Blog / AI & search /
What AI can help with in product taxonomy work
Say a catalogue has “travel mug”, “commuter cup”, “coffee tumbler” and a few misspellings of each. Someone asks an AI tool to tidy up the categories.
It comes back with a beautifully neat taxonomy. It also lumps different products together and splits one product family across three branches.
AI is good at the detective work in taxonomy: surfacing patterns, clustering candidate terms, spotting inconsistent labels and preparing a list for review. It can’t decide what a category means without business context, and it can’t answer for what customers see. Taxonomy is information architecture, merchandising and data governance, not just naming, and I wouldn’t hand it to a model unsupervised.
Define what the taxonomy needs to do
Start with the use cases. Is the taxonomy for storefront navigation, internal reporting, product feeds, search filters, inventory management or several of these? The same catalogue may need different classifications for different jobs. A category that works for a supplier report may not help a shopper choose.
Write down the rules for categories: what belongs, what doesn’t, whether branches overlap and how products get assigned. Decide whether a product can sit in several categories and how synonyms work. Without those rules, a model will infer a tidy structure from the text and ignore how the business really runs.
Name the decision owners. Merchandising may own product grouping, operations may need stable codes, and SEO may advise on search demand and crawl structure. Set those boundaries before asking AI for a hierarchy. Otherwise the output looks authoritative and answers to nobody.
Use AI to explore the catalogue
Give it a controlled sample with only the fields the task needs: product name, type, approved attributes and current category. Strip out personal, supplier-sensitive or confidential data you don’t need, and use only an approved tool. Catalogue exports can hold information that shouldn’t travel.
Ask it to group similar labels, surface spelling variants, find items with missing category data, or suggest questions for a human reviewer. Ask for the evidence behind each suggestion, such as the attributes or names that look alike. The aim is candidates for review, not a silent change to the production taxonomy.
Keep the existing hierarchy and the proposed one separate. Ask AI to flag uncertain matches instead of forcing every product into a branch. If two categories look close, it can explain the overlap and flag examples. That beats one confident list that hides the ambiguous cases.
Validate the clusters with product knowledge
Review samples from each proposed group. Compare the actual product details, not only names. Two products may share keywords but serve different purposes; a single product may be described with different terms across markets. I’d always have someone who understands the assortment check the boundary cases.
Look at the shopper’s language as well. On-site search, customer questions, sales conversations and Search Console queries can suggest how people describe products. These signals may include spelling variation or broad terms that do not map neatly to internal categories. A fluent local editor should review regional language rather than relying on literal equivalence.
Test the suggested taxonomy against common tasks. Can a shopper find a product by its purpose? Can they compare meaningful alternatives? Does the structure help staff maintain the catalogue? A hierarchy that looks logically symmetrical may still make navigation worse.
Keep category and product data distinct
AI may suggest a category based on a product title even when the underlying attributes are missing. Do not use a category label as proof of a fact. If an item is assigned to “waterproof bags”, verify that the product specification supports that claim. Classification does not replace data verification.
Use stable identifiers when mapping recommendations back to records. Product names change and may not be unique. Record the proposed category, confidence or uncertainty, evidence, reviewer and final decision. Avoid relying on a model-generated row order or title string as the update key.
If the taxonomy is multilingual, keep the underlying category identity distinct from each display label. Provide approved translations and local review. A translated term may be natural in one context and unsuitable in a storefront filter. Do not create separate taxonomy branches only because labels differ by language unless the product grouping itself differs.
Make changes in a controlled way
Before changing the live navigation, model the new structure and check the number of affected products, links, filters and feeds. Some taxonomy changes affect URLs or breadcrumbs; others only change labels. Understand which behaviour the platform will alter. Keep old-to-new mappings if category URLs change and plan relevant redirects.
Test a small section or staging copy. Check product assignment, collection pages, filters, internal links, breadcrumbs, structured data and sitemap output. Confirm that no important products become orphaned or appear in irrelevant collections. A taxonomy cleanup can change search visibility and customer journeys even when product descriptions remain the same.
Keep a human approval step for any change that affects product claims, availability or navigation. Use a reversible update method where possible and record the previous mapping. If an import tool supports dry runs, review the proposed changes before applying them.
Measure whether the structure helps
After launch, observe whether shoppers can find the intended products and whether internal search or support questions reveal new confusion. Search impressions and clicks may change, but they do not prove that the taxonomy alone caused the change. Consider URL redirects, seasonality, product availability and other page updates.
Track exceptions that required human decisions. If the model repeatedly misclassifies a product type, the taxonomy rules or source attributes may be unclear. Improve the source data or the decision guide rather than simply tightening a prompt until the model produces the expected labels.
Keep the human definition of the categories
AI makes a large catalogue easier to inspect, but the business owns the meaning and consequences of its taxonomy. Use the tool to make inconsistencies visible and reviewable, and keep the approved structure, product facts and change ownership in human hands.
Related reading
- How to verify product facts in AI-assisted ecommerce copy
- AI and SEO: AI and multilingual product descriptions
- AI and SEO: AI answers versus a store’s product pages
So choose one product group with inconsistent labels. Can AI surface the candidates, and can a product owner explain the rule that decides the final category? If not, write that rule before you import any suggested change.