Answers people and AI can both trust.

Start from replies that already worked

The best article is usually something an agent wrote well last Tuesday. Publishing should begin with that text, edited for a general reader, at the moment of resolution — not from a blank page during a documentation project that will not happen.

Write the question, not the topic

Title articles with what the customer actually asks: "Why was I charged twice?" rather than "Billing discrepancies". That is what search matches against, what the assistant matches against, and what a customer recognises.

Every article needs an owner who is not support

Support can draft anything. The refund policy belongs to whoever runs finance, and the troubleshooting guide to whoever builds the thing. An article support wrote and nobody confirmed is a liability that will eventually be quoted at a customer.

Verification is a date, not a status

"Verified" means a named person confirmed this was true on a specific date. Give each article a review period appropriate to it — pricing quarterly, troubleshooting annually — and let it lapse visibly when nobody re-checks it.

• Verified means a person, on a date
• Lapsed articles stop being used by the assistant
• The owner is warned before it lapses, not after
• Most-used articles get reviewed most often

Short articles stay true

A long article is one nobody updates and nobody finishes reading. One question per article, answered directly, is cheaper to re-verify and easier for both a person and an assistant to use correctly.

Find gaps by counting, not by guessing

Three sources between them give a precise picture: cases about a topic with no article, help centre searches that returned nothing, and questions the assistant declined. Rank the result by how much demand each gap generates, not by how easy it would be to write.

Measure whether the answer worked

After publishing, watch contact on that topic. Most knowledge bases never check, which is why they fill up with articles that are used often and followed by a second contact — confidently wrong or unclear, and invisible if you only count usage.

Delete things

A knowledge base grows monotonically unless something forces removal. Zero usage over a long period is that prompt. Fewer, better-maintained articles beat comprehensive coverage nobody trusts.