Chatbot Knowledge Base Built From Questions Customers Ask
A bike shop I helped last spring connected its new website chatbot to a 40 page product catalogue and a returns policy written in 2019. Within a week the bot had told three customers that returns were free for 60 days.

A bike shop I helped last spring connected its new website chatbot to a 40 page product catalogue and a returns policy written in 2019. Within a week the bot had told three customers that returns were free for 60 days. The real window was 30 days, and the 60 day rule had been dropped two years earlier but never deleted from the document.
The bot was not broken. It was faithfully repeating what it had been given. Most chatbot problems start the same way, with a chatbot knowledge base that nobody wrote for a bot. Here is how to build one from the questions your customers already ask, in plain language, in about a day of work.
What a Chatbot Knowledge Base Really Is
Modern chatbots answer by searching a set of documents you provide and then writing a reply based on what they find. That set of documents is the knowledge base. It can be a folder of text files, a help centre, or a list of question and answer pairs pasted into the platform.
The bot can only be as accurate as that material. If two pages disagree, it may pick either. If an answer is buried in paragraph nine of a long policy, it may miss it. Writing the knowledge base well is less about technology and more about good, short, current writing.
Start With Questions People Really Ask
Do not begin with your website. Begin with your inbox. Pull the last three months of customer emails, chat logs and phone notes, and list every question that came in more than twice.
For a small business this usually produces 30 to 60 questions, and a handful dominate. At the bike shop, five questions covered nearly half of all contacts:
- How long does a repair take?
- Do you have this size in stock?
- What is your returns policy?
- Do I need to book a service?
- When are you open on public holidays?
Write those first. A bot that answers your top ten questions perfectly is far more useful than one that answers 200 questions vaguely.
How to Write Each Answer So a Bot Can Use It
Each entry should be one question and one answer, short enough to stand alone. I aim for 40 to 120 words per answer. Longer than that and the bot starts mixing details from different parts of the text.
A simple template
- Question: written the way a customer would type it, such as "can I return a helmet I wore once".
- Answer: the direct answer in the first sentence, then the conditions.
- Exceptions: one line covering what is not included.
- Next step: what the customer should do, such as bring the receipt or book online.
- Owner and date: who checked this and when, kept outside the text the bot reads if your platform allows it.
Use exact numbers. "Repairs usually take 2 to 4 working days" beats "repairs are quick". Name things the way customers do, so if people say "tune up" and your price list says "standard service", include both words.
Why Uploading Everything Is Bad Advice
Many platforms encourage you to upload your whole website, every PDF and the staff handbook, and let the bot sort it out. Vendors like this approach because setup takes ten minutes. In my experience it is the single biggest cause of wrong answers.
Old documents contradict new ones. Internal notes leak into public replies. Marketing pages full of "fast", "easy" and "best" give the bot nothing concrete. The bike shop's 60 day error came straight from an old PDF in an upload folder.
Be selective instead. Include only material you would be happy to read aloud to a customer today. Everything else stays out, even if it feels like the bot should "know" it. If you are still choosing a tool, check whether it lets you control sources this way, a point covered in choosing the right chatbot platform for a small business.
Keeping the Knowledge Base From Going Stale
A knowledge base decays quietly. Prices change, staff leave, opening hours shift for summer. Nobody remembers that the bot also needs to know.
| Task | How often | Time needed |
|---|---|---|
| Read ten random bot conversations | Weekly | 15 minutes |
| Add answers for new repeated questions | Fortnightly | 30 minutes |
| Check prices, hours and policies | Monthly | 30 minutes |
| Delete or merge overlapping entries | Every quarter | 1 hour |
Tie updates to events you already handle. When you change a price on the till, change it in the knowledge base the same day. One owner per topic works better than a shared responsibility that ends up belonging to no one.
Testing Answers Before Customers See Them
Before launch, ask the bot each of your top 30 questions in three different ways: the polite version, the short version and the annoyed version. "What's your returns policy", "returns?" and "I want my money back for this" should all reach the same correct answer.
Mark every reply as right, partly right or wrong. Partly right is the dangerous one, because it sounds confident and misses the key condition. Fix the source entry, not the bot settings, and test again. A fuller testing routine is laid out in how to test a website chatbot before it goes live.
Also decide which questions the bot should never answer alone, such as refunds over a certain amount or complaints about injuries. Write a short entry for each that tells the bot to pass the customer to a person.
Frequently Asked Questions
How many entries does a small business chatbot need?
Most small businesses do well with 30 to 80 entries. Start with the ten most common questions, launch, and add new entries as real conversations reveal gaps. Quantity matters far less than accuracy.
Should I write answers in a formal or casual tone?
Write the way your best staff member talks to customers. The bot will often borrow the tone of the source text, so stiff legal wording produces stiff replies. Keep exact terms where the policy demands it.
Can the bot learn new answers on its own?
Not reliably. Some tools suggest new entries from past chats, which can save time, but a person should approve each one. Letting a bot write its own facts is how wrong answers spread.
The Practical Takeaway
Block out one day. Pull three months of customer questions, write short single topic answers for the top 30, leave out every document you would not read aloud today, and test each answer three ways before going live. Then spend 15 minutes a week reading real conversations. That small habit keeps your chatbot knowledge base honest long after launch.
More in Chatbots
Chatbots
Chatbot Privacy and What Your Bot Should Not Store
I read fifty transcripts from a small physiotherapy clinic's website bot last spring, as a favour.
Chatbots
Chatbot Buttons vs Free Text and When to Use Each
A friend who runs a two person accountancy practice replaced her old button menu with a shiny new text box last spring, because...
Order Tracking Chatbot for a Small Online Shop
If you run a small online shop, a large share of your support inbox is probably one question asked fifty ways: where is my...