How to Train a Chatbot on Your Business FAQs — The Method That Actually Works

My first version of the chatbot was frustrating to use. Not broken — it replied quickly, it was polite, it understood the questions. But it kept escalating. "Let me check that for you, our team will reply shortly." Over and over, for questions I had answered a hundred times manually. The chatbot had all the right plumbing but the training document was too thin to be useful.

The FAQ document I had given it had maybe 25 entries. I had written what I thought guests asked. It turned out I had written what I was comfortable answering — not the full range of what guests actually wanted to know. The gap between those two things was where the chatbot kept failing.

The Real Threshold: Under 100 FAQs Is Still Frustrating

This is the number most guides skip because it sounds like a lot of work: a chatbot trained on fewer than 100 FAQs is not yet reliable enough to leave unsupervised on a business Messenger page. It will handle the most common questions — rates, availability, basic house rules. But real guest conversations go in directions you did not predict when you wrote the first 25 entries, and every unexpected question produces an escalation.

Once the FAQ document hit 100 entries, something changed. The escalation rate dropped significantly. The chatbot started handling full conversations — multi-message threads where a guest went from asking rates, to asking about parking, to asking whether they could bring extra guests, to asking how to send a deposit — without once needing to hand off to a human. The document had become dense enough to cover the realistic range of guest conversations.

Getting from 0 to 100 FAQs sounds like a week of work. It is not. It took one full day using a method I have not seen described anywhere else.

The Method: Ask Claude to Generate the Questions

Here is the approach that changed everything: instead of trying to think of all the questions guests might ask — which is limited by what you can recall under pressure — I asked Claude to generate them.

I gave Claude a brief description of my transient house: the location, the capacity, the basic setup, the type of guests we host. Then I asked: "Generate 200 questions that a potential guest might ask before booking a transient house like this." Claude produced the full list — organized by topic, covering angles I would not have thought of on my own: questions about early check-in, questions about what happens if there is a typhoon, questions about whether the bonfire needs extra payment, questions about corkage, questions about how to get there from the bus terminal.

My job was to answer each one. Not to edit the questions — to answer them, accurately, based on my actual property. Some answers were one sentence. Some were a short paragraph. The whole process took a full day: roughly six to seven hours of focused writing, with breaks.

What I ended up with was a document built from the perspective of the customer — questions phrased the way a guest would phrase them, not the way I naturally think about my own property. That phrasing difference is critical. A chatbot trained on "the bonfire is included in the weekend rate" answers the question "do you have bonfire?" but handles "bonfire ba included sa price?" much better when the FAQ is written in the guest's language from the start.

How to Structure the FAQ Document

Once you have your answers, structure matters. A flat list of 100 Q&A pairs is hard for a language model to navigate efficiently. Organizing by topic gives Claude a logical structure to work from and makes the document easier to maintain as you update it.

The section structure I use for a transient house:

• Rates — all rate configurations: weekday, weekend, holiday, per head, minimum headcount, extra person charges
• Availability and Booking — how to check dates, how to reserve, what deposit is required, cancellation policy
• Inclusions — exactly what is included at each rate: parking, bonfire, linen, kitchen, wifi, cooking equipment
• House Rules — no outside food rules, noise policy, curfew, maximum headcount, pet policy
• Getting There — directions from common starting points, parking instructions, landmarks
• Common Concerns — what happens if it rains, early check-in availability, late check-out, extra guests

Each section should be dense with specifics. Vague answers like "rates vary depending on the season" are not useful training data. Specific answers like "weekday rate (Monday to Thursday) is ₱5,500 for up to 10 persons, ₱250 per additional person up to 20 persons maximum" are what the chatbot needs to answer accurately.

The Three Meta-Instructions That Make or Break the Chatbot

Beyond the Q&A content, three instructions at the top of the document determine whether the chatbot is reliable or dangerous.

The Escalation Rule

This is the most important line in the entire document: if you are not completely certain of the answer based on the information in this document, do not guess. Tell the guest: "Let me check that with our team — we will reply within a few minutes." This single instruction is the difference between a chatbot that builds trust and one that erodes it. A guest who receives an honest "I will check" and then gets a correct human reply has a better experience than a guest who receives a confident wrong answer.

The Persona Instruction

Define who the chatbot is before the FAQ content begins. Not a long character description — a clear, functional identity: "You are the booking assistant for [Property Name]. You are warm, direct, and helpful. You do not discuss topics unrelated to the property." The persona sets the tone for every response and prevents the chatbot from wandering into conversations that have nothing to do with bookings.

The Language Instruction

For a Philippine business, this instruction is essential: "Respond in the same language the guest uses. If they write in English, reply in English. If they write in Filipino or Tagalog, reply in Filipino. If they mix both — Taglish — match their mix." Without this instruction, Claude defaults to English even when guests write in Tagalog, which feels impersonal and transactional. With it, the chatbot naturally mirrors the guest's communication style.

How to Test the Chatbot — The Right Way

The most common mistake after building the FAQ document is testing it yourself. You know your property. You phrase questions the way someone who already knows the answers would phrase them. Your tests will pass because you are not simulating the experience of a stranger.

The correct way to test is to ask someone who has never seen your property — a friend, a family member, ideally someone who has never even heard of it — to have a real conversation with the chatbot as if they were a potential guest. Give them no coaching. Let them ask whatever occurs to them naturally.

Every escalation that happens during this test is a gap in your FAQ document. Every wrong answer is a specific entry that needs to be corrected or added. Run this test before you go live, and run it again whenever you make significant changes to rates or policies. The chatbot is only as accurate as the last time you updated the document it was trained on.

Maintenance: 30 Minutes Per Month

A FAQ document is not a one-time project. Rates change. Policies change. You add a new room configuration. You change the check-in time. Every change in your business that guests would ask about needs to be reflected in the FAQ document — otherwise the chatbot confidently gives guests outdated information.

Once the initial document is built and working, maintenance takes about 30 minutes per month. Review any escalations from the previous month, identify the questions the chatbot could not answer, add those entries. Update any information that has changed. Re-test the affected sections. The document compounds over time: the longer you maintain it, the fewer gaps remain, and the lower the escalation rate.

After more than a year of monthly updates, the transient house chatbot handles over 90% of all Messenger inquiries without any human intervention. The FAQ document now has well over 200 entries across all sections. That depth is what makes the difference between a chatbot guests trust and one they work around.