How to write knowledge your chatbot can answer from correctly
By The AI Chatbot team4 min read
Chatbot knowledge is every document, web page and note your chatbot uses to answer customers. An AI chatbot only answers well from what you give it: short, clear, specific and current knowledge produces correct answers from day one.
This article sets out eight rules for writing knowledge, with before-and-after examples you can apply to your own store today.
Why knowledge decides answer quality
Modern chatbots search first and answer second: they find the section of knowledge most relevant to the question, then write the answer from it. If that section is vague, outdated or contradictory, the answer will be too. We explain this mechanism in what is an AI chatbot.
8 rules for writing chatbot knowledge
1. One topic per section
Keep returns, shipping and warranty in separate sections. When one section covers several topics, the chatbot can pick up the wrong detail.
2. Use specific numbers
Write "exchanges within 7 days of delivery" instead of "exchanges within a short time". Write "local delivery fee 20,000 VND" instead of "reasonable shipping fees".
3. State conditions and exceptions
"Items discounted more than 50% cannot be exchanged" stops the chatbot making wrong promises. Exceptions are what customers ask about most, so do not leave them out.
4. Give promotions an effective date
"10% off from 1 to 11 November" lets the chatbot know whether a deal is still running.
5. Remove old information instead of adding a new version
If there are two policy files, one old and one new, the chatbot may answer from the old one. Replace or edit the old source.
6. Write the way customers ask
If customers ask "can I pay COD?", use the word "COD" in your knowledge, not only "cash on delivery".
7. Say what your shop does not do
"We do not ship overseas" is a useful answer that saves the chatbot from handing off or guessing.
8. Add guidance for situations that need a person
"For orders over 20 items, please contact our staff for wholesale pricing" tells the chatbot when to hand off.
Before and after examples
| Before | After |
|---|---|
| We support returns if there is a problem with the product. | Exchanges within 7 days of delivery if the item is faulty or the wrong item was sent. We pay the exchange shipping. Items discounted more than 50% cannot be exchanged. |
| Fast nationwide delivery. | Ho Chi Minh City: 1-2 days, 20,000 VND. Other provinces: 3-5 days, 30,000 VND. Free shipping on orders from 500,000 VND. |
| 100% genuine products. | Products are sourced from the official distributor in Vietnam. Customers can ask staff to see the purchase invoice. |
The "after" versions are longer, but every sentence answers a real customer question.
How to add knowledge
AI Chatbot takes knowledge from several sources, each suited to a type of information:
| Source | Best for | Note |
|---|---|---|
| Website | Existing product and policy pages | Re-checked weekly or daily depending on plan |
| PDF, DOCX, TXT files | Policies, service price lists, guides | Scanned PDFs are still read |
| CSV files | Size charts, long price lists | Give every column a clear header |
| Written notes | Promotions, information that lives nowhere else | Quickest to edit |
| Product sync | Price and stock from WooCommerce | Updates by itself |
Keeping knowledge correct
- Websites are re-checked on a schedule, and product price and stock are re-checked the moment a customer asks.
- Changes to money or time terms in a policy, such as fees or return days, are held for your approval, so a website mistake does not become a wrong answer.
- Every version is kept, so you can restore an earlier one if needed.
- Review unanswered questions each week and add knowledge for the repeat ones.
Conclusion
A chatbot is never smarter than the knowledge you give it. These eight rules make every section of knowledge answer one real customer question. If you are just starting, follow how to create an AI chatbot in 30 minutes, then use this guide to refine your knowledge.
Frequently asked questions
Should I upload long files or split them into small sections?
Either works, but each file should focus on one topic and not mix in old information. One clear return policy file beats one file that collects everything from several years.
Do I need to write in question-and-answer form?
No. The chatbot understands normal text. Q&A form helps for questions customers phrase in a particular way, such as 'can I pay cash on delivery'.
What do I do when a policy changes?
Edit or replace the old knowledge source; do not add a new version next to the old one. For websites, changes to money or time terms in a policy are held for your approval, and every version is kept so you can restore it.
How do I know what knowledge is missing?
Check unanswered questions from the last 30 days, and the coverage check in the private preview. Both point to topics customers ask about that your knowledge does not cover.
