Guide
How to train an AI chatbot on your website content, so it answers correctly
Every AI chatbot now promises it can be “trained on your website in minutes”. The crawl is the easy part. Whether it then answers correctly depends on what it reads, how your pages are written, and what it does when your site doesn’t have the answer. This guide covers all three.
Maqsood R.Founder, HolaReply · Updated · 8 min read
The short answer
- To train an AI chatbot on your website, you let it index your published pages, posts and products.
- It splits them into passages, turns each into an embedding (a numeric fingerprint of its meaning) and, for every question, finds the most relevant passages and answers only from them.
- Prepare by removing outdated and conflicting pages, writing key facts in plain text and adding short answers for questions no page covers.
- Choose a chatbot that re-indexes when you edit and says “I don’t know” when nothing matches.
What “training” really means
You are not retraining the AI model. That would cost a fortune and take weeks. Modern website chatbots use a method called retrieval-augmented generation (RAG):
Split
Your pages are split into short passages, each about one topic.
Index
Each passage becomes an embedding: a list of numbers that captures its meaning, stored in an index.
Retrieve
When a visitor asks something, the question is turned into an embedding too, and the closest passages are found.
Answer
The AI model writes an answer using only those passages, and ideally shows which page they came from.
So “training” is really indexing. The quality of the answers depends on three things: what is in the index, whether the right passage is found, and whether the model sticks to it.
Step 1: Decide what it should read
| Index | Leave out |
|---|---|
| Product pages and product data | Drafts, private pages and anything password-protected |
| Delivery, returns, payment and warranty pages | Prices or terms shown only to logged-in or trade customers |
| FAQs and help articles | Old promotions, expired offers and outdated policies |
| About, contact and opening hours | Blog posts with opinions you wouldn’t want quoted as policy |
| Service and pricing pages (if prices are public) | Internal documents and anything not meant for the public |
A public chatbot should only know public content. HolaReply indexes published content only, and reads WooCommerce prices as a signed-out visitor sees them, so a trade price is never quoted.
Step 2: Clean up before you index
The most common cause of wrong answers isn’t the AI. It is the website. Before you index:
- Remove or update pages with old prices, old delivery times or old policies.
- Pick one page as the source for each policy; delete or redirect duplicates that disagree.
- Put important facts in text, not only in images, sliders or PDFs.
- Write numbers with units: “Delivery in 2–3 working days in the UK”, not “fast delivery”.
- Give each product its key facts on the page: size, weight, materials, what’s included.
Step 3: Write pages a chatbot can use
Retrieval finds passages, not whole pages. Pages that work well for people usually work well for a chatbot, with a few habits:
- One topic per section, under a clear heading that says what it answers (“Delivery to Ireland”, not “Good to know”).
- Say the thing directly. “We deliver to Ireland in 3–5 working days for €9” retrieves better than a paragraph about “our European logistics network”.
- Use the words customers use. If people say “postage”, mention postage as well as shipping.
- Keep tables simple, with a heading row, so each row still makes sense on its own.
Step 4: Fill the gaps with a knowledge base
Some of the questions customers ask most are written nowhere on your site: “Do you do gift wrapping?”, “What’s the cut-off for next-day dispatch?”, “Can I collect in person?”. Add them as short question-and-answer pairs. Keep each answer to a few sentences and one fact.
Answers you write yourself. The assistant uses them exactly like pages of your site.
The knowledge base: short answers for questions your pages don’t cover.
Step 5: Keep it in sync
An index is a snapshot. When you change a price or a policy, the chatbot must learn it too, or it will confidently repeat the old one. Check how your chatbot updates:
| Method | What it means |
|---|---|
| Re-index on edit | Each page or product is re-indexed as soon as you save it (HolaReply for WordPress works this way) |
| Scheduled re-crawl | The whole site is crawled again daily or weekly; changes can be days late |
| Manual re-sync | Nothing changes until you press a button |
Step 6: Make it admit what it doesn’t know
Even a well-trained chatbot will be asked things your site doesn’t cover. What happens next decides whether it helps or hurts:
- It should answer only from the passages it found, never from the model’s general knowledge.
- It should show the source, so visitors and you can check.
- When nothing relevant is found, it should say so and offer a person, capturing the lead.
- Topics like prices, contracts or allergy advice should be locked, so a person always answers.
straight from the growers.
Trade packs for kitchens, bakeries and shops. Order by the kilo or by the pallet.
Shop wholesaleView product
An answer with its source; nothing found means an honest “I don’t know” and a person.
Step 7: Test, then keep improving
Ask the 20 questions your team hears most, in the words customers use. For each: is the answer right, and is the source the page you expected? Then look at what visitors actually ask. HolaReply Pro groups unanswered questions in a report, so you can add the missing answers to a page or the knowledge base and the chatbot answers them next time.
What visitors asked that your site doesn’t answer yet — most asked first.
The unanswered questions report (Pro): the gaps in your content, grouped.
What it costs
Indexing is cheap: creating embeddings for a 500-page site costs roughly $0.10 on Google’s Gemini embedding model, once, plus a fraction of a cent each time a page changes. Answering costs about $1.70 per 1,000 answers on Gemini Flash, depending on the model.
With HolaReply the cost depends on where you use it. On any website the AI is included in the plan, with no AI key: Pro is $19 a month with 500 AI conversations, Business $49 with 2,500 (pricing). On WordPress the plugin is free and you use your own AI key, billed by the provider, or HolaReply AI credits (pricing). A Shopify app is coming soon to the Shopify App Store.
Test a trained one. The chat on this page is HolaReply, indexed on this website. Ask something the site covers, then something it doesn’t.
“What happens when HolaReply can’t answer?”“What’s your office address in Tokyo?”
Open the chatQuestions people ask
How long does it take to train a chatbot on a website?
Indexing a typical small-business site takes minutes. Cleaning up outdated pages and adding missing answers takes longer, and makes the biggest difference.
Can a chatbot learn from PDFs?
Some can. Text on your web pages usually works better and stays in sync more easily. If an important document is only a PDF, publish its key facts as a page too. HolaReply on any website answers from files you upload (PDF, Word, text, CSV price lists); the free WordPress plugin answers from your pages, posts and products.
Why does my chatbot give wrong answers even after training?
Usually because two pages disagree, a page is out of date, the fact is only in an image, or the chatbot fills gaps from the model’s general knowledge. Clean up the content and choose a chatbot that answers only from sources.
Does training a chatbot send my content to the AI provider?
The passages are sent to create embeddings and, when relevant, to answer questions. HolaReply removes names, emails and phone numbers from visitors’ questions before they are sent.
Do I need to re-train it when I update my site?
With a chatbot that re-indexes on edit, no. With scheduled or manual syncs, yes, after every important change.
Sources
Checked on 25 September 2026. Prices and features change; tell us in the chat if something here is out of date and we will fix it.
Founder of HolaReply. He builds AI assistants that answer only from a business’s own website, and checks every fact in these guides against the products’ own pages.
More by Maqsood R. →An assistant that knows your site, and admits what it doesn’t.
One line of code on Wix, Squarespace, Webflow or any website: 7 days free, no card. Free plugin for WordPress; Shopify app coming soon.