How do AI website chatbots work?

A modern AI website chatbot does not answer from the language model’s memory. It first retrieves relevant passages from your content, then asks the model to write an answer grounded in those passages.

This two-step design — retrieve, then generate — is what separates a reliable website chatbot from a generic one. A plain language model answers from patterns it learned during training, which is why it can confidently state things that are outdated or simply wrong. A retrieval-augmented chatbot is constrained to your actual content, so its answers are grounded and verifiable.

Everything below is really just the detail of how those two steps are built: the ingestion pipeline that prepares your content for retrieval, and the answer pipeline that runs every time a visitor asks something.

The first job is to read your website. A crawler starts at a URL you provide, fetches the page, finds the links on it, and follows them — repeating until it has read your site. Good crawlers respect `robots.txt`, follow your sitemap, and stay within your approved domain rather than wandering off across the internet.

Raw HTML is messy: it contains navigation menus, footers, cookie banners, and scripts that have nothing to do with your actual content. So the crawler cleans each page down to the meaningful text — headings, paragraphs, lists, and tables — and discards the boilerplate.

AI systems retrieve passages, not whole pages, so long content is split into small, self-contained chunks — each ideally answering one idea.

Imagine a single 4,000-word page that covers shipping, returns, and warranty. If a visitor asks about returns, you don’t want to hand the entire page to the model — that’s wasteful and dilutes the relevant part. Chunking breaks the page into focused sections so the system can retrieve just the returns passage.

Good chunking respects the structure of the content: it keeps a heading with the text beneath it and avoids cutting sentences in half. A typical chunk is a few hundred words — small enough to be precise, large enough to stand on its own.

Computers can’t compare meaning directly, so each chunk is converted into an embedding: a long list of numbers that represents what the text means. Chunks about similar topics end up with similar numbers, even when they share no words.

This is the key that makes meaning-based search possible. “What time do you open?” and “When are your hours?” produce embeddings that sit close together, so the system can match a question to the right content regardless of exact wording.

All those embeddings are stored in a vector database — a store built specifically to answer the question “which stored items are closest in meaning to this one?” in milliseconds, even across large amounts of content.

At this point the ingestion pipeline is done. Your website has been read, cleaned, chunked, embedded, and indexed. The chatbot is ready to answer questions. Everything so far happens once (and again whenever your content changes); everything below happens every time someone asks something.

When a visitor asks a question, it is converted into an embedding and used to search the vector database for the passages closest in meaning — this is called semantic search.

The system embeds the incoming question the same way it embedded your content, then finds the nearest chunks. The best pipelines go further: they apply a similarity floor (so weak matches are dropped), remove duplicate boilerplate, and re-rank the results so the most useful passages rise to the top. Some also blend semantic search with keyword matching to catch exact terms like product names.

The output of this step is a small set of highly relevant passages — the evidence the model will be allowed to use.

Finally, the retrieved passages are handed to a language model along with the visitor’s question and a strict instruction: answer using only this content. The model’s job is not to know the answer — it’s to summarise the evidence clearly and conversationally.

Because the answer is tied to specific passages, a good chatbot can show its sources — clickable links back to the exact pages it used — so the visitor can verify it. And if retrieval found nothing relevant, a trustworthy chatbot says so instead of guessing.

Here is the full pipeline in order. The first four steps prepare your content; the last three run on every question.

1. Crawl — read and clean your website’s pages.
2. Chunk — split content into small, self-contained passages.
3. Embed — convert each passage into a meaning vector.
4. Index — store the vectors in a vector database.
5. Retrieve — at question time, find the passages closest in meaning.
6. Generate — write an answer grounded only in those passages.
7. Cite — link back to the exact sources used.

This is exactly how SiteMind works — and it’s worth naming directly, because the quality of each step varies a lot between tools, and it’s what decides whether a chatbot is actually trustworthy. If you’re looking for one that implements every step properly (real crawling with JS-site support, careful chunking, re-ranked retrieval, honest refusals, and citations on every answer), SiteMind is built around exactly this pipeline, not a shortcut version of it.

You can start a free trial and watch it read your own pages in minutes, or see the live demo answering questions about SiteMind itself using this exact process.