← Back to all tipsNovember 10, 2026

Can a classifier actually build your product taxonomy

Sorting a catalogue into categories sounds like a one-time project — define the hierarchy, place every item once, done. In practice it's never one-time: new items arrive continuously, and a hierarchy that was correct on day one quietly drifts as nobody re-checks old placements against new categories. That's the actual case for automating it, more than raw speed.

What a classifier can do with what it's given

A classification model trained on titles, descriptions and existing attributes can place a new item into a known hierarchy without anyone opening a spreadsheet — the same underlying technique used for any multi-class labelling problem, just applied to "which category" instead of "which outcome." Give it enough correctly-labelled examples per category and it generalises reasonably well to items it hasn't seen, including ones described in slightly different words than anything in the training set.

Where unrestrained classification drifts

Left alone, a model will still make locally-reasonable calls that are wrong in aggregate — placing near-identical products in different categories because of a stray adjective, or inventing a category boundary nobody asked for. None of that shows up as an error; it shows up months later as a catalogue that looks inconsistent to anyone browsing it. Validation rules — a fixed category list, confidence thresholds below which a human reviews the call, consistency checks against similar items — are what keep automated classification inside the taxonomy you actually intended, rather than the one the model drifted toward.

before · unsorted
Earl Grey 100g tin Dark roast beans 250g Cold brew concentrate Green tea bags ×20 Ceramic drip filter
→
after · classified
Food & drink
Beverages
Tea 2 Coffee 2
Brewing equipment 1
Attributes and descriptions in, a placed, confidence-scored item out — the review step is what catches the rest.

Mapping onto a standard you don't control

Most catalogues eventually need to speak someone else's taxonomy too — Google's product categories for shopping ads, GS1 or UNSPSC codes for B2B trading partners, a marketplace's own browse structure. That's a second, separate classification problem: not "where does this live in our hierarchy" but "what's the correct external code for it," against a list you don't get to edit. Keeping one internal taxonomy and mapping it outward, rather than maintaining several parallel hierarchies by hand, is what actually scales as the number of external standards grows.

What still needs a person

None of this removes review — it moves where review happens, from "place every item" to "resolve the ones the model flagged as uncertain, and periodically audit a sample of the ones it didn't." That's a materially smaller job than classifying a catalogue by hand, and it's the part worth keeping a person on.