What “training” actually means
The AI model itself isn't retrained on your site. Instead, most website chatbots, HeyKiko included, work like this:
- Step 1
Read
Each page's text is fetched, without menus, footers and code.
- Step 2
Split
The text is cut into short passages, so an answer can use just the part that matters.
- Step 3
Index
Each passage is stored in a way that can be searched by meaning, not just by matching words.
- Step 4
Search
When a customer asks, the passages closest to their question are found, often by searching several ways at once.
- Step 5
Answer
The model writes a reply from those passages, and can say which page it came from.
This is why a change to your website reaches the chatbot only once the page is read again, and why the chatbot can only be as right as the pages it reads.
Which pages to include, and which to leave out
| Include | Leave out | |
|---|---|---|
| Selling | Services, products, prices, plans | Old promotions and expired offers |
| Help | FAQ, delivery, returns, warranty, how-to articles | Release notes nobody asks about |
| Contact | Opening hours, locations, how to book | Careers and team bios, unless customers ask |
| Policies | Terms customers ask about, such as refunds | Pages behind a login, admin pages |
An out-of-date page is a common cause of wrong answers. Our own agent tests keep an archived page with last year's delivery fee next to the current one, because it's exactly the trap a real website sets. Remove or unpublish old prices rather than leaving them for the chatbot to find.
Step by step
- Step 1
Find your pages
Your sitemap lists them: usually yoursite.com/sitemap.xml. It's the standard file websites use to tell search engines which pages exist, and most chatbot builders read it first. Sitemaps.org: protocol
- Step 2
Import and choose
Give the chatbot your address, then untick the pages from the leave-out list above.
- Step 3
Add what your site doesn't say
Price lists, menus, brochures and policies that live in PDFs or documents, uploaded as files. Many builders, HeyKiko included, don't follow links to PDFs on your site, so upload those.
- Step 4
Test with real questions
Twenty questions your customers really ask, including a few your site doesn't answer (see below).
- Step 5
Fix the source
When an answer is wrong, change the page or file it came from rather than adding a rule to the chatbot's instructions.
- Step 6
Keep it fresh
Have it read your site again on a schedule, such as weekly, so a new price is picked up without you remembering.
Write pages that answer well
- Put the number in the sentence: “Delivery is $49, free on orders of $1,500 or more” answers better than a table cell with no context.
- Say the exception next to the rule. “Clearance items can't be returned unless damaged” stops the chatbot promising a refund it shouldn't.
- Keep text as text. A price that only appears in an image, or a scanned PDF with no text layer, gives most chatbots nothing to read; HeyKiko doesn't read text inside images.
- One topic per page or section, with a clear heading, so the right passage is easy to find.
- Write how your customers talk. If they say “refund”, don't only say “reimbursement”.
What a chatbot can't learn from your website
- Anything live: stock levels, an order's status, today's open slots. These need a connection to your systems, such as your database or your calendar.
- Anything not written down. If your team “just knows” it, write it in a page or a document first.
- Pages a site hides from automated readers. Some sites block bots or build their pages only in the browser. A robots.txt file asks crawlers to stay away from parts of a site, and well-behaved ones listen. Google: robots.txt
When HeyKiko can't find a site's pages because the site blocks page finders, it checks the home page in a real browser and tells you, so you can add the other pages by hand.
How to test it before customers do
Ask a mix like this, and check where each answer came from:
| Kind | Example | A good answer |
|---|---|---|
| A plain fact | How long do I have to return something? | Gives the number from your returns page |
| An exception | Can I return a clearance item? | Applies the exception, not the general rule |
| Two pages at once | What does delivery to an island cost on a big order? | Combines the free-delivery rule and the surcharge |
| Not on your site | Do you price match? | Says it doesn't know and points to your team |
| Another language | The same question in your customers' other language | Answers in that language, from the same pages |
An answer that invents a policy for the “not on your site” question is the one to fix first: tell the chatbot to admit what it doesn't know, or hand the chat to a person.
Doing it in HeyKiko
In HeyKiko you type your website address, choose the pages, and set how often they're read again: daily, weekly, monthly or never. Files can be PDF, Word, Excel, CSV, text and Markdown. The Playground shows which source each answer used, and every page shows how many answers it has supplied. See training on your website and files and testing your agent.
You can try it on your own website free for 14 days, with no card.
Sources
Checked on 5 November 2026. We make HeyKiko, which this guide uses as its example.
