A product taxonomy is the agreed structure for classifying what you sell. It decides which categories exist, which products belong in them, and what attributes a product carries so that buyers, systems and partners can find and compare it.
It is also the one most often inherited rather than designed. Most catalogs carry a structure somebody built for a different range, a different channel or a different company, and the current shape is the residue of decisions nobody present was part of.
It is the most commercially exposed taxonomy most organizations own. A structural mistake in a document taxonomy makes material hard to find. A structural mistake in a product taxonomy loses orders, because a product nobody can filter to is a product nobody buys.
Three layers get conflated, and separating them is most of the work.
| Layer | What it holds | Who usually owns it |
|---|---|---|
| Category structure | The browsable tree a buyer navigates | Merchandising or commercial |
| Attributes | Size, material, voltage, certification, the facets buyers filter on | Product data or engineering |
| Identifiers | SKU, GTIN, manufacturer part number, and how they resolve to one product | Master data |
Treating these as one artifact is how product taxonomies become unmaintainable. The category tree moves when the market moves. Attributes grow as the range grows. Identifiers almost never change and must not. Different rates of change want different governance.
Because the business changes faster than the structure, and nobody is funded to keep up.
That last one is worth dwelling on. Marketplaces impose their own taxonomies, and mapping to them is mandatory. The failure is letting an external taxonomy become the internal one, because then every channel change becomes a restructuring project and no two channels agree.
The working pattern is one governed internal structure with explicit mappings out to each channel. The mapping is a published artifact that somebody owns, and it changes when the channel changes without touching the core.
The question that generates the most argument, and it has a reliable test.
If a buyer would browse to it, it is a category. If a buyer would filter by it, it is an attribute. Colour is an attribute. Footwear is a category. The test fails only when the business has a genuine browse path built on something you would expect to filter, which happens in regulated markets where buyers navigate by compliance class.
Hierarchical vs faceted taxonomy covers the general version of this choice, and the answer there holds here: most catalogs run a shallow tree for orientation with the real narrowing done by attributes.
Three things, and they all raise the cost of an ambiguous structure.
Agents resolve categories literally. An agent asked to find a compliant replacement part applies the classification as written. Where two categories overlap, it picks one and does not report the ambiguity.
Retrieval depends on attribute consistency. A filter is only as good as whether the same property is recorded the same way across the range. Three fields meaning one thing is a quiet failure that surfaces as poor results.
Answer engines quote your structure. When a model summarises what you sell, it works from how you have described it. Structure written in marketing categories instead of buyer language gets paraphrased into something you did not choose.
There is a fourth effect that gets missed. Agents do not ask clarifying questions about your catalog. A merchandiser who hits an ambiguous category pings someone and the ambiguity surfaces. An agent resolves it silently and moves on, so the structure stops generating the complaints that used to tell you it was wrong. The feedback loop that kept the taxonomy honest was human confusion, and automating the consumer removes it.
None of this is new work. It is the existing work, with the tolerance for ambiguity removed. Data governance for AI covers the wider frame.
A distributor with sixty thousand SKUs ran two category trees for four years, one inherited from an acquisition. Nobody merged them because the merge had no owner and no deadline.
The visible symptom was search. A buyer looking for a fitting found one of two near-identical categories and saw half the range. Conversion on those terms was poor, and the obvious diagnosis was a search problem, so the team bought a better search product. It did not help, because the structure underneath still split the range in two.
The merge took one person six weeks once it had an owner. The four years of delay were not a resourcing problem. Nobody could name whose decision it was, so the decision never got made.
How should it be governed?Separately per layer, because the layers move at different speeds.
Record the reason with every category change. Six months later somebody will ask why the range is shaped this way, and a recorded reason is the difference between a structure people trust and one they route around. Taxonomy governance covers what that trail needs to hold.
A product taxonomy is three artifacts wearing one name, and the governance fails when they are treated as one. Separate the browsable structure from the attributes from the identifiers, and let each change at its own speed under its own owner.
TopQuadrant's position is that the mapping layer is where the value concentrates. Channels will keep imposing their own taxonomies and you will keep mapping to them, so the thing worth governing is one internal structure that every mapping points back to. Build that in an open standard and the mappings survive a replatform. Skip it and every channel becomes a fork. Taxonomy management covers the discipline around it.