[object Object]

Your Knowledge Base has 8,000 articles, three top-level categories, and an average article-find rate of 12% from search. The taxonomy is wrong, everyone knows it, and nobody wants to rebuild because every existing URL, saved search, and Now Assist citation will break. Here is how to rebuild without that fallout.

Diagnose before redesigning

Pull the search log: kb_search_log if you have it enabled, otherwise the Now Assist or AI Search logs. Bucket the top 200 unsuccessful queries. Three patterns will emerge:

  1. Vocabulary mismatch (users say “wifi”, articles say “wireless network”)
  2. Granularity mismatch (one article should be three)
  3. Discoverability failure (article exists, no path to find it)

Each pattern has a different fix. Do not redesign the tree if the problem is vocabulary.

The dual-tree migration pattern

Do not delete the old taxonomy. Build the new one alongside it. Every article gets tagged with both old and new categories during transition.

Old: kb_category (single-select, deprecated)
New: u_kb_topic (multi-select, taxonomy v2)
Bridge: u_kb_legacy_category_map (old to new, many-to-many)

Search reads both for 90 days. Saved filters using the old field continue to work. New content uses the new tree.

Use the search log to validate

After 30 days on the new tree, re-run the unsuccessful query analysis. If the rate has not dropped at least 20%, the new tree is wrong. Iterate the tree, not the migration plan.

Synonym lists are the secret weapon

Most search misses are vocabulary. The platform supports synonym dictionaries — use them aggressively before changing structure.

wifi -> wireless network
laptop -> notebook
vpn -> remote access

A focused synonym list will lift findability more than any tree restructure.

Owners per topic, not per article

Taxonomy debt accumulates because nobody owns the categories. Assign every top-level topic to a named owner with a quarterly review cadence. The owner is responsible for stale-content sweeps and merge decisions.

Now Assist and AI Search rely on metadata

In the AI era, the taxonomy is RAG context. A clean, consistent topic field makes every Now Assist response better. The old fuzzy categorization will be cited in answers — and look amateur.

What to do this week

Pull the top 200 failed searches from the last 30 days. Categorize by vocabulary, granularity, or discoverability. The plurality wins your sprint. Do not start any work until you know which problem you are solving.

[object Object]
Share