Bar Stools, Wall Sconces, Patio Tables and so on. Then we asked each about 2,844 products it had never seen, from the name and the first 400 characters of the description alone, as they'd arrive in a supplier feed.

StayChartedAI Model Trainer · model comparison
StayCharted AI Classifier Model Dedicated AI Model
Exact category (one of 158) 95.3% 97.3%
Right department (Furniture, Lighting, Outdoor…) 96.9% 98.4%
Wrong answers, of 2,844 134 77
Always guessing the most common category 0.7%

How much a person has to check. Every answer comes with a confidence. Pick a cutoff, and everything below it goes to your team:

AI Classifier Model

Cutoff Sent straight through Right among those
0.9 41.4% 99.7%
0.7 70.5% 99.5%
0.5 (its default) 85.2% 98.8%
0.3 93.4% 97.6%

Dedicated AI Model

Cutoff Sent straight through Right among those
0.9 (its default) 95.5% 99.0%
0.7 98.1% 98.3%
0.5 99.5% 97.7%

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 is surer more often: at its default it sends fewer than 5 in 100, and gets 99 in 100 of the rest right.

Real answers from the AI Classifier Model test:

Product (from the feed) Category the model chose Confidence
ripley twin daybed Daybeds & Guest Beds 1.00 Right
folkes 1-light wall sconce Wall Sconces 0.99 Right
dimmable led 79" under cabinet strip light Under Cabinet Lighting 0.93 Right
meinhardt 9-light shaded tiered chandelier Chandeliers 0.70 Right
schiller vanity Makeup Vanities 0.34 Right, but sent to a person: "vanity" could be bathroom furniture
isai rectangular corner plant stand Planter Accessories 0.78 Wayfair filed it under Plant & Telephone Tables
fuiloro pineapple garden plaque wall décor Wall Décor 0.82 Wayfair filed it under Outdoor Wall Decor

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. Three of them:

Product AI Classifier Model Dedicated AI Model
tuten rooster overdoor wall décor Wall Décor (0.41) Outdoor Wall Decor (0.97), right
deniece armoire Accent Chests / Cabinets (0.45) Armoires & Wardrobes (1.00), right
primm storage vanity set with mirror Vanities (0.36) Makeup Vanities (0.95), right

It also got 26 wrong that the AI Classifier Model had right, mostly at low confidence, so they'd go to a person. Both models got 51 products wrong.

Those aren't random errors. They're the borderline calls in the catalog itself, and exactly the ones your team should decide once and write down. Fix those, retrain, and the model follows your decision.

Measured 23 October 2026 on StayCharted's production service. One run of each model, on one public catalog; your catalog and categories will differ. The report for your own model shows the same numbers, measured on your own products.

Dataset and method

Source: Wayfair WANDS, Chen et al., ECIR 2022, under the MIT license. We selected categories with at least 50 products, capped each at 100, and used a fixed-hash 80/20 split within each category. We excluded 2,247 products with two class labels because this test asks for one category. The separate test file contained 2,844 products; 11,370 were used for training. These figures are from the separate test file, not the model’s internal validation score.

PREPARE YOUR EXAMPLES

Start with a file your team has already labeled.

Start with approved product descriptions and their categories. Keep SKU identifiers for your own tracking, but choose meaningful fields such as the description for the model to read. Train on the categories you want back; classification does not create a new catalog taxonomy for you.

Use one row per example with an approved answer. Keep the input representative of what the model will receive later, and agree on how to label ambiguous cases before training.

Product descriptionApproved category
Insulated stainless steel bottle, 750 mlDrinkware
Canvas shopper with reinforced handlesTote bags
Cotton crew-neck short-sleeve shirtT-shirts

From examples to reviewed results

1

Prepare

Resolve inconsistent labels and remove unnecessary sensitive information.

2

Train

Choose a model on your plan and train on representative examples.

3

Validate

Inspect held-out accuracy by category and the mistakes that remain.

4

Use and improve

Publish, fill a file or call the API, then review and feed corrections into the next version.

KEEP REVIEW IN THE WORKFLOW

Know where the model stops.

Check new product families, ambiguous category boundaries, and unusual image backgrounds. Keep SKU identifiers for tracking; classification does not manage stock levels, identify every exact SKU, or build a taxonomy automatically.

Evidence and implementation

Add pictures with Vision

Train a separate picture model from category folders or a spreadsheet of image links and labels. Use it to suggest categories for new arrivals or organize catalog photos. Check duplicates and mislabeled images before training. Vision reads the image, not text extracted from a photographed label.

Compare against relevant evidence

For product text: measured on Wayfair’s catalog, 95.3% right across 158 categories with StayCharted’s AI Classifier Model, and 97.3% with its Dedicated AI Model. For product photos: our image benchmark compares SigLIP 2 and DINOv2. Either way, test on your own products: the report measures each version on products it never trained on.

Wayfair catalog results →
Product photo benchmark →

Review before updating your catalog

Fill a file with predicted categories and confidence values, then review exceptions before importing it into your catalog system. For ongoing arrivals, use the API from your own software. Keep examples current as you add product lines or change category boundaries.

File fills return category suggestions for your existing import process. API calls return results to your own software; your integration decides how to apply them. Plan allowances and workspace rate limits still apply.

Common questions

Can text and pictures use the same categories?

You can define the same labels for separate text and picture models. Choose examples suited to each model; this workflow does not assume a combined text-and-image model.

Does it extract text from product packaging?

No. Vision classifies what a picture shows. Use an existing OCR or data-entry process if you need text extracted from packaging.

Can results go back into our store?

Download a filled file for your import process, or build an integration with the API. Review and apply the changes in your own catalog system.

Is this a service or software?

StayCharted is software. Your team trains it on reviewed examples, checks the results and reviews uncertain suggestions before applying them. It is not an outsourced catalog-classification service.

How accurate is AI product categorization?

On 2,844 held-out Wayfair products across 158 categories, StayCharted’s AI Classifier Model was 95.3% accurate overall, and its Dedicated AI Model was 97.3% accurate on the same test products. At their default cutoffs, accuracy was 98.8% on the 85.2% above the AI Classifier cutoff and 99.0% on the 95.5% above the Dedicated cutoff. Your results depend on your categories and product data; evaluate on your own products before relying on them.

How many products do I need per category?

This test used up to 80 training products per category. About 20 well-chosen examples per category can be a starting point, with more for similar categories. That is a starting guideline, not an accuracy guarantee; inspect your model’s report to identify gaps.

Can I map products to Google's or Amazon's product categories?

Yes, when your training examples carry the reviewed category labels from that taxonomy. New products receive a suggested category and confidence. Use the appropriate taxonomy version and maintain your labels when it changes; the model learns the categories represented in your examples.

Explore the details

Try it on your own examples.

Free includes text Classifier Models and one AI Classifier Model. Essentials adds API access, the Review Queue and starting without answers; Business adds Dedicated AI Models. Vision is an optional add-on on paid plans.