Website AI chatbots: making them answer from your own content

How do you add an AI chatbot that answers from your own content?

Collect the content you want it to answer from, split it into small passages, index them for search, and for each question retrieve the most relevant passages and give them to the model with instructions to answer only from them, say when it does not know, and hand off to a person when needed. This approach is called retrieval-augmented generation. It keeps answers grounded in your pages, policies and documents instead of the model's general knowledge, and it lets you update the chatbot by updating your content.

This is the build we use in our AI automation projects for website assistants.

Which content should the chatbot use?

Start with the content that already answers customer questions well:

• Service and product pages, pricing explanations and FAQs.
• Help articles, policies such as returns or delivery, and onboarding guides.
• Selected internal documents, such as product sheets, once you have checked they are current.

Leave out anything outdated, contradictory or confidential. If two pages disagree, the chatbot will too, so content clean-up is usually the first real task. Answer quality depends more on the sources than on which model you choose.

How does retrieval work?

1. Chunking: each document is split into passages of a few paragraphs, keeping headings so each passage makes sense on its own.
2. Indexing: passages are turned into embeddings, numerical representations of meaning, and stored in a vector index, often combined with keyword search.
3. Retrieval: for each question, the system finds the passages closest in meaning.
4. Answering: the model receives the question, the retrieved passages and the rules, and writes an answer that can cite the source page.

Re-index automatically when content changes, so the chatbot never quotes last year's prices.

What guardrails does it need?

• Answer only from the provided sources, and say clearly when the answer is not in them.
• Stay on topic: decline questions unrelated to your business politely.
• No promises on prices, deadlines, legal or medical matters beyond what your content states.
• Link to sources so visitors can check the answer themselves.
• Escalate anything sensitive, such as complaints, refunds or account changes, to a person.

How should hand-off to a person work?

A good chatbot knows its limits. When it cannot answer, or the visitor asks for a person, it should collect the visitor's name, email and question, and pass the whole conversation to your team through email, your help desk or your CRM. During business hours it can offer live chat or a booking link; outside them, it promises a reply time your team can keep. These connections are usually built with workflow automation or directly into your systems.

What about privacy and data protection?

Through the business APIs of the main providers, your data is not used for training by default. OpenAI states that data sent to its API is not used to train or improve its models unless you explicitly opt in, and that abuse-monitoring logs are kept for up to 30 days by default (OpenAI: your data). Anthropic's commercial terms state that it may not train models on customer content from its services (Anthropic commercial terms). Still, avoid sending personal or sensitive data the chatbot does not need, tell visitors they are talking to an AI assistant, and cover the chatbot in your privacy policy.

Why do logging and review matter?

Log every conversation with the sources used. Each week, review questions it could not answer, answers visitors rated poorly and conversations that ended in a hand-off. The fixes are usually content gaps, which you solve by writing a clear answer on your website, improving both the chatbot and your search visibility at once.

How do you keep costs under control?

• Set daily and monthly usage limits on the API keys.
• Use a smaller model for simple questions and a larger one only where needed.
• Cap conversation length and the number of passages sent per question.
• Cache answers to very common questions.

How should you launch it?

Test the chatbot with a list of real questions from your inbox and sales calls, including awkward ones, before it goes live. Launch on a few pages first, watch the logs closely for two weeks, then expand. Make sure it loads without slowing the page, which our web development team checks as part of any build, and see our guide to AI search optimization for making the same content easy for external AI assistants to cite.

Frequently asked questions

Last updated November 1, 2026