Product tag hygiene: the invisible taxonomy that runs your search, filters, and recommendations

Count the mess before you fix it

Export every tag with its product count. Sort by similarity and read the list out loud. The duplicates announce themselves: singular and plural, hyphenated and not, the abbreviation one supplier uses, the campaign tag from a sale two summers ago still sitting on four hundred products. In most stores a fifth to a third of tags are variants of another tag, and a long tail of tags sits on exactly one product, which means they organise nothing.

Map each tag to the job it does: search synonym, filter value, collection rule, recommendation signal, or reporting category. Tags with no job get deleted, not archived. An archived tag in most platforms is a tag that will be back on a product by Christmas.

One concept, one tag, written in the rule book

Build the controlled list: the approved tag for each concept, its format, lowercase, hyphens, no abbreviations unless the abbreviation is the customer word, and an owner who approves additions. Migrate products to the approved forms with a bulk edit, then test the surfaces that consume tags: run your top searches and compare result counts before and after, click the filters, check the automated collections that build from tag rules. A migration that breaks a collection rule and empties a landing page is the failure mode to test for first.

Keep customer language in the synonyms, not in the tags. Shoppers search joggers and sweats and your tag says whatever your rule book says. The search tool's synonym mapping is the right home for that variety, one layer up, where it can be generous without polluting the data underneath.

Two rules keep it clean

Rule one: no tag is created during product upload. New products use the approved list, and a genuinely new concept goes to the owner for a same week decision. Rule two: imports map to the list or queue for review, using the same mapping table discipline as facet cleanup. Supplier tags arrive uninvited with every feed. A mapping table is the difference between a catalogue and a suggestion box.

Review the tag list quarterly with the product count export. It takes an hour. The quarter where nobody does it is the quarter the one off campaign tags and the new supplier's synonyms move back in, and they bring friends.

Search recall rises because the synonym layer maps onto consistent data. Filters agree with collections because both read the same controlled values. Recommendations start to look intentional, because similar now means something. None of this needs a new platform. It needs a list, an owner, and the refusal to let a thousand small exceptions reaccumulate, which is less a technical project than a house rule the team actually keeps.