How to Add an AI Chatbot to Your Website in 2026
How to add an AI chatbot to your website in 2026: plug-in widgets vs custom LLM builds vs RAG, real costs, GDPR and EU AI Act rules, and what makes one useful.
Legacies is a software and web studio from Romania, founded by Horia Stan and Alexandru Talnaci, and adding an AI chatbot is now one of the most common requests we get from businesses with an existing website. Most of them have seen a demo that looked magical. Fewer have seen what happens when the bot confidently invents a refund policy.
There are three ways to add an AI chatbot to a website in 2026: install a ready-made widget, build a custom chat on a model API, or build a custom assistant that answers from your own content using retrieval (RAG). Widgets are fastest and cost tens to hundreds of dollars a month. Custom builds cost more up front but give you control over data, behaviour and running costs.
Quick answer
- Plug-in widget (Intercom Fin, Tidio Lyro, Chatbase): live in a day, priced per month or per resolved conversation.
- Custom chatbot on an LLM API: your own interface and logic, pay per token, usually cents per conversation.
- RAG over your content: the bot answers from your pages, docs and policies, which cuts made-up answers.
- Compliance: since August 2, 2026, EU users must be told they are talking to an AI, and GDPR applies to every chat log.
Option 1: plug-in chatbot widgets
A widget is a script you paste into your site. You point it at your website or upload documents, and it starts answering. This is the right first step for many small businesses.
Prices at the time of writing, from official pricing pages:
| Tool | Pricing model | Entry price |
|---|---|---|
| Intercom Fin | Per outcome, plus seats | From $0.99 per Fin outcome; seats from $29 a month billed yearly |
| Tidio Lyro | Monthly, by AI conversations | Lyro from $32.50 a month for 50 AI conversations |
| Chatbase | Monthly, by message credits | Hobby $40 a month for 700 credits; Standard $150 for 4,000 |
Good for: support FAQs, opening hours, shipping questions, simple lead capture.
Watch out for: per-conversation pricing that grows with traffic, limited control over tone and answers, your chat data stored on the vendor's servers, and design that does not match your site. Many widgets also add heavy scripts that slow down your pages, which hurts both users and search rankings.
Option 2: a custom chatbot built on an LLM API
Here you build the chat interface into your site and call a model API from your own backend. You choose the model, write the instructions and decide exactly what the bot can and cannot do.
Model prices per million tokens at the time of writing, from Anthropic's pricing page and OpenAI's pricing page:
| Model | Input | Output | Good for |
|---|---|---|---|
| Claude Haiku 4.5 | $1 | $5 | Fast, cheap support answers |
| Claude Sonnet 5.5 | $2 | $10 | Better reasoning, longer conversations |
| gpt-5-mini | $0.25 | $2 | Very low cost, simple tasks |
| gpt-6.1-sol | $2 | $10 | Stronger general model |
What does that mean in real money? Our own rough estimate: a typical support message with some context is about 3,000 input tokens and 300 output tokens. On Sonnet 5.5 that is under one cent per reply. At 5,000 replies a month, that is roughly $45 in model costs. On Haiku 4.5, about half that. Your numbers will depend on how much context you send.
Good for: bots that need to take actions (check an order, book a slot, create a ticket), match your brand closely, or keep data on your own infrastructure.
Watch out for: you own the reliability, the safety rules and the cost limits. Always put rate limits and a monthly spending cap on the API key. Never call the model directly from the browser, because the key would be visible to anyone.
Option 3: RAG over your own content
RAG means retrieval-augmented generation. Before the model answers, your system searches your own content, such as service pages, help articles, PDFs and policies, and passes the most relevant passages to the model. The model answers from those passages, not from its general memory.
This is what turns a chatbot from "sounds smart" into "is correct about your business".
How it works
- Your content is split into small chunks.
- Each chunk is turned into an embedding, a list of numbers that captures its meaning, and stored in a vector database.
- When a user asks something, the question is embedded and the closest chunks are found.
- The model gets the question plus those chunks, with an instruction to answer only from them and to say when it does not know.
Embeddings are cheap. OpenAI's text-embedding-3-small costs $0.02 per million tokens, so indexing a typical small business site costs cents. Postgres with the pgvector extension handles the storage, so you often do not need a separate database.
Keeping RAG honest
- Re-index whenever your content changes, or the bot will quote old prices.
- Show sources under each answer, so users can check.
- Tell the bot to say "I do not know" and offer a human handoff, instead of guessing.
- Test it with the 50 questions your customers actually ask, before launch.
Cost comparison at a glance
| Approach | Setup cost | Monthly running cost | Control |
|---|---|---|---|
| Plug-in widget | Low, often DIY | About $30 to $500+, grows with volume | Low to medium |
| Custom LLM chat | Medium, developer build | Model tokens plus hosting, often tens of dollars | High |
| Custom RAG assistant | Medium to high | Model tokens, embeddings, database | Highest |
The pattern is simple. Widgets are cheap to start and can become expensive at volume. Custom builds cost more on day one and are usually cheaper and more accurate over time. If you are weighing a bigger AI product, our guide on how much it costs to build an app in 2026 covers the wider picture.
Privacy, GDPR and the EU AI Act
A chatbot collects personal data. People type names, emails, order numbers and sometimes much more. That brings real obligations.
GDPR
If you have EU visitors, you need a lawful basis for processing chat data, a privacy policy that mentions the chatbot, and a data processing agreement with every vendor that handles the data, as required by Article 28 GDPR. Check where the vendor stores data, how long it keeps logs and whether it uses your conversations to train models. Set a retention period and delete old logs.
EU AI Act Article 50
Since August 2, 2026, Article 50 of the EU AI Act requires that people are told they are interacting with an AI system, unless it is obvious, at the latest at their first interaction. In practice, label the chat clearly, for example "AI assistant" in the header and a short line in the first message. This also applies to businesses outside the EU whose chatbot serves EU users. We cover the details, dates and fines in our EU AI Act Article 50 transparency checklist.
What makes an AI chatbot actually useful
Most website chatbots are ignored or annoying. The useful ones share a few traits.
- A narrow job. "Answer questions about our services and book a consultation" beats "help with anything".
- Grounded answers. RAG over your real content, with sources shown.
- Real actions. Checking an order or booking a slot is worth more than ten paragraphs of text.
- A clear handoff. An easy path to a human by email or ticket when the bot is unsure.
- Fast and light. The widget should not slow down your site.
- Measured. Track resolved questions, handoffs and the questions it failed on, then fix the content.
This is the same idea we argued in AI agents should start with one boring workflow. Pick one job, do it well, expand later.
Also remember that your website content now feeds AI assistants outside your site too. Clear, structured pages help your own chatbot and help ChatGPT or Google's AI answer correctly about you. More on that in how AI crawlers are changing what a business website is for.
When to do it yourself and when to bring in a team
Do it yourself with a widget when you mostly need FAQ answers, your traffic is modest and your content is already clear. Set it up in an afternoon, label it as AI, and review the conversations weekly.
Bring in a team when the bot needs to take actions in your systems, answer from a lot of your own content, keep data under your control, or stay predictable on cost. That is a small web app, and we build those in Next.js with the model behind a secure backend. Our web apps start from 3,499 lei, about EUR 665, with a fixed price after a short written chat. See our services page, and if your site itself needs work first, run the free website audit. Choosing a partner for this? Our guide on how to choose a web design agency applies here too, and you can compare the cost of the site itself in how much a business website costs in 2026.
Frequently Asked Questions
How much does it cost to add an AI chatbot to a website?
A plug-in widget usually costs from about $30 to a few hundred dollars a month, depending on conversation volume, and some charge per resolved conversation, like Intercom Fin from $0.99 per outcome. A custom chatbot needs a one-time build, then model costs that are often under one cent per reply. At higher volumes, custom builds are often cheaper to run.
Can I add ChatGPT to my website?
Yes, but not by pasting your ChatGPT account into a page. You either use a widget built on a model like GPT or Claude, or a developer calls the model API from your backend and builds a chat interface on your site. The API key must stay on the server, never in the browser.
What is the best AI chatbot for a small business website?
For simple FAQ answers and lead capture, a ready-made widget such as Tidio Lyro, Chatbase or Intercom Fin is usually enough. If the bot needs to answer accurately from many of your own documents or take actions like booking, a custom RAG chatbot works better. Choose based on the job, your volume and where your data is stored.
Do I have to tell users they are talking to an AI chatbot?
In the EU, yes. Since August 2, 2026, Article 50 of the EU AI Act requires that people are informed they are interacting with an AI system, at the latest at the first interaction, unless it is obvious. A clear "AI assistant" label in the chat is the simple way to comply.
How do I stop my chatbot from making things up?
Use retrieval-augmented generation so the model answers only from your own content, tell it to say when it does not know, and show the sources under each answer. Test it against the real questions customers ask and review failed conversations every week. Keep your content up to date, because the bot can only be as correct as your pages.
Is an AI chatbot GDPR compliant?
It can be, but compliance depends on how you set it up. You need a privacy policy that covers the chatbot, a data processing agreement with each vendor, a clear retention period for chat logs and a check on whether the vendor trains models on your data. Avoid collecting more personal data than the bot needs.