Short answer: A modern AI chat widget does not get retrained on your business. It reads your website, breaks the text into passages, stores those passages in a searchable index, and then — every time a visitor asks a question — it looks up the most relevant passages and hands them to a language model as context. The model writes the answer using your words, not its memory.
That distinction sounds academic until something goes wrong. If you believe the model "learned" your pricing, you will not understand why it answers from a page you deleted last month. Once you understand that answers come from a retrieval step, everything about maintaining a chatbot becomes obvious.
What actually happens when you install the widget
You paste one script tag into your site. Behind that tag, three things happen in sequence.
First, a crawler walks your website the way a search engine would: it follows internal links, reads the visible text of each page, and skips navigation boilerplate where it can. This is the raw material. Your homepage, your services pages, your FAQ, your about page — all of it becomes text.
Second, that text is split into chunks and converted into embeddings. An embedding is a list of numbers that represents meaning rather than spelling. Two passages that say the same thing in different words end up numerically close together. This is why a visitor can type "how much do you charge" and match a page headed "Investment" — no keyword overlap required. Those embeddings live in a vector store, which is just a database built to answer the question "what is closest in meaning to this?"
Third, when a visitor asks something, the question is embedded the same way, the closest passages are pulled out, and those passages are sent to the language model along with instructions about who your business is and how to behave. The model composes a reply grounded in that retrieved text. The whole pattern is called retrieval-augmented generation, and it is the reason a general-purpose model can talk accurately about a plumbing company in Leeds it has never heard of.
Why nobody retrains a model on your website
Retraining a large language model means adjusting billions of internal parameters. It is expensive, slow, and — critically — irreversible in any practical sense. If you retrained a model on your old pricing, you could not simply delete the old price. It would be smeared across the weights.
Retrieval avoids all of that. Your content sits outside the model in a knowledge base you control. Change a page, recrawl it, and the next answer uses the new text. Delete a page and the passage disappears from the index. Nothing has to be unlearned because nothing was ever learned in the first place.
There is a second benefit that matters more than most owners realise: attribution. Because the answer was built from specific retrieved passages, the system can point at where the answer came from. A model answering from memory cannot do that, which is exactly how confident-sounding nonsense gets produced.
What the model contributes
If your pages supply the facts, what is the model for? Three things.
It handles language. A visitor asks "do you lot come out on a Sunday" and your page says "weekend appointments available by arrangement." Matching those two is the model's job.
It handles synthesis. A good answer often needs two facts from two different pages — your service area from one, your call-out policy from another. The model stitches them into one paragraph rather than dumping two excerpts.
It handles tone. The same fact can be delivered curtly or warmly. The instructions wrapped around the retrieved passages set that tone, so the widget sounds like your business rather than like a generic assistant.
Filling the gaps your website does not cover
Websites are marketing documents. They are rarely complete answers to what customers actually ask. Most businesses discover a gap within days of going live: nobody wrote down the cancellation policy, or the parking situation, or whether you take same-day bookings.
There are two ways to close a gap. Write the answer into a page — good for anything customers should be able to find themselves — or upload a knowledge file. Clerkzo accepts ten file types at up to ten megabytes each, so an internal FAQ document, a price sheet, or a policy PDF can go straight in without ever appearing on your public site.
The third mechanism is the corrections loop, and it is the fastest. When you spot a wrong or vague answer in your conversation history, flag it and type what the answer should have been. That correction becomes an authoritative override — the next time a similar question arrives, your version wins. It is closer to training a new hire than to configuring software: you review the transcript, you say "no, tell them this instead," and it sticks. This is what chatbot training means in practice for a retrieval-based system.
How long it takes to be useful
The crawl is the slow part, and it is measured in minutes, not weeks. After that you have a widget that can already answer the questions your site answers well. The gap between "working" and "genuinely good" is usually a week or two of reading real conversations and correcting the handful of answers that come out thin.
That review habit is the whole discipline. There is no model tuning, no intent tree to maintain, no scripts to write. You read what visitors asked, you notice what your business never explained clearly anywhere, and you fix it once. Often the fix improves your website too, because the gap was real for human readers as well.
You can test all of this before it goes live. A playground that mirrors the live widget lets you ask the awkward questions yourself — the pricing edge cases, the competitor comparisons, the things you would rather it declined to answer — and see exactly what a visitor would get. The mechanics are walked through step by step on how it works.
Frequently asked questions
Does my content get used to train the underlying AI model?
No. Your content is stored in your own index and retrieved at answer time. Clerkzo does not use your data to train AI models, and the model provider is not given your pages as training material.
What happens if I change my prices?
Update the page, recrawl, and the old passage is replaced. If you want certainty on a specific question, add a correction so your exact wording becomes the authoritative answer.
Can it answer questions my website does not cover?
Only if you give it the information — through a knowledge file or a correction. If nothing in your knowledge base covers the question, a well-configured widget says it is not sure and offers a handoff rather than guessing.
Do I need any technical skill to set this up?
No. Installation is one script tag, and everything after that is reading conversations and typing corrections in plain English.