Most taxonomies are not badly designed. They are operationally dead: the schema says one thing, the tagged content says another, and the editorial team has been working around the system for two years.
An audit starts with the data, not the documentation — what is actually in the database, how term usage is distributed, which branches are empty and which have ballooned. Only against that picture is it worth discussing whether to repair the structure or rebuild it.
What gets inspected
- Quantitative distribution and dead zones. I export the actual tagging records from your CMS or database and measure the distribution: how many categories account for 80% of assignments, which leaves sit at zero, and what is hiding inside the residual bucket (“other”, “general”) or the first alphabetical option in a mandatory dropdown.
- Logical division errors in the schema. I look for places where sibling nodes at the same level answer different questions — mixing what an object is with its format, audience or market frequency. Wherever two independent axes are glued into a single select list, taggers are forced to guess.
- Inter-tagger (or inter-model) agreement. Using a sample of borderline records, I test whether two people — or two independent model runs — assign the same item to the same leaf using the written rules alone, and where unwritten team folklore is compensating for missing scope notes.
- Identifier and layer hygiene. I separate genuine domain concepts from campaign tags, workflow flags and navigation labels that have accumulated inside the core vocabulary over time.
When to repair and when to rebuild
A clean-slate rebuild is expensive: it breaks historical reporting continuity, requires re-tagging the archive and forces URL migrations. If the primary axis of division is sound and the failure comes from fuzzy borders between neighbours, uncontrolled synonyms and missing scope notes, targeted surgery on the existing tree is enough — merging duplicates into altLabel aliases, writing explicit yes/no boundary rules, and extracting attributes into a separate facet.
Rebuilding from scratch is warranted only when the audit shows that the trunk itself tangles independent dimensions and no amount of definition tuning can make the branches mutually exclusive.
Scope and timeline
Turnaround depends on three variables: vocabulary size (hundreds vs. thousands of concepts), how quickly a clean export of tagged records can be pulled, and how many teams or systems share the schema. Most audits — from data export and tagger interviews to the findings report and closing decision workshop — run between one and three weeks.