Adding an AI chatbot to your site without the hallucinations

1. Why chatbots make things up

A language model predicts plausible text. If it does not have your prices, policies or opening hours in front of it, it will still produce an answer that sounds right. That is a hallucination, and on a business website it can mean a wrong promise to a customer.

2. Retrieval first, then answer

Instead of asking the model what it remembers, we search your own content first — pages, FAQs, product sheets — and give the model only the relevant passages. The instruction is simple: answer from these passages, and say so when they do not contain the answer.

• Keep the source content current; the bot is only as good as what it reads.
• Show the source page next to each answer so users can verify it.
• Test with the real questions your customers ask, not invented ones.

3. Guardrails that actually work

• Scope: limit the bot to topics you cover, and make it decline everything else politely.
• Fallback: when confidence is low, offer a contact form or WhatsApp instead of guessing.
• No promises: never let it quote prices, discounts or deadlines unless they come straight from your data.
• Review: read the conversation logs regularly and fix the gaps they reveal.

4. When not to use AI

If your questions are few and repetitive, a clear FAQ page is cheaper and more reliable. If the answer needs a person’s judgement — a custom quote, a complaint, anything legal — route it to a person. AI earns its place when there is a lot of content, many different questions and a team that cannot answer them all quickly.

5. Questions people ask

Related