A Free AI Chatbot Trained on Your Own Data: What That Really Means
"Trained on your data" sounds like you are building your own AI. You are not, and that is good news. What actually happens is much simpler, much faster to fix, and much easier to check. Here is the plain version, including the parts that go wrong.
The short version
- It reads, it does not learn. The bot looks up your pages at the moment someone asks a question.
- Your website is the knowledge base. If a page is wrong, the answer is wrong, in about a minute.
- Fixing an answer means fixing a page. That is the biggest practical advantage over retraining a model.
- Disclosure: EzyConn is our own product. Its free plan trains by crawling your website and includes 500 messages a month.
What "training" means here
Start with the word most people get stuck on. Fine-tuning means changing a model's own internal settings by showing it thousands of examples. It is expensive, slow, and awkward to correct. Almost no website chatbot does it, and yours should not.
What website chatbots do instead is called retrieval. The tool visits your pages, saves the text in a searchable store called an index, and waits. When a visitor asks a question, it finds the few passages that look most relevant and sends them to the AI along with the question. The AI then writes an answer using that text. Practitioners call this grounding: keeping the answer tied to a source you control.
So the model has not memorised your business. It is reading a handful of your own paragraphs each time, the way a new hire would read the handbook before replying. That is why a chatbot can be live an hour after signup, and why updating a page updates the answers.
| Approach | What happens | What it costs you | Good for |
|---|---|---|---|
| Reading your pages (retrieval) | The tool crawls your site, stores the text, and looks up the closest passages before answering | Minutes to set up, and it is what free plans mean by trained | Support answers that change whenever your website changes |
| Fine-tuning a model | The model itself is adjusted using thousands of example conversations | Weeks, real money, and a data science habit | A house writing style, or a very unusual task |
| Instructions in the prompt | A short set of house rules sent with every question | Free, and takes about ten minutes to write | Tone, refusals, and when to hand over to a person |
Most good setups use the first and third together. Retrieval supplies the facts. A short instruction block supplies the manners. If you want the longer technical comparison, we wrote one on fine-tuning versus retrieval.
What to feed it, and what to keep well away
The instinct is to give the bot everything. Resist it. More text is not better text, and a bot with too much to search through gets vaguer, not smarter. Here is the split we would use.
| Source | Verdict | Why |
|---|---|---|
| Pricing and plan pages | Feed it | The single most asked question, and it is already public |
| Shipping, returns and refunds | Feed it | Long, boring, and exactly what people ask at 11pm |
| Product and service pages | Feed it | Gives the bot the words your customers already searched for |
| Help articles and FAQs | Feed it | Written as question and answer, so retrieval works well |
| Opening hours and contact details | Feed it | Cheap to include, and wrong answers here look careless |
| Blog posts older than two years | Think twice | Old claims resurface as current advice |
| Customer records and past tickets | Keep out | One retrieval mistake becomes a privacy incident |
| Negotiated or per-customer pricing | Keep out | The bot will quote one customer the other one deal |
| Internal escalation notes | Keep out | Written for colleagues, not for the person complaining |
| Anything under an NDA | Keep out | A retrieval bot cannot reliably keep a secret it holds |
The rule underneath the table is short: if it is not fit to be a public page, it is not fit for the bot. A retrieval system decides what to surface based on how similar the text looks to the question. It has no sense of who is asking or why. Assume that anything in the index can be said out loud to a stranger, because one day it will be.
How to check the answers before customers do
Write down twenty questions before you launch, and keep the list. Ask them again after every content change. It takes fifteen minutes and it catches nearly everything. Five of the twenty should be these.
| Type | Example | A good answer looks like |
|---|---|---|
| The easy one | What are your opening hours? | Exact hours, no hedging, no invented holiday closures |
| The buying one | How much is the mid plan? | The current number, or a link, never an estimate |
| The awkward one | Can I get a refund after 40 days? | Your actual policy, including the word no if that is the answer |
| The missing one | Do you ship to Iceland? | An admission that it does not know, plus a handover |
| The trick one | My friend said you do free installs, right? | A correction, not agreement with the premise |
The fourth and fifth rows matter most. A chatbot that invents an answer when it has no source is doing what people call hallucinating: producing confident text that is not grounded in anything real. The fix is not clever wording. It is telling the bot to say it does not know and pass the chat to a person, then testing that it actually does. If your tool cannot be made to admit ignorance, that is a reason to choose a different tool.
When your content is wrong or out of date
This is the part nobody warns you about. A retrieval chatbot is a mirror. It repeats what your website says, in a confident voice, to everyone who asks. Every stale price, every policy you changed but never rewrote, every page from a product line you dropped, becomes a live answer.
Old blog posts are the usual offender. A post from three years ago announcing a promotion reads, to a retrieval system, exactly like a current fact. So does an out-of-date help article that nobody deleted because nobody remembered it existed. If two pages disagree, the bot may pick either one, and it will not tell you it noticed the conflict.
Three habits fix most of this. Put a review date on any page with a price or a policy in it. Delete or clearly date old announcements rather than leaving them floating. And re-crawl after any content change, then run your twenty questions again. Our walkthrough on training a chatbot on your website goes step by step.
Where your data goes, and what free actually covers
Being upfront: EzyConn is our product. Our free plan is $0 forever with no card at signup and no expiry date. It covers 2 seats, 500 messages a month that reset monthly, one chatbot with a chat widget for your site, and the mobile app with notifications. The AI runs on GPT-4o and Claude, and it is trained by crawling your website.
Five hundred messages a month is roughly sixteen a day. That is enough to judge answer quality honestly, which is the thing you cannot judge from a demo. Slack, Microsoft Teams, WhatsApp and Zapier all work on the free plan. Paid plans are Starter at $25 a month and Professional at $95 a month, with 20% off for annual billing. The pricing page has the full list.
On the question everyone should ask any AI vendor, which is what happens to your data: if keeping your data out of model training is a contractual requirement for you rather than a preference, the Enterprise plan comes with a data processing agreement. Confirm it in writing before you route real customer conversations through any AI tool.
Other tools for a free start, compared
Two other free tiers are worth understanding before you compare. Crisp's free tier includes no AI at all, so you cannot train anything on it; treat it as human live chat, and note that watermark removal only starts on the Plus plan at $295 a month. Tidio's free tier does include 50 Lyro AI conversations, but they are described as a one-off allowance rather than a monthly reset, so free AI stops permanently once they are used. Removing Tidio branding costs roughly $16.67 a month.
Intercom has no free tier at all. It charges $39 per seat per month, and its Fin agent bills $0.99 for each resolution on top of that. For any vendor we have not verified, we describe how they bill rather than quoting a number, because an invented figure is worse than no figure. We keep a broader shortlist on our best free AI chatbot page.
One last piece of context. When we scanned thousands of small business websites, only 7.5% had any chat at all. Reading a week of real transcripts often teaches you more than the bot does. The numbers are in our live chat adoption study.
Common questions
What does it actually mean to train an AI chatbot on your own data?
For almost every website chatbot, it means the tool reads your pages, stores them in a searchable index, and looks things up before it answers. The model itself is unchanged. Your content is pasted into the question behind the scenes, so the answer comes from your words rather than the model's memory of the internet.
Is this the same as fine-tuning a model?
No. Fine-tuning adjusts the model's own settings by showing it thousands of examples, which is slow, costly, and hard to correct. Retrieval just looks up your pages at the moment someone asks. Retrieval is what you want for support, because you fix a wrong answer by fixing the page, not by retraining anything.
Can I train a chatbot on my data for free?
Yes. EzyConn is our product, and its free plan crawls your website, runs on GPT-4o and Claude, and gives you 500 messages a month on one chatbot, with no card at signup.
What should I keep out of a chatbot knowledge base?
Anything you would not put on a public page. That means customer records, pricing you negotiate case by case, internal escalation notes, staff phone numbers, unreleased plans, and anything covered by a confidentiality agreement. A retrieval chatbot has no reliable way to keep a secret it has been handed.
Does EzyConn use my conversations to train AI models?
If keeping your data out of model training is a requirement for your business rather than a preference, put it in writing: the Enterprise plan comes with a data processing agreement. Check this with any AI vendor before you send real customer conversations through a tool.
Point it at your website and see
The free plan crawls your pages, runs on GPT-4o and Claude, and gives you 500 messages a month. No card, no expiry date. Ask it your twenty questions before anyone else does.
Start freeEzyConn plan details are current as of August 19, 2026. Competitor pricing and free tiers change often, so check with the vendor. We quote only figures we have verified.