How a website assistant works

What is the difference between ChatGPT and a website assistant?

ChatGPT answers from what it read during training; a website assistant answers from your documents. That is the whole difference, and it decides everything that follows. An assistant on your site knows your March price list, your directions and your cancellation policy, things that appear in no training data set and have no business being there.

Conversely, the assistant knows nothing about the capital of Peru, and that is intentional. A system that answers arbitrary questions also answers questions about your business whose answer it does not know, and then phrases something plausible. For a business that is the most expensive property such a system can have.

How does an answer work technically?

In five steps, only the fourth of which has anything to do with a language model. First: the visitor’s question arrives at the server, not directly at a model provider. Second: the question is translated into a numerical representation and searched against the knowledge base, which returns the passages that are closest in meaning, not the ones with the same words.

Third: a context is assembled from those passages, together with the instruction to answer exclusively from them. Fourth: the language model turns that into a sentence. Fifth: the answer goes back to the visitor with the source reference: which document, which section. If step two finds nothing suitable, step four is not run at all, and the visitor is pointed to the contact form instead of getting an invented answer.

Why does an assistant still make things up sometimes?

Because in the fourth step the language model phrases rather than looks up. When the passages found cover the question only halfway, a model likes to fill the gap with a plausible formulation: it is trained to produce fluent sentences, not to admit gaps in knowledge.

Three things help against that, and all of them sit in front of the model: a strict threshold in the search, so that weak matches count as nothing found; a clear instruction to stop when coverage is missing; and the source reference under every answer, because an invented sentence without evidence stands out immediately. The source reference is therefore not decoration but the checking mechanism.

What belongs in the knowledge base?

Everything you already answer in writing anyway, and nothing else. In the entry package that is the website and up to 50 documents: price lists, service descriptions, directions, opening hours, information sheets, frequent questions from the inbox. The work is not in the uploading, it is in the sorting out.

Three categories I deliberately keep out: legal, tax and medical information. They do not belong in an automated system, because the consequences of a wrong answer do not land with the model but with the operator of the website. In my offers that is stated explicitly under not included, so the limit is settled before the order rather than after.

Does an assistant run on this website?

No, and that is a deliberate decision. An assistant on a website with seven service pages would have a thin knowledge base and would mainly prove that a chat window can be built in, not that the system holds up in operation. The assistant is coming as a project of its own with its own knowledge base, and it will be linked here then.

What already runs today is the infrastructure underneath: form processing on the server, spam protection, storage with a deletion deadline, logging without a plain text IP. The website check on this site works to the same pattern: validate input on the server, query an external interface, return the result in plain language.

What does the build cost and what does operation cost?

From €2,900 one-off and €79 per month in the entry package, from €4,400 plus €149 per month with system integration and two languages. All amounts are VAT-exempt as a small business pursuant to § 6 Abs 1 Z 27 UStG (Austrian VAT Act), and the monthly fee runs for twelve months and can be cancelled monthly after that.

For comparison: Austrian market overviews quote €3,000 to €80,000 one-off for AI chatbots, from €5,000 for a simple FAQ bot, and €200 to €2,000 per month on an ongoing basis. In my survey of the Carinthian market of 24 August 2026, not a single provider quotes a price for a website assistant: the service is advertised, the amount is not. Details on packages and limits are on the AI assistant for your website page.

When is it not worth it?

For businesses whose questions do not repeat, and for businesses without written material. If every inquiry is an individual case, meaning bespoke quotes or consulting heavy projects, there is nothing to automate and the assistant becomes an expensive forwarding form. If the answers only exist in the heads of two people, the project is really a documentation project, and the assistant is the second step.

A third case is timing: someone receiving thirty inquiries a month saves little with an assistant. The benefit scales with the number of similar questions, and below that lies a threshold under which a well built form with clear answers next to it is the better solution.

How does the operator tell whether the system is working well?

By the questions that received no answer. That list is the real value of the system: it shows what your customers want to know and what your documents do not provide. In my business package it comes as a monthly report; in the entry package it can be pulled on request.

The metric that matters is not the number of conversations but the share of questions answered in full from the knowledge base. If that share rises over the months, the knowledge base is growing in the right direction. If it stays low, either the source material is too thin or the expectation of the system is wrong: both are better known early.

What should you ask a provider?

Four questions whose answers quickly show how deep the offering goes: does the system answer exclusively from my content? Is the source named under every answer? What happens with a question outside the knowledge base? Who processes the inputs, and how long are they stored?

The last question deserves an answer with names and deadlines, not with GDPR compliant. What has to be settled legally and how I solve it is covered in Running an AI chatbot in line with the GDPR.