·5 min read
How to build an AI chatbot for your website
A clear blueprint for shipping a useful AI chatbot on your site — data sources, guardrails, model choice, and the metrics that matter.
AI chatbotRAGcustomer support
A useful chatbot is not a wrapped LLM
If you drop a generic chat UI on your site, you get hallucinations and frustrated users. A useful chatbot is grounded in your content and bounded by clear rules.
The four pieces
- Sources — docs, help center, product pages, and FAQs ingested into a vector store
- Retrieval — semantic search + re-ranking so the model sees the right 3–5 chunks
- Generation — a small, fast model with a strict system prompt and forced citations
- Escalation — a "talk to a human" handoff with the conversation attached
What to measure
- Containment rate (resolved without human)
- Citation coverage (% of answers backed by a source)
- CSAT on bot conversations vs human ones
- Cost per resolved conversation
Common mistakes
- Letting the bot answer anything — restrict it to your domain
- Ignoring stale content — wire ingestion to your CMS updates
- Skipping evals — write 30 test questions before launch, run them on every prompt change
