New — Sohay now answers WooCommerce product, policy & cart questions. See how →

Writing a good Knowledge Base article

What makes an article retrieve well and answer well — written for a reader who is also a chatbot's only source.

Every Knowledge Base article has two readers: the visitor who finds it on your site, and the chatbot that has to answer from it. The good news is that what helps one mostly helps the other.

Title it like the question

The single highest-leverage thing you can do. Titles are weighted heavily when matching a visitor’s question, so a title shaped like that question retrieves better.

  • Good: “How long does delivery take, and what does it cost?”
  • Weaker: “Shipping”

“Shipping” is a filing label. Nobody types it into a chat box.

Answer in the first paragraph

Put the actual answer at the top, then elaborate. Two reasons: a visitor who skims gets what they came for, and if the article is trimmed for length, the part that survives is the part that answers.

Resist the run-up. “At Acme we pride ourselves on fast delivery” is not an answer; “Orders ship within two business days” is.

Cover one topic, thoroughly

Prefer one thorough article per topic to three fragments.

Retrieval hands the model a small number of articles. A single article covering delivery times, costs, regions, and delays lets it answer a follow-up. Three short articles mean it might get the one about costs and not the one about delays, and answer half the question confidently.

Length is not the enemy. Padding is.

Write for someone who arrived cold

An article that only makes sense after reading another one retrieves badly, because retrieval will hand it over on its own. State the premise in a sentence, then answer.

Avoid “as mentioned above” and “see the previous article” — there is no above and no previous when the chatbot is holding it.

Use ordinary structure

Headings, short paragraphs, and lists all help. So do tables for anything genuinely tabular — delivery times by region, price tiers.

What does not survive is anything that carries meaning visually rather than in words. An image of a pricing table is invisible to the chatbot; the same table in text is not. If a screenshot is load-bearing, describe what it shows in the surrounding text.

Say what is not true

Explicit exclusions prevent confident wrong answers. “We do not ship to PO boxes” and “gift cards cannot be refunded” are among the most useful sentences you can write, because without them the model has nothing to contradict a hopeful assumption.

Keep it current

An out-of-date article is worse than a missing one — a missing article produces an honest “I don’t know”, while a stale one produces a confident wrong answer with your name on it. When a policy changes, the article is part of the change, not a follow-up task.

Editing a published article re-syncs it automatically.

Where to go next

Organising articles with categories and tags, and Article length and how content is processed for the limits.

Mithun B.
Mithun B.

More articles by Mithun B..

View all posts

Give your website an assistant that answers from your content

Grounded answers, a team inbox, and store-aware tools — free on WordPress.org.

Add Sohay — Free

Free plugin · You bring your OpenAI key · Works with any theme