How to train a chatbot on your products and policies

How to train a chatbot on your own products, prices, and policies: the steps, what to give it, how often to retrain, and what training really means.

By Matt · · 6 min read

How to train a chatbot

  1. List the questions it has to answer. The 20 your customers ask most, from your inbox, your phone, and your staff.
  2. Gather the content that answers them. Products, prices, sizes, stock, shipping, returns, payments, opening hours, and contact details.
  3. Clean it up. One fact in one place, no old prices, and no two versions of the same policy.
  4. Write its rules. What it must never say or promise, and when to hand over to a person.
  5. Load it in. Point the chatbot at your website, connect your product feed, and upload your documents.
  6. Test it with hard questions, then retrain it whenever something changes.
  7. Done! Your chatbot answers from your own products and policies.
Woman in a black t-shirt writing on a clipboard beside a row of shoe boxes at a timber packing bench in a sunny stockroom
Everything your chatbot needs to know is already in your stockroom, your inbox, and your head.

Every chatbot builder promises the same thing: "train it on your website in minutes". Switch it on, though, and it quotes last season's price, offers a size that sold out yesterday, or tells a customer something your returns policy never said.

That's not the AI being dumb. A chatbot only knows what you give it, in the shape you give it, as of the day you gave it.

Matey builds custom chatbots for Australian businesses, trained on each store's own products, prices, and policies. This guide covers what training a chatbot really means, what to give it, and how to keep it right. It's one step in how to build a chatbot, and the step that decides whether the rest works.

How are chatbots trained?

Most business chatbots aren't trained the way the AI model inside them was. The model already knows how to read and write. Training your chatbot means giving it your content to look up when a question comes in, plus rules on how to answer.

There are three things you can change, and only two are worth it for most businesses:

What you changeWhat it meansGood forNot good for
Its rulesWritten instructions it follows in every chatTone, what it must never promise, when to hand overFacts that change
Its knowledgeYour content, looked up fresh for every questionProducts, prices, sizes, policies, FAQsAnything you forget to keep current
The model itself (fine-tuning)Retraining the AI on your examplesA fixed writing style at huge volumeStock, prices, or anything that changes

When a chatbot builder says "train it on your website", it means the middle row: it reads your pages, cuts them into small pieces, and finds the few pieces that match each question. Older rule-based chatbots were trained differently, on lists of example phrasings for every question. AI chatbots don't need those lists.

What to train a chatbot on

Train a chatbot on everything a good staff member would need to answer your customers: your products, your policies, and the rules your team knows but nobody wrote down.

  • Products: name, price, sizes, colours, what's in stock, and what each one is for.
  • Buying advice: the guides on your blog, size notes, and how your staff steer people to the right product.
  • Shipping: costs, free shipping rules, express, times by state, and where you don't ship.
  • Returns and exchanges: the rule and every exception. Worn items, sale items, missing boxes, and who pays postage.
  • Payments: card, PayPal, Afterpay, Zip, and anything you don't take.
  • Contact: phone, email, hours, and what happens after hours.
  • Its rules: never promise a refund, never make up a code, never guess a restock date.

The small notes matter most. A footwear store we tested prints "Birkenstock sizing runs large. We recommend ordering one size down from your normal size for Regular widths" on every Birkenstock page. A chatbot that misses that note gives the wrong size to every customer who asks.

How to train a chatbot on your own data

To train a chatbot on your own data, connect the sources it should answer from (your website, your product feed, and your documents) and treat them as the one place each fact lives.

A website crawl reads what's printed on your pages, and that's where most DIY chatbots trip. When we built a DIY chatbot from an Australian footwear store's website, a work boot's page listed every size from a men's 7 to a 13, but the 13 was sold out. To anything reading the page's text, a sold-out size and one on the shelf look exactly the same. A product feed (the list of products, sizes, and stock that most store platforms can send to Google Shopping) carries the stock for every size, so it's the better source for anything that sells out.

Three rules keep your own data clean:

  1. One fact, one place. If your returns page says 30 days and an old FAQ says 14, the chatbot picks one, and you won't know which. Our DIY chatbot told a customer shipping was free over $100: the store's shipping policy says $160, but an old line on 26 of its collection pages still says $100.
  2. Write down what's in your team's heads. "We don't ship sale items to New Zealand" or "the black ones run small" is a rule only if it's written somewhere the chatbot can read.
  3. Date your policies. When a policy changes, replace the old page. Don't add a new one beside it.

How to train a chatbot on your own documents

To train a chatbot on your own documents, upload them to your chatbot builder: PDFs, Word documents, spreadsheets, or plain text. Keep each one plain and focused on one topic.

OpenAI's own help centre gives the same advice for files uploaded to ChatGPT: prefer clear, text-heavy files, because complex layouts make uploaded content harder to use (OpenAI). In practice:

  • Copy the text out of designed PDFs, like a catalogue or a brochure, into a plain document.
  • Turn tables of sizes or prices into simple rows, one product per row.
  • Skip scans and photos of text. The chatbot can't read them reliably.
  • Name every file for what's in it, like "returns policy" or "boot size guide", so you can find and replace it later.
  • Leave staff-only documents out. Supplier price lists, payroll, and internal notes don't belong in a customer chatbot. They belong in a company brain for your team.
Young bearded man on a stool at a surf shop counter highlighting a printed returns policy, laptop and drink bottle beside him
If your returns policy needs a highlighter to make sense, it'll confuse a chatbot too.

How often should you retrain a chatbot?

Retrain a chatbot whenever your products, prices, stock, or policies change. For an online store, that means every day.

Stores change faster than they feel like they do. When we counted an outdoor gear store's range over 11 days in October, 476 products changed, most of them sizes selling out (outdoor gear store case study). A chatbot trained once on day one was wrong about hundreds of products by day 11. Most chatbot builders let you resync your content, but check whether yours does it on its own or waits for you to press the button.

What goes wrong when you train a chatbot yourself?

The four things that go wrong most when you train a chatbot yourself are stale stock, clashing answers, missing rules, and too much noise.

  • Stale stock. It was trained on Monday and it's now Friday.
  • Clashing answers. Two versions of a policy, and it picks the wrong one.
  • Missing rules. The thing everyone on your team knows, which isn't written anywhere.
  • Noise. Menus, footers, and pop-ups read on every page, burying the answer.

Then test it. Our guide on how to test a chatbot has the 20 questions in the Matey break-it test, which find every one of these.

ironbarkworkwear.com.auSat 7:03 am · support opens Monday
Sunny
Live
Message Sunny…

Or skip the training. Matey builds custom chatbots for Australian businesses that read every product, price, size, and colour fresh from your site every night, so there's nothing for you to upload or keep up to date. Send us your website and we'll build a free demo from it, usually within 24 hours.

FAQs

How do you train a chatbot?

You train a chatbot by giving it your own content: products, prices, stock, shipping, returns, payments, opening hours, and the answers your staff give every day. Most chatbot builders read your website and any documents you upload, then answer from them. You also write its rules, test it with hard questions, and update it whenever something changes.

How are chatbots trained?

Most business chatbots aren't trained the way the AI model inside them was. The model already knows how to read and write. Training your chatbot means giving it your content to look up when a question comes in, plus rules on how to answer. Changing the model itself, called fine-tuning, is rarely worth it for a business chatbot.

How do you train a chatbot on your own data?

Connect the sources it should answer from: your website, your product feed or catalogue, your FAQ and policies, and any documents. The chatbot finds the parts that match each question and answers from them. Keep one current version of every fact, because a chatbot can't tell which of two clashing answers is the right one.

Can you train a chatbot on your own documents?

Yes. Most chatbot builders let you upload PDFs, Word documents, and spreadsheets. Plain, text-heavy documents with one topic each work best. Scanned pages, photos of text, and complex layouts are harder for a chatbot to read properly.

How often should you retrain a chatbot?

Retrain a chatbot whenever your products, prices, stock, or policies change. For an online store that means every day, because sizes sell out and prices change faster than most owners think. A chatbot trained once and left alone slowly turns wrong.

Can you train a chatbot with ChatGPT?

Not for your website. You can give ChatGPT your files inside a project, but that only helps you inside ChatGPT, and OpenAI is retiring custom GPTs on 11 December 2026. To train a chatbot for your website on the same models, use a chatbot builder that runs on them or connect your content to OpenAI's API.