How to add an AI chatbot to your website (and what we learnt)

To add an AI chatbot to your website, you either switch on one from a chat platform, use the one built into your help desk or website builder, or have one built around your own pages and systems. Whichever you choose, it should answer only from what your business has actually written, show where each answer came from, and hand anything else to a person. You can see what that looks like on your own site in about a minute with our free website chatbot demo.

Key takeaways

• A good website chatbot answers from your own pages and documents, and says so when they don't cover a question.
• There are three ways to add one: a chat platform, a built-in feature of software you already use, or a custom build.
• Your business is responsible for what it says. A tribunal held Air Canada liable for its chatbot's wrong answer.
• Fix your website first. A chatbot can only be as accurate as the pages it reads.
• Test it on real websites and real questions. Our own demo passed every test on a tidy site and then hit five problems on real ones.

What does an AI chatbot on a website actually do?

It sits in the corner of your pages and answers visitors' questions in plain language, at any hour. The useful kind doesn't rely on what an AI model happens to know about the world. It reads your website, cuts it into short passages, and each time someone asks a question it looks up the passages that match, then writes an answer from those alone. That method is called retrieval-augmented generation, or RAG, and our plain-English guide to RAG explains it with an open-book exam.

The difference matters. A chatbot that answers from general knowledge will guess your opening hours. One that answers from your pages either finds them or says it doesn't know.

Three ways to add an AI chatbot to your website

For a sense of platform prices: at the time of writing (November 2026), Tidio's free plan includes 50 conversations with its Lyro AI agent in total, and the Lyro add-on starts at $32.50 a month. Help desk products usually charge per resolved conversation, and our comparison of AI customer service software sets out what each charges for 1,000 resolutions.

Under the hood, the computing is cheaper than most owners expect. Our own demo runs on Cloudflare's Workers AI, which gives every account 10,000 "neurons" of computing a day free, and charges $0.011 per 1,000 neurons above that on its paid plan. The real costs of a chatbot are setting it up well and looking after it.

How to add one, step by step

1. Pick the jobs. Write down the ten questions you answer most often, by email, phone and message. Those are the chatbot's job description.
2. Fix your pages first. If your prices, opening times or delivery areas aren't written on your site, or are out of date, the chatbot can't get them right. This step improves your site for people and search engines too.
3. Choose the option from the table above. If it only needs to answer questions, a platform may do. If it needs to book, quote or check stock, it needs connecting to your systems.
4. Set the rules. What it must never answer (advice, complaints, anything about a specific customer's account), and when it hands over to you. Every handover should arrive with the visitor's name, contact details and question already taken.
5. Test it with 30 real questions, including awkward ones: a question your site doesn't answer, a rude one, one in another language, and someone trying to make it say something it shouldn't. Read every answer.
6. Install it. Usually a short snippet of code before the end of each page, or a plugin on WordPress, Shopify, Wix or Squarespace. Check your page speed before and after.
7. Read what people ask, every week. The questions it couldn't answer are a free list of what to add to your website.

What can go wrong?

It makes things up. In February 2024, a tribunal in British Columbia ruled on Moffatt v Air Canada. The airline's website chatbot told a grieving customer he could apply for a bereavement discount after booking, which wasn’t the airline’s policy. Air Canada argued the chatbot was responsible for its own actions. The tribunal rejected that, held the airline responsible for all the information on its website, chatbot included, and awarded the customer the difference in fare. If your chatbot says it, your business said it.

It gets talked into things. Visitors type "ignore your instructions" and worse. A chatbot that will change its role on request will eventually say something you'd never put on your website.

It leaks what it shouldn't. A chatbot connected to bookings or customer records needs the same care as any system holding personal data: it should only see what it needs, and never another customer's details.

It fails quietly. A slow site, a page that changed, a provider limit: chatbots break in ways nobody notices unless someone reads the answers.

What we learnt building one

We built our free website chatbot demo in October 2026: you give it a web address, it reads up to 15 public pages, and it builds a chatbot that answers from them and shows the page behind each answer. It passed every test on our own tidy test site. Then we ran it on real websites with the real model, and found five problems in one day:

TEST LOG · 29 OCT 2026REAL SITES, REAL MODEL

FIVE THINGS THE TEST SITE NEVER SHOWED

1. 01

   It thought before it spoke

   The model reasoned silently before every answer: about 9 seconds and twelve times the computing, and now and then an empty reply.

   FIXED: Thinking switched off for short, sourced answers.

2. 02

   It invented an email address

   A security service had hidden the site’s email from automated readers, so the page said “[email protected]”, and the model filled the gap with a plausible address.

   FIXED: Hidden emails decoded, and any email or phone number that isn’t on the site is removed from answers.

3. 03

   A new topic got “couldn’t find”

   Every short question was searched together with the one before, so “and do you deliver?” after a question about prices found nothing.

   FIXED: The last question is only added when the new one is a follow-up.

4. 04

   One slow website ran 84 seconds

   A site that answered slowly kept the reader waiting, and our host stops any request at 60 seconds.

   FIXED: A firm time limit for each page, and the build goes ahead with the pages already read.

5. 05

   A false alarm switched it off

   A check for “daily allowance used up” matched digits inside normal replies, and closed the demo for the rest of the day.

   FIXED: The check now reads the error code, not the text.

5 FOUND · 5 FIXEDNONE VISIBLE ON A TIDY TEST PAGE

The second one is the lesson for anyone buying a chatbot. The page didn't contain the email address, so the model produced a believable one. Nothing in a quick demo would have shown it. It took a real website and someone reading every answer.

After that, we built the rules every answer has to pass, in four layers:

1. Questions are screened first. Attempts to change its instructions, requests for anything harmful and requests for other people's data are turned down before any AI sees them. Anyone who mentions self-harm gets crisis lines, not a refusal.
2. The instructions are narrow. Answer about this business, only from the passages found, with no opinions or advice beyond what the site says.
3. Every answer is checked. It must point to a passage from the site, or it isn't shown. Links, emails and phone numbers that aren't on the site are removed.
4. A separate safety model reads each answer before the visitor sees it, and reads a sample of each website before a chatbot is built from it.

The rules are about intent, not topics, so a soap shop can still be asked about its bath bombs and a phone repair shop about cracked screens.

Try it on your own website

The demo is free, needs no code, and takes about a minute. You can ask it 20 questions, and the link lasts 7 days, so you can send it to a colleague. Every demo is labelled as a preview built by Tamkoz from your public pages, hidden from search engines, and deleted after 7 days. It never pretends to be your business.

Use it as a test of your website as much as of the chatbot. Ask it what your customers ask. Where it says your site doesn't say, that's a page worth writing.

What the demo can't do is what a real one is for: book appointments into your calendar, quote from your price list, answer from documents that never made it online, sound like you, hand conversations to your phone, and work on WhatsApp, Instagram and Facebook messages as well as your website.

What we would do

For most small businesses:

1. Run the demo on your own site and note every question it couldn't answer.
2. Fix those pages. Some businesses stop here, and that's fine.
3. Start with answers only, from your pages, with a clear handover to you, for a month. Read every conversation in the first week.
4. Then connect it to bookings, quotes or orders, one at a time, once you trust the answers.

That's how we build AI chatbots and assistants for clients, with the same four layers of rules. When the chatbot needs to answer from handbooks, price sheets and PDFs, a private knowledge base sits behind it. If most of your enquiries arrive by phone rather than by message, our guide to the cost of missed calls is the better place to start.