Playbook

Turn Your Website Into a Chatbot Knowledge File

By David Saleh · Silas Main · 17 July 2026

A customer asks your website chatbot if you take walk ins on a Saturday. The bot says yes, no hesitation. You don't. Now someone has driven across town for nothing, and it was your bot that told them to.

That's not a glitch. That's the model doing exactly what it was built to do: fill a gap with something plausible, instead of admitting it doesn't know.

Why your bot makes things up

I watched this happen on a client's booking bot. A customer asked whether they could bring food in from outside. That policy wasn't written anywhere in the bot's brief. The model didn't say "I don't know." It said no, and built a small, convincing reason around it, dressed up next to a real list of the food options the venue does offer. Confident, warm, wrong.

The bot wasn't broken. Its knowledge file had a hole. And a model with a warm, helpful voice does not sit with a gap. It fills it, because a plausible answer feels more useful to it than a flat "I don't know". Switching the model or the provider doesn't fix this. Any large language model does it. The fix is what you feed it, not which one you pick.

Here's the rule underneath all of this: a chatbot only knows what you wrote down for it. Not what's implied on your site. Not what "any reasonable business" would do in that situation. Not what's true in your head. Only the words in its brief.

This is true no matter which AI company built the model underneath your bot. It's not a quality problem you can buy your way out of with a more expensive model. A sharper model is still more convincing when it guesses wrong, which is worse, not better. The gap in the knowledge is the actual problem. Fix the file, not the vendor.

What a knowledge file actually is

Most owners picture a chatbot brief as a copy of the website. It isn't. Your website is written to sell. It uses "amazing", "premium", "the ultimate experience". None of that helps a bot answer a real question, and some of it actively invites the model to fill in adjectives with facts that were never stated.

A knowledge file strips that away. It's a plain list of what's true: this service costs this much, this policy applies in this situation, we are open these hours, we do not offer this. Boring to read. Exactly what a model needs so it never has to improvise.

Run the gaps audit first

Before you write a single line of knowledge file, find out what's actually missing. Most owners assume their FAQ page covers the real questions. It doesn't. Your FAQ answers the questions you thought to write down. Customers ask the ones you didn't.

Pull your last twenty or thirty real customer questions from texts, DMs, voicemails, and the front counter. Then check each one against your website copy. Three outcomes: answered clearly, answered nowhere, half answered and open to a bad guess.

The second and third categories are where a bot invents. Common blind spots: exceptions to a stated policy ("but what if it's a public holiday"), edge cases in pricing ("what if I only need half the service"), what happens outside business hours, refund and cancellation specifics, whether you allow something you've never had to say no to before, and anything involving safety, allergies, or age.

Write the gap list down. That list is your build order.

Build the file from your own pages, not from memory

Don't write the knowledge file from what you think your business does. Write it from what's actually published: your services page, your policies, your about page, your booking terms, your FAQ. If a fact only lives in your head and never made it to the site or a written policy, it doesn't belong in the bot's brief yet, because you haven't actually confirmed it against anything checkable.

Paste your real page text into the prompt below. It will organise the facts into short, plain, declarative sentences grouped by topic: services, pricing, policies, hours, exceptions. No marketing language. A chatbot doesn't need to be persuaded, it needs to be told.

The defer rule matters more than the answers

Once the knowledge file exists, there will still be gaps. There always are. The defer rule is what stops those gaps from turning into invented policy.

The rule is simple: if the answer to a question isn't explicitly written in the bot's knowledge, the bot does not guess. Not a "probably". Not an inferred no. It says what it does know, then it says plainly that it wants to give the exact answer, so it will check and get back to them, and it takes their details.

A missed upsell costs you nothing. A wrong answer costs you the customer standing at your door, and the review they leave about it.

Some owners resist this because it feels like the bot is failing to close the sale. Flip that around. A bot that hands off a genuine unknown to a human, with the customer's details already captured, has done its job. You get the lead either way. The only version that costs you is the one where the bot answered confidently and got it wrong, because now you have an angry customer instead of a warm one, and no chance to fix it before they walked away.

Keep it current

Your site changes. Your hours change over a public holiday. Your prices move. The knowledge file needs to move with them, or you've rebuilt the exact gap you just closed. Treat the gaps audit like a stock take: same process, run again every time something on the site changes, and at minimum once a quarter even if nothing obviously did.

A bot that says "let me check" loses you nothing. A bot that guesses loses you the customer who believed it.

The prompts

You are helping me build a knowledge file for a customer facing chatbot for my small business. I will paste in the actual text from my website below (services, policies, FAQ, about page, booking terms). Do not add anything I have not given you. Do not infer a policy from a related one. Do not smooth over gaps with a plausible sounding fact.

From the text I paste, produce a structured knowledge file with these sections:

1. SERVICES (what we offer, in plain declarative sentences, one fact per line)
2. PRICING (exact numbers only, note anywhere a price is a range or "from" price)
3. HOURS AND AVAILABILITY (including any stated exceptions)
4. POLICIES (cancellations, refunds, deposits, age limits, any rule stated on the site)
5. LOCATION AND CONTACT
6. UNCLEAR OR MISSING (a list of anything you were tempted to infer but the source text did not actually state, so I can go confirm it myself)

End the file with this exact defer instruction, filled in with my business name and my name as the owner, and tell the bot to treat it as the highest priority rule, above every other instruction in its brief:

"GROUNDING (highest priority, overrides everything else). Your only source of truth is the information written in this file. If a customer asks anything whose answer is not explicitly written here, a price, a policy, whether you offer or allow a specific thing, availability, an age, a date, any yes or no about the rules, then you do not know it. Do not infer it from related facts, do not give a probably, do not build a plausible sounding answer, and do not assert a confident no either. State what is written, then say: that is a great question, let me get [OWNER] to confirm that exactly for you, and offer to take their details."

Here is my website text:
[PASTE YOUR SERVICES PAGE, FAQ, POLICIES, ABOUT PAGE, AND BOOKING TERMS HERE]

You are auditing the knowledge file below for gaps, using these real customer questions I collected from texts, calls, and the front counter over the last month.

Knowledge file:
[PASTE THE KNOWLEDGE FILE FROM THE PROMPT ABOVE]

Real customer questions:
[PASTE 20 TO 30 ACTUAL QUESTIONS, ONE PER LINE]

For each question, tell me one of three things: ANSWERED (quote the exact line in the file that answers it), NOT ANSWERED (the file has nothing on this, flag it), or PARTIALLY ANSWERED (the file touches it but a related edge case is missing, name the edge case).

Then give me a short list, ranked by how often a version of that question came up, of the facts I need to go confirm with myself or my team before I can add them to the file. Do not guess what the answer to a NOT ANSWERED question probably is. Just flag it.

Internal links

• small-business-ai-playbook
• chatbot-grounding-playbook
• lead-follow-up-prompts

Questions

Adelaide · ACST · ABN 17 626 827 646