AI and automation
You train a website chatbot by giving it your own pages, documents and written answers, then testing it with real questions and fixing the gaps. The work is mostly gathering clear, current information, not anything technical.
By Look North Digital · Published
When chatbot tools talk about training, they usually mean feeding the chatbot your business information. You are not teaching the AI to read and write. It can already do that. You are giving it the facts it should use.
Most modern tools work through a method called RAG (retrieval-augmented generation). Amazon Web Services describes it as having the AI check an authoritative knowledge base outside its training data before it responds. In plain terms, when a visitor asks a question, the chatbot searches your information, finds the relevant parts and writes an answer from them.
This is good news for a small business. AWS points out that it is a more cost effective way to add new information than retraining an AI model, and that you can change the sources whenever you need to. If your prices change, you update the information, not the AI.
Start with what customers actually ask. Look through your emails, phone notes, social media messages and reviews, and list the questions that come up most often.
Then gather the material that answers them. Most chatbot tools accept a few common types of content.
| Content type | Best for | Watch out for |
|---|---|---|
| Your website pages | Services, about us, areas covered, general information | Old pages with out of date prices or offers |
| Uploaded documents (PDF or Word) | Detailed guides, terms, treatment or product information | Scanned documents with no selectable text |
| Question and answer pairs | Anything that must be exact: prices, policies, hours | Forgetting to update them when things change |
| Pasted text or short notes | Seasonal information, temporary changes, new services | Leaving them in after they stop being true |
Write for the chatbot the way you would write for a new member of staff. Use plain words, short paragraphs and one topic per section. Put the answer first, then the detail.
Vendor guidance gives some practical rules. Intercom says each uploaded document should have a unique name with unique content. It also says its AI can only use text you can select in a PDF, can't read images, and doesn't support multi-column layouts. Chatbase says scanned PDFs with no text layer can't be used. So if your price list is a picture or a scan, type it out.
For answers that must be exactly right, use question and answer pairs. Chatbase recommends them for questions that need a precise, consistent answer, such as pricing and policies. It also suggests adding several ways of asking the same question to one answer, for example "How much is it?", "What do you charge?" and "Price?".
Remove contradictions before you start. Search your website and documents for every place a price, opening time or policy appears, and make them all match. A chatbot can't tell which version is the current one.
Most tools can read your website for you. You enter your home page address or your sitemap (a file that lists every page on your site), and the tool fetches the pages. Chatbase says that from a home page it discovers every public page, and lets you review the links before training.
Be choosy. Leave out pages that could confuse it, such as old blog posts about past offers, job adverts or test pages. Intercom and Chatbase both let you exclude pages by their web address.
Remember that changes to your site won't always show up straight away. Intercom says its website syncs run each week, and that you can run a manual sync when you need to. Chatbase re-reads websites weekly on some plans and offers a resync button on others. After you update a price on your site, resync the chatbot too.
Facts are only half the job. The chatbot also needs instructions on how to behave. Most tools have a settings box for this, sometimes called instructions, guidance or a system prompt.
Anthropic, the company behind the Claude AI model, advises telling the AI it is allowed to say "I don't know" and to use only information from the documents provided. Both rules reduce made up answers. Write them into your chatbot's instructions in plain words.
GOV.UK guidance says you should test the accuracy of a bot's answers before it goes live. Do this yourself and ask a couple of people who don't know the business well to try it too.
Test the awkward cases as well as the easy ones. You want to know what happens when it doesn't have the answer, not just when it does.
Training is not a one-off job. GOV.UK warns that a bot's accuracy might drop as it becomes exposed to requests beyond its scope. Real visitors will ask things you never thought of.
Read the conversation log every week for the first month or two, then monthly. Look for questions it couldn't answer, answers that were wrong, and places where people gave up. Each one is a gap to fill.
Tie chatbot updates to your normal routines. When you change prices, update your opening hours for Christmas or launch a new service, update the chatbot on the same day. Put a reminder in your calendar to review it every quarter.
| When | What to do |
|---|---|
| Every week at first | Read the conversation log and fix wrong or missing answers |
| Whenever something changes | Update prices, hours, services or policies, and resync your website |
| Before holidays and busy seasons | Add temporary notes about closures or availability, then remove them after |
| Every quarter | Re-run your test questions and remove out of date content |
Most problems come from the information, not the AI. These are the ones that catch people out most often.
Our Website Chatbot is taught only your business information: your pages, documents or pasted text. It says it doesn't know rather than inventing an answer, captures name, email and the enquiry mid-conversation, hands off to a person when needed and logs every conversation, so you can see what to improve. It works on any website with one line of code, and the price is on application.
If you would rather learn to work with AI tools yourself, our AI training sessions are built on your own work, and you keep the prompt library. Tell us about your project through our quote form.
No. Most of the work is gathering and tidying your business information, and writing clear answers to common questions. Adding your website or documents is usually a matter of pasting an address or uploading a file.
Start with enough to answer your 20 to 30 most common questions well. It is better to cover fewer topics accurately than many topics badly. Add more as you see what visitors actually ask.
Most tools accept them. Chatbase, for example, accepts PDF, TXT, DOC and DOCX files. Make sure the text can be selected, because scanned documents with no text layer can't be read.
Usually not in the way people imagine. Most business chatbots answer from the information you give them, and improve when you add or fix that information. Reading the conversation log shows you what to add.
Whenever your prices, hours, services or policies change, and at least every quarter. Some tools re-read your website weekly, but it is safer to resync after any important change.
Fix the website first, or exclude those pages from the chatbot. The chatbot will repeat whatever it is given. Tidying your site also helps visitors who never open the chat.
Yes. Most tools let you add instructions, such as never discussing refunds or giving medical advice. Tell it where to send those questions instead, such as your phone number or a contact form, and test that it follows the rule.
Yes, if customers ask about it. Write it down as short question and answer pairs. This is often the most useful part of the setup, and it helps new staff too.
Keep reading
More on the same subject, and where we can help with it.
Next step
Tell us what eats your time or where enquiries slip away. We will suggest the right mix of website chatbot and automation, with a fixed price.
We use cookies that are essential for this site to work. With your consent, we may also use analytics cookies to understand how the site is used. See our cookie policy for details.