Automatic Product Tagging: Consistency at Catalogue Scale

A catalogue is a set of claims about what you sell. Search, filters and recommendation slots can only act on the claims that were actually written down.

If no field on a frying pan says it works on an induction hob, then for every shopper who filters on induction, that pan is not in your shop. The stock is on the shelf. The attribute is missing, so the sale is not available.

What manual tagging really costs

Manual tagging is not the wrong idea. It is a slow, expensive answer that degrades in a specific way.

It drifts between people. Two taggers looking at the same jacket will disagree about whether the colour is rust or terracotta, and both will be defensible. Multiply that across a team, then across staff turnover, and this year’s labels stop lining up with last year’s. Filters are comparisons, so inconsistent labels break them from the inside without ever looking broken.

It is slow enough to matter. A tagging pass measured in weeks cannot keep pace with a range that turns over weekly. A trained model runs against the catalogue on demand, in minutes.

It makes changing your mind expensive. Adding one new attribute across two hundred thousand products means commissioning the whole exercise again, so most retailers simply do not, and the taxonomy stays frozen at whatever seemed right two years ago.

Automation flips the last point in particular. Redefining the criteria and re-running across the entire catalogue costs close to nothing, which means the taxonomy can follow the business instead of constraining it.

Two readers on the same product

MLArt puts two models on every item, because products state their attributes in two different places.

An image model looks at the photography. A supplier feed might describe a lamp as “table lamp, gold”. The photograph shows a brushed brass base, a fluted glass shade and a fabric flex, none of which reached a field. In categories where the look is the product, the picture holds more usable attributes than the text does.

A language model reads the title, the description and whatever marketing copy the supplier sent. That is where the practical facts hide: hob compatibility, whether something folds flat, what fits inside it, who it is meant for.

The attribute nobody wrote down

The second reader does more than extract stated facts. It recovers attributes and categories that were never entered anywhere.

A stockpot listed with no capacity field, described as “big enough for a whole chicken”, belongs in your large-capacity band. A garden chair described as folding flat for winter storage carries a storage attribute that no field on it declares. A product whose category is missing or plainly wrong can be placed by what its words are about, not by the label someone typed in a hurry.

This works because the model compares meaning rather than matching strings. Description copy rarely shares vocabulary with your taxonomy. It does not have to.

Why the rest of the stack cares

Attribute coverage is not a data-hygiene project. It sets the ceiling on three things downstream.

Search and filtering. Every empty attribute is a facet that hides products from shoppers who are explicitly asking for them. Zero-result queries over stock you hold are the most avoidable loss in retail.

Recommendation retrieval. Ranking a million products inside a page load is impossible, so a cheap step first narrows the catalogue to a plausible shortlist, largely on attributes. A weak shortlist cannot be rescued by a good ranking model, as covered in predictive personalization.

Cold start. A product launched this morning has no sales history, but it has photographs and a description. Rich attributes let the engine place it next to items it resembles and estimate its performance on day one. In a 2022 fashion engagement, mapping product attributes and descriptions this way was what brought products with no sales history into recommendations at all, alongside a click-through lift of around 30%. The detail is in the case study.

The visual and language modules behind this sit on the technology page, and the wider product offer on solutions. To talk about your own catalogue, email office@mlart.biz.