By Rakibul Hasan, Moonlab
An AI chatbot for your website answers visitors' questions using your own content, captures leads and passes complex conversations to a person. Done well it saves repetitive replies. Done badly it frustrates visitors and makes things up.
This guide covers how to add an AI chatbot to a website: deciding what it should do, setting it up, training it on your content, building a clean handoff, and the pitfalls that catch most small businesses. Chatbot products and pricing change often, so check each vendor's current details.
Decide what the chatbot is for
Start with the job. A chatbot that tries to do everything does most things poorly. Pick one to three purposes and write them down.
- Answer common questions about services, hours, areas, delivery and returns.
- Capture leads by asking for name, contact details and what the visitor needs.
- Book appointments or calls through a link to your calendar.
- Help visitors find the right page or product.
- Route support requests to the right person.
How the setup works
Most website chatbots follow the same pattern. You sign up with a chatbot platform or build one in a tool such as n8n, connect your content, set instructions, style the chat widget, and paste a small code snippet or install a plugin on your site. WordPress, Shopify, Wix and Squarespace all support this in some way, though the steps vary.
Before you go live, decide where the widget appears. Putting it on every page can distract from your main call to action. Often it suits pricing, contact and product pages best.
Training it on your content
The common method is retrieval: the chatbot searches your content for passages relevant to the question and writes an answer from them. It is not retraining a model. That means the quality of the answers depends on the quality of your content.
Gather your service pages, FAQs, policies, price list and any documents customers ask about. Remove outdated material. Fill gaps by writing short answers to the twenty questions you hear most. Tell the chatbot to answer only from this material and to say it does not know when the answer is missing. Re-sync whenever your content changes, because a chatbot quoting last year's prices is a liability.
Handoff to a human
A chatbot should know when to stop. Define triggers: the visitor asks for a person, expresses frustration, mentions a complaint or a refund, asks about something outside its content, or the bot has failed to help twice.
At handoff, offer a clear route: a live chat with your team during hours, or a form that sends the full conversation to your inbox with a promise of a reply time. Do not make visitors repeat themselves. Pass the transcript along. If you also run phones, the same principle is covered in moonlab.space/guides/ai-receptionist-small-business.
Pitfalls to avoid
Most chatbot problems are predictable.
- Made-up answers. Without strict instructions the bot may invent policies or prices. Limit it to your content and test with trick questions.
- No way out. Always provide a visible option to contact a person.
- Pretending to be human. Say clearly that it is an automated assistant.
- Stale content. Assign someone to update the sources monthly.
- Pop-ups that block the page. Keep it unobtrusive, especially on mobile.
- Ignoring conversations. Read a sample of chats each week to find gaps and fix them.
Privacy and consent
A chatbot collects personal data: names, emails, and whatever visitors type. Tell visitors what is collected and why, link to your privacy policy, and avoid asking for sensitive information. If you serve EU or UK visitors, GDPR rules apply, and your cookie and consent setup may need to cover the chat widget. This is not legal advice, so check with a professional.
Ask the vendor where chats are stored, how long they are kept, and whether they are used to train models. See moonlab.space/guides/gdpr-ai-automation and moonlab.space/guides/cookie-consent-uk-eu for more.
A worked example: a chatbot for a small dental practice
Say a dental practice in Leeds wants a chatbot that handles the questions reception hears daily, without giving clinical advice. The owner collects the content: treatment pages, opening hours, new patient information, price ranges the practice is happy to publish, and the cancellation policy. She writes a short list of twenty questions patients ask most and a clear answer for each.
The instructions say: answer only from the supplied content, never give medical advice or diagnoses, and if someone mentions pain, swelling or bleeding, show the emergency contact details and offer a callback. For booking, the bot links to the online booking page. For anything else it collects a name, phone number and reason, and emails the reception team. She tests thirty tricky questions, including ones about treatments the practice does not offer, and fixes the answers that were wrong. Only then does the widget go live, on the contact and treatment pages.
- Mistake: letting the bot answer outside your content. Tell it to say it does not know and offer a person.
- Mistake: giving advice in regulated areas such as health, law or finance. Add hard rules and test them.
- Mistake: no monitoring. Read a sample of conversations weekly for the first month.
- Mistake: forgetting mobile. Check the widget does not cover buttons or forms on a phone.
- Mistake: leaving old pricing in the sources. Re-sync after every change.
- Decision rule: if you cannot staff the handoff, offer a callback form rather than live chat.
How to know if it is working
Pick measures that match your goal: number of leads captured, questions answered without a human, handoffs that were needed, and visitor ratings if you ask for them. Keep them modest and honest. A chatbot that answers half the routine questions correctly and hands off the rest cleanly is a success.
If you want a second pair of eyes on your setup, the free audit at moonlab.space/contact is a good starting point, and our web work is described at moonlab.space/services/web-development.
