KB gap analysis

Find topics customers keep asking about that you haven't documented yet.

Your KB is a moving target. Customers ask new things; your docs lag behind. This feature reads recent tickets and chat openers and tells you what to write next.

How it works

KB → ✨ Find gaps → pick a window (30 / 90 / 180 days) → Run analysis.

The service:

  1. Pulls up to 200 recent ticket subjects + first visitor messages from chats, within the window.
  2. Pulls the titles of your published KB articles.
  3. Sends Claude both lists with a prompt: "cluster the questions, skip topics already covered by an existing title, suggest article titles for the gaps".
  4. Returns a JSON array of suggestions — each with a title, a one-line rationale, and an evidence_count (how many sample questions fit this cluster).

The result renders as a list ranked by evidence_count. The number "12" next to a title means roughly 12 different customer questions in the sample mapped to this gap.

Acting on a suggestion

Each row has a Write this article → button that opens the new-article form with the suggested title pre-filled. From there it's a normal article write — save as draft, iterate, publish.

When it shines

  • Workspaces with 30+ resolved tickets in the window — enough volume to cluster meaningfully.
  • Long-tail products with topic drift over time — the gaps reveal where your KB is stale.
  • After a feature launch — run a 30-day window to find what customers couldn't figure out.

When it doesn't

  • New workspaces with under 5 tickets. Returns an empty array.
  • Very narrow products with one or two FAQs everyone repeats. The "gaps" are just "you have no KB at all".
  • High-noise inboxes where every customer question is unique. Clusters won't form; suggestions will be vague.

Cost

A single run is ~5000 input + 800 output tokens. About $0.01 per run. Worth doing once a week or once a month, not once an hour.

Off-switch

Settings → AI → uncheck KB gap analysis. The Find Gaps page still loads but returns "no clear gaps detected" (the analyzer short-circuits).

What it isn't

  • It's not a writer. The suggestions are just titles with rationale; you write the actual article.
  • It's not a customer-facing tool. This view is for KB authors only.
  • It's not real-time. You trigger a run; it's not running every time a ticket comes in.