By Danijel Balog, Founder, Fragon Studios · Last updated: 2026-11-13
An AI chatbot for your website
A chatbot that answers visitors' questions from the content of your own website, with clear rules on which data goes where. For websites with many recurring questions, including WordPress sites.
| Language model | What leaves your server | What it needs | Running costs |
|---|---|---|---|
| with an external provider | at least the visitor's question and the text passages found | a data processing agreement, with the transfer to third countries checked | per call, passed on by us at cost |
| self-hosted, for example via Ollama | nothing to an external AI provider, if indexing and cache also run on your side | a dedicated server with suitable hardware | none per call, but the hardware instead |
In short
- The bot answers from the content of your website. What is not there, it cannot know.
- Our WP FAQ Bot runs as a WordPress plugin on ordinary shared hosting.
- It gives no binding prices, legal information or commitments, and wrong answers can never be ruled out entirely.
- Workshop from €1,500, feasibility analysis from €3,000, implementation from €5,000.
How does an AI chatbot on your own website work?
It searches the content of your website for matching passages and has a language model write an answer from them.
The method is called RAG (retrieval-augmented generation): search first, then answer.
Index
Split pages and posts into passages.
Search
Find the passages that best match the question.
Answer
A language model writes the answer from them.
Cache
A cache answers similar questions.
You set the tone of the answers in the system prompt.
During indexing, pages and posts are split into passages and stored as vectors, that is, as sequences of numbers that represent the meaning of a text. When content changes, the index updates automatically.
The tone, language and behaviour of the answers are set by a system prompt that you can adjust yourself. Identical and similar questions are answered by a cache, without calling the language model again.
Is there an AI chatbot for WordPress that runs on normal hosting?
Yes, our WP FAQ Bot. The WordPress plugin runs on ordinary shared hosting and needs no external services apart from the language model.
- 500 msSearch time stays below this with 1,000 to 5,000 content passages, directly in MySQL.
- 2 sThe first answer arrives in less than this, including the call to the language model.
- 60-80%of answers come from the cache after a warm-up phase.
The cache cuts calls to the language model by a factor of 3 to 5.
The vector search runs directly in MySQL, without a separate vector database. The cache does the most to cut costs. After a warm-up phase, 60 to 80 percent of answers come from the cache, and calls to the language model drop by a factor of 3 to 5. All the details are in the WP FAQ Bot case study.
What should you watch out for on data protection?
Above all, where the language model runs. For GDPR-sensitive projects we do not use US cloud providers.
Notice in the chat window
Since 2 August 2026, visitors must be able to tell that they are talking to an AI system (Art. 50 AI Act). Please check with your legal adviser whether further obligations apply to your use.
Six points need clarifying, from hosting to the privacy policy.
- Hosting the bot on your own server or in a data centre in the EU
- The language model self-hosted or with a provider under a data processing agreement, with the transfer to third countries checked
- Indexing self-hosted or through an external service, which then also receives your website texts
- How long chat histories and the cache are kept, because visitors also type in personal details
- Only public pages in the index, no internal documents or customer data
- A notice in the privacy policy
With an external provider, at least the visitor's question and the text passages found go to its servers; with a self-hosted model, the data stays on servers you control. The WP FAQ Bot works with any API that follows the OpenAI specification. If the language model and indexing are self-hosted, for example via Ollama, no call goes to an external AI provider. Such a model does not run on simple shared hosting; it needs a dedicated server with suitable hardware.
The notice about the AI system therefore belongs visibly in the chat window. More on where to host in the guide GDPR-compliant hosting in Austria.
What should a website chatbot not answer?
Anything you would be liable for. The bot gives information from your content; it does not conclude contracts.
From your content
- Questions your website has an answer to
To your contact
- when your content has no answer
What would bind you
- no legal advice, no information on individual cases
- no binding prices or discounts
- no commitments on delivery times or dates
Before launch we test with real questions.
Language models sometimes write an answer even when your content says nothing on the subject. That is why clear rules belong in the system prompt, and why we test with real questions before the bot goes live.
Wrong answers cannot be ruled out entirely. We would rather tell you that up front.
How does a chatbot project work?
First a workshop from €1,500 or a feasibility analysis from €3,000; implementation starts at €5,000.
If you only get a handful of enquiries a week, a well-maintained FAQ page is often enough.
- from €1,500 or from €3,000
Workshop or feasibility analysis
- fixed price or time and materials
Quote
- a demo every week
Build and tests
- support optional
Go-live and operation
The workshop ends with a written outline and risk assessment.
- In the workshop or the feasibility analysis we clarify which content, which model, which hosting and which questions. We hold the workshop on site in Linz and Upper Austria or remotely. Details on both formats under Architecture workshop and feasibility analysis.
- The quote comes in writing, at a fixed price for the agreed scope or on a time and materials basis. We pass on the running costs of an external language model at cost.
- During the build we index your content, write the system prompt and test with questions from your daily work. There is a demo every week.
- At go-live we embed the bot on your website and train you. Ongoing support is an option, not an obligation.
Whether a bot pays off depends on the number of recurring questions.
Frequently asked questions
How much does an AI chatbot for a website cost?
Implementation starts at €5,000; we do not take on development projects below that.
We state the exact effort after an architecture workshop from €1,500 or a feasibility analysis from €3,000. If a ready-made chatbot tool covers your case, that is almost always cheaper. In that case we recommend it.
How does the bot know what to answer?
From the content of your website.
It searches the index for matching passages and answers on that basis. What is not on the website, it cannot know.
Can the chatbot run entirely locally?
Yes, with a self-hosted model, for example via Ollama.
If indexing and cache also run on your side, no call goes to an external AI provider. The model does need a dedicated server with suitable hardware, though; ordinary shared hosting is not enough for it. We clarify this in the analysis.
Does this work without WordPress too?
Yes. The principle carries over to other websites.
The WP FAQ Bot is a WordPress plugin. We work out in the workshop which variant suits your website.
Are there running costs?
With an external language model, you pay the provider per call, and we pass that on at cost.
The cache cuts the number of calls considerably. With a local model these costs disappear, but the hardware costs money instead.
Further reading
Let's talk.
A chatbot for your website?
Tell us what questions your visitors ask. In the initial consultation we work out whether a bot is the right answer to them, in person in Linz or remotely.