Skip to main content
Back to selected work
E-commerce · Information architecture · GCC market

Avnzor — Product Taxonomy Rebuild

A GCC health & beauty retailer whose product structure had stopped matching its business — and the intake system that keeps the fix from decaying.

Avnzor — Product Taxonomy Rebuild
Outcome

Avnzor's category tree still described a beauty store while the revenue had moved to nutrition — and with roughly one visitor in twenty using search, almost all product discovery ran through it. A five-pass audit set the catalog against the business and produced the constraints for a tree capped at three levels. But a one-off cleanup would not hold: a single import re-broke the catalog weeks later. So what shipped was an intake path — a daily tripwire, a classifier that proposes a category, and a queue where merchandising approves. Categorisation stopped being a project and became a queue somebody works.

Finding 01 — the catalog ranked by product count, then re-ranked by revenue per product, inverting completely

Every storefront inherits its taxonomy from whatever the business was when the catalog was first loaded. Avnzor’s had never been revisited — so it still described a beauty store, while the revenue had quietly moved to nutrition.

The Constraint

The category tree was the least maintained system in the store and the one carrying the most weight. Roughly one visitor in twenty used search; the rest browsed, which meant almost all product discovery ran through a menu nobody owned.

Underneath it sat more than a thousand product types nested several levels deep. 589 of those types held three products or fewer — shelves with nothing on them. Duplicate labels split single concepts across multiple branches, so the same idea lived in several places at once and none of them completely.

Finding 02 — before and after taxonomy structure

The System I Introduced

Rather than redraw the tree from intuition, I measured the catalog against the business in five passes — shelf space against revenue, structure against depth, search demand against actual sales, query language against index language, and browse against search. Each pass produced one finding, and each finding constrained the shape of the new tree.

Shelf space ran opposite to earning power. There were 17× more Makeup SKUs than Sports Nutrition SKUs, and each Sports Nutrition SKU earned 80× what a Makeup SKU did.

Evidence That Shaped the System

The restructure caps depth at three and organizes around how the catalog actually earns: Sports Nutrition and Supplements get real branch structure — Protein split by whey, isolate and plant; Performance split by creatine, pre-workout and amino — while the long tail of near-empty product types stops being navigation and becomes filtering.

That gap is the whole argument in one chart. Demand and revenue were being generated by two different halves of the store, and the taxonomy was organized around the half that searched rather than the half that bought.

اوردیناری returned nothing while The Ordinary sat in the catalog with dozens of products. میکب returned nothing against 4,350 makeup products. اشوقندا returned nothing against a stocked Herbals shelf. The fix is a synonym layer between query and index — no catalog rewrite required, and no dependency on the taxonomy work landing first.

Finding 03 — most searched brands versus best selling brands, with no overlap
Finding 04 — Arabic and Persian queries returning zero results for products the catalog already held
Finding 05 — 72% of store sessions were browsing rather than searching
Taxonomy audit — collection depth and navigation finding

What Shipped

The audit set the order of work: cap the tree at three levels around what actually earns, retire the empty product types into filters, and ship the intake and synonym layers first so new imports stop re-breaking the store while legacy collections move over.

The intake side is now live in admin: a daily check flags uncategorised supplier imports, proposes a placement, and puts it in a review queue where merchandising approves in one click. Alongside it, a weekly job reads zero-result Arabic and transliterated searches (اورديناري, اشوقندا, ميكب) and writes the missing synonyms straight into Algolia.

Product types → max tree depth
1,000+ → 3
Catalog reachable after restructure
~95%
Product types holding 3 products or fewer
589
Revenue per product · Sports Nutrition vs Makeup
80×
Top-searched brands that were also top sellers
0 of 5
Zero-result synonym updates
Weekly

What’s Next

Finishing the remaining collection URL moves, then running the same five passes again once the dust settles — the browse/search split is still the number to watch, since a taxonomy that works should pull discovery back into the menu rather than forcing people into the search bar.

Want to build something like this?

Let's talk about your project.

Next case study

My Stack