Evaluation · Ecommerce
Product data classification for ecommerce
In a large catalogue the hard part is not listing a product but attaching it to the right category and the right attributes. Zero-Shot builds classifiers that read titles, descriptions and images, and measures accuracy per category rather than on average, because errors always cluster in specific categories.
What we build
- 01A classifier that assigns categories from title, description and image
- 02Attribute extraction for colour, size, material and compatibility
- 03Matching records of the same product arriving from different suppliers
- 04Per-category accuracy reporting and a review queue for low-confidence records
- 05Mapping to each marketplace's own category tree
How it runs
01
Sample the catalogue
Samples are taken from every category and every supplier. Sampling only the largest category makes the measurement look better than it is.
02
Label with the catalogue team
Correct categories and attributes are labelled by your catalogue team. Cases two people label differently get resolved first.
03
Confidence threshold
The model attaches a confidence score to each prediction. Anything below the threshold is never written automatically — it goes to the review queue.
04
Shrink the queue
Errors from the review queue feed back into the set. The shrinking queue is the real return curve of the project.
What we measure
- Accuracy per category, as a breakdown rather than an average
- Share of records entering the review queue, and how it falls over time
- Attribute extraction accuracy for colour, size and material separately
- Returns and complaints caused by wrong categorisation
- Time for a new product to go live
When this is the wrong fit
If your category tree is undefined, or the same product legitimately sits in two categories, the model learns that confusion and multiplies it. Fix the tree first. And for a catalogue of a few hundred products, doing it by hand is cheaper than building a model.
Common questions
At what catalogue size does this pay off?
Roughly from ten thousand products up, manual work starts to be a real cost. It can be built below that, but the gain often does not cover the setup — and we say so up front.
Do you classify from images too?
Yes, and images make a serious difference when the text is weak. Using text and image together noticeably improves accuracy, especially in apparel and home decor. Which signal works better in which category is measured and reported.
How do we catch misclassification?
Three ways: low-confidence records already go to the queue, per-category accuracy reporting shows which categories are slipping, and returns and complaint data feed back in. A wrong category usually comes back as a return reason.
Other work in this sector
Other services
What we've written on this
- Where to start with enterprise AIHow to choose a first AI project: which tasks qualify, in what order to proceed, and what to expect in the first eighty days.
- Writing prompts for productionWhy a prompt that works in a demo breaks in production. Treating prompts like code: structure, versioning, boundaries and measuring every change.
- Why AI makes things up, and how to reduce itWhy language models state false things confidently. The real cause of hallucination, which countermeasures work, and which do not.
Last updated: 2026-09-05
Let's talk about this one
Tell us what you want to build and within two days we'll tell you who takes it and how we'd measure it.