Turn Customer Questions Into a Better Chatbot Knowledge Base

To create a chatbot knowledge base from customer questions, collect about 50 recent questions from calls, inboxes and staff, strip out personal details, tag each by intent, and turn every repeated question into a short knowledge card with an approved answer, an owner and a review date. The result is better than training on marketing pages alone, because it reflects what customers actually ask.

Your best source isn't your website#

Most chatbot set-ups start with the website, which is sensible: it is public, approved and easy to import. But marketing pages are written to persuade, not to answer. The questions customers ask every week, such as "Can you come tonight?", "Do I need to be home?" or "Is the callout fee included?", are often missing or buried.

If you have already worked through our checklist of what to train your chatbot on, this is the next step: finding the answers that are not yet written down anywhere.

Step 1: collect 50 real questions#

Gather recent questions from:

• Call notes and voicemail summaries.
• Quote and enquiry emails.
• Website form submissions.
• Reception or office scripts.
• The questions apprentices and new staff keep asking.
• The answer the owner gives ten times a week.

Remove names, addresses, phone numbers and any sensitive details before this material goes anywhere near an AI tool. You want the pattern, not the customer record. The OAIC has advised that organisations should not enter personal information, particularly sensitive information, into publicly available generative AI tools (OAIC), so do the clean-up by hand or in a spreadsheet first.

Step 2: tag by intent#

Tag each question with one intent:

• Service: do you do this?
• Location: do you come to my area?
• Price: how much, and what is included?
• Availability: when can you come?
• Process: what happens next?
• Preparation: what should I do before you arrive?
• Policy: cancellations, warranties, payment.
• Safety: is this dangerous?
• Escalation: complaints and anything needing a person.

Counting the tags tells you where the effort should go. If a third of questions are about availability, that is where the knowledge base needs the clearest answers.

Step 3: compare with what's published#

Put the list next to the website:

• A repeated question with no approved answer is a content gap. Write the answer.
• A repeated question with three different staff answers is an operational gap. Agree the answer before anyone trains a chatbot on it.

The second kind is more common than most owners expect, and a chatbot will expose it quickly.

Step 4: write knowledge cards#

Turn each approved answer into a compact card:

• Customer wording: how people actually ask it.
• Canonical question: the clean version.
• Approved answer: short and specific.
• Source owner: who keeps it current.
• Last reviewed: a date.
• Exceptions: when the answer changes.
• Handoff trigger: when a person must take over.

For example, "Can you come tonight?" should not get a blanket yes. The card might state normal emergency hours, the areas covered, the details to collect and that attendance must be confirmed by a person.

Step 5: govern it#

Upload only material the business has the right to use. Keep prices, policies and service areas in sources with a named owner. Review volatile information, such as prices and availability, monthly, and stable information, such as services and licences, quarterly. When something changes, remove or replace the old source rather than leaving two versions for the chatbot to choose between.

Step 6: test with the original wording#

The customer wording column becomes your test set. Ask the chatbot each question the way a customer did, typos and all, and check the answer against the card. Where the answer is wrong, fix the card or the source, not just the prompt.

The payoff goes beyond the chatbot#

A question inventory improves much more than a chatbot:

• Service pages: add the answers customers cannot currently find.
• Staff scripts: everyone gives the same answer.
• Google Business Profile: keep hours and services consistent.
• Onboarding: new staff learn the real questions on day one.
• Quote follow-up: pre-empt the questions that stall a decision.

One real question becomes a reusable business asset.

Loading the cards into Pivra#

Pivra accepts websites, sitemaps, files (PDF, DOCX, TXT, Markdown and CSV up to 10 MB), YouTube transcripts and plain text as training sources; see training sources. Plain-text sources suit knowledge cards well: give each a descriptive title so workspace members know who owns it.

Website, sitemap and YouTube sources can be retrained automatically on a daily, weekly or monthly schedule. For prices and policies that need approval before they change, keep retraining manual. Keep your chatbot up to date explains the routine.

For answers that must be word for word, such as a licence number or a cancellation policy, use refinements. Improving responses covers Strict mode and testing in the dashboard before you publish. You can start building and testing on the free plan.

Next step: collect 50 anonymised questions this week and publish the ten answers customers cannot currently find, on your website first and in the chatbot second.