Knowledge base health audit

A run-it-yourself audit of your knowledge base — the articles agents pull from and the ones customers hit in self-service. It tells you three things: where the KB is genuinely helping, where it's confidently wrong, and where it's quietly rotting and costing you tickets. Run it against a small, prioritized sample, not the whole library at once.

How to run it

• Don't audit everything. Pull your priority set: the top 20–30 articles by views over the last 90 days, plus one article per item on your top-contact-driver list. That's usually 80% of the value for 20% of the effort.
• Two kinds of checks below. Per-article checks you run on each article in the priority set. Library-level checks you run once across the whole KB. Both are marked.
• Time-box it. Budget ~2 minutes per article for the mechanical checks (links, dates, images) and more for the judgment calls (is this actually correct? is it findable?).
• Cadence. Priority set quarterly; a wall-to-wall pass once a year.
• Do it logged out. Test search and links as a customer sees them, not from your admin view.

The audit

1. Coverage of top contact drivers library-level

• You have a current, ranked list of your top 15–20 contact drivers from the last 90 days (from ticket tags, chat intents, or IVR reasons).
• Every top-20 driver maps to at least one published article. List the drivers with no article — that gap is your priority backlog.
• The top 10 drivers each have an article a customer could act on unaided — a real answer, not a "contact us" stub or an internal-only note.
• You pulled the "no results" search log and the top 10 zero-result queries each have a decision: write it, redirect it, or consciously accept it.
• New drivers from the last quarter (new features, price or policy changes, recent incidents) have articles, not just agent macros.

Score: __ / 5

2. Freshness & last-reviewed per-article

• Every article carries a visible "last reviewed" date — distinct from "last edited." A typo fix is not a review.
• Nothing in the priority set was last reviewed longer ago than your threshold. Suggested: 6 months for pricing / policy / security, 12 months for stable how-tos.
• No article references a deprecated feature, retired plan, old UI, or renamed setting.
• Steps and screenshots match the current product. Actually follow the steps for your top 5 — don't assume.
• Every article has a review cadence set, not "review when a customer complains."

Score: __ / 5

3. Findability & search terms per-article + library

• The title is what a customer would type or say, not internal jargon — "Card declined," not "Payment gateway 402 handling."
• The body includes the customer's words and the product's words — synonyms, verbatim error text, and old feature names — so search matches either.
• Searching the top 3 phrases for this topic returns this article in the first 3 results. Test it in your real search.
• The first paragraph answers the question. No throat-clearing, no "in this article we will" before the answer.
• The article is reachable by browsing (correct category and tags), not only by exact search.

Score: __ / 5

4. Duplicates & conflicts library-level

• No two published articles answer the same question. Where they do, pick one canonical article and merge or redirect the rest.
• No two articles give conflicting numbers or instructions — refund windows, SLAs, limits, eligibility. Reconcile to a single source of truth.
• The internal/agent version and the public version of a topic don't contradict each other.
• Near-duplicates (same topic, slightly different scenario) are either merged or clearly scoped in the title and cross-linked.
• Superseded articles are unpublished or redirected — not left live to compete in search with the current one.

Score: __ / 5

5. Broken links & assets per-article

• No broken internal links to other articles, categories, or app pages. Run a link checker — don't eyeball it.
• No broken external links, and no links to third-party docs that have since moved or changed.
• Images, GIFs, and videos load and are current — no missing images, no two-year-old screen recordings of the old UI.
• Anchor / deep links still point to the right section after edits.
• Embedded forms, widgets, and downloads still work and aren't gated behind a dead login or expired share link.

Score: __ / 5

6. Ownership & accountability per-article + library

• Every article has a named owner — a person or a specific team — not "Support" in the abstract.
• The owner can actually verify the content is right — usually the team closest to that feature or policy.
• Agents have a one-click way to flag a wrong or missing article, and flags get triaged — not dropped in a channel nobody reads.
• Orphaned articles (owner left, team dissolved) are reassigned, not silently unmaintained.
• The KB as a whole has an owner, and this audit has a name and a date on the calendar for next time.

Score: __ / 5

7. Self-service success signals library-level

• You can see, per article: views, plus at least one of — a "was this helpful?" signal, search click-through, or follow-on contact rate.
• For the top 10 articles you know the self-service success rate: viewed the article and did not open a ticket on the same topic within ~48 hours.
• High-traffic, low-helpfulness articles are on the rewrite list. High views + low "helpful" + high follow-on tickets = an article failing loudly.
• Any "deflection" claim is backed by a real signal, not raw article views. Views are not deflection.
• You close the loop: articles that resolve the issue get promoted; articles that generate contacts get fixed or killed.

Score: __ / 5

Scoring

Each ticked box is 1 point, so each dimension scores out of 5 and the KB scores out of 35. Read the dimensions individually before you read the total — a KB can be perfectly fresh and still un-findable.

Per dimension:

Overall (out of 35):

After the audit

• Turn every unticked box into a ticket, with the owner from dimension 6 attached. An audit that doesn't leave a backlog was a wasted afternoon.
• Fix in this order: wrong beats missing beats ugly. Conflicting or incorrect info first (it's actively hurting customers), then coverage gaps, then findability, then cosmetics.
• Re-run the priority set next quarter and compare. The trend matters more than the absolute number.
• Feed the gaps to whoever writes the content. Your top zero-result searches and your highest follow-on-contact articles are the content roadmap — you don't have to guess.