Language:

How to Train a Chatbot on Your Own Data: 4 Methods Compared

Train an AI chatbot in four ways: from pasting your website into a ready-made bot to running your own model. We tested the three no-code options with one’s store content so you don’t have to.
See what ChatGPT thinks
How to Train a Chatbot on Your Own Data: 4 Methods Compared

There are four ways to train a chatbot on your own data, and the main difference between them is the level of effort required from your side. We tried the three no-code methods with the same store’s content, and in all three, a wrong answer meant the bot was missing a fact or an instruction, never that it needed a new model.

You want a chatbot that answers customers like someone who knows your business. Out of the box, it knows nothing about your products or policies, and it doesn’t always admit it: in our tests, one bot passed off a generic cleaning tip as the store’s own advice, and another just made up its first answer.

Training a chatbot for your business most often means feeding the bot your content, so it can look things up in a knowledge base before it responds, and telling it how to behave. The simplest tools already do the retrieval part for you, so most small businesses don’t need to look beyond website chatbots.

The four approaches differ in what they ask of you: how much you set up initially and what level of maintenance is required in the long run. We tested the three no-code routes hands-on with the same 12-page store. The fourth is for developer teams, and we describe it from documentation and builders’ accounts.

TL;DR

  • A ready-made AI chatbot is enough for most businesses: you give it your website, then add what’s missing.
  • Workflow builders also start from your website, and suit you if the bot must follow multi-step procedures you design.
  • A RAG pipeline you assemble needs no code, but you connect, fix and update every part yourself.
  • Your own stack or model is for developer teams that need full control; fine-tuning changes how a model answers more than what it knows.
  • Whatever you pick, check what the bot actually read and how it picks up changed pages before you go live.

Four ways to train a chatbot on your own data, compared

Think of it this way: starting from a simple URL-paste, each following approach adds to what’s required from you to set it up. Below are the four methods and what you need to start with each:

Method
What you need
Example toolsStarting price (monthly)Time to a first working answer
Ready-made AI chatbotYour website and documentsElfsight AI Chatbot, Chatbase, Tidio Lyro$6 (Elfsight Basic)About 20 seconds to a preview
Workflow builderThe above, plus a clear list of the situations the bot must handleBotpress, Voiceflow$60 (Voiceflow Pro)Under 3 minutes of onboarding steps
RAG pipeline you assembleThe above, plus a basic grasp of APIs, AI models and retrievaln8n, Dify, OpenAI file search$24 (n8n Cloud Starter) + model useAbout an hour for a beginner
Your own stack or modelThe above, plus developer skills and your own servers or GPUsLlamaIndex + Pinecone; Hugging Face TRL, UnslothAbout $10–$25 to run a small setupWeeks of developer time
Prices on monthly billing, checked November 2026.

Some tools technically sit between two rows. For example, the AI agents built into helpdesks, such as Intercom Fin or Zendesk AI agents, learn from your help center the way a ready-made chatbot learns from your website, and several website chatbots now add step-by-step workflows too.

1. Ready-made AI chatbot: load your content, then fill the gaps

For a small business, this is often the whole job: you paste your website URL, and the tool stores your pages into a database it runs for you. Training it means making sure the right content got in, adding what’s missing, and keeping it up-to-date. No advanced knowledge of AI models is needed, and no developer.

The process below works similarly in any website chatbot or helpdesk AI, only the button names change. We show it in Elfsight AI Chatbot (that’s us), plus describe what we saw while testing six other tools given the same content: a 12-page test store, 10 test questions, one run per tool.

Step 1. Create the chatbot from your sitemap

From a single URL, the chatbot builder drafts a name, role and tone of voice for you to edit, and starts reading your pages (it can take in up to 200). Our initial preview was ready in about 20 seconds.

Open Web Pages and use View Content to see what the bot read. Watch for typos: a mistyped URL was accepted without a warning, and the bot learned the 404 page.

Tip: Check that your website scan gets everything you need it to, and add extra URLs if necessary. Given the same 12 pages, the tools in our test took in anywhere from none to all of them. HubSpot couldn’t read the pages, so got the same information as a file. Tidio needed 5 pages added by hand.

Step 2: Ask the questions your customers ask

Ask your most common customer questions, each in a fresh chat, and check every answer against your pages. In our test, a wrong or missing answer almost always meant a page the bot hadn’t synced.

Include questions on topics your content only touches. All seven bots said “I don’t know” when a fact was clearly missing, like whether the store’s glazes were lead-free. But when a topic sounded familiar, a bot could fill the gap with general knowledge: asked about tea stains, ours passed off a generic baking-soda method as the store’s own advice.

Chatbot answering from model knowledge because a fact was missing from training

If you have customers who write in other languages, ask in those too. With default settings, Zendesk’s agent replied “I don’t speak your language” to a question in Spanish, and Freddy replied in English (both have language settings you can change).

Step 3. Add what your pages don’t say

Fill each gap with the right kind of content:

  • A Q&A for a one-line fact. “Can’t confirm” on our product-safety question became a direct answer.
  • A file for a longer guide. A short care guide replaced the bot’s improvised tea-stain advice.
  • A text block for something temporary. It gave the bot the holiday closure dates it didn’t have.
Chatbot answer before and after uploading a care guide file
Chatbot answer before and after uploading a care guide file

Step 4. Write rules for behavior, not fact-checking

Rules tell the chatbot how to act in a situation: what to ask, what it may promise, when to send the customer to you; keep the facts in the knowledge base.

Elfsight chatbot answered a cracked-bowl question correctly from the policy page but didn’t ask for the order number. One Conversation Rule later, it asked for the number first, and still took the facts from the page.

Step 5. Review what the bot couldn’t answer

Most tools collect the questions the bot couldn’t answer. In Elfsight it’s Knowledge Gaps: each one shows a summary, the transcript, and three actions (Add Knowledge, Mark as Resolved, Ignore). Our three unanswered questions all appeared there, but treat the list as a starting point, not a verdict:

  • Read the transcript, not just the summary. One summary described our tea-stain question as a dishwasher question.
  • Close gaps yourself. They don’t close on their own after you add content, and a correct answer after a fix opened a new gap anyway.
  • Keep testing. A question the bot answered partly wrong never appeared in the list.
The chatbot's list of questions it couldn't answer

Step 6. Retrain a page after you change it

Training isn’t finished when the bot answers well. When a policy changes on your website, most tools keep answering from the old text until you act. We changed the store’s return window from 30 to 45 days: every tool gave the new number on the first question after its update step, but the step differed.

ToolWhat you do after a changeAutomatic update by default
Elfsight AI ChatbotClick Retrain on the changed pageNone
Intercom FinRe-sync, one clickWeekly
ChatbaseResync now, two clicksWeekly, but switched off by default
Zendesk AI agentsSync now, one clickSet to “Never”
Tidio LyroRe-sync the page; this replaces your edits to its Q&AsNone; locked for about 22 hours after the last sync
Freddy AI AgentResync and relearn the whole sourceNone; once every 24 hours at most
HubSpot’s customer agentRemove the file and upload the new versionWeb pages weekly; files never

Find this setting on day one, and make the update step part of changing a page on your website.

You can try steps 1 to 3 on your own content right here. Paste your website’s address into the editor below, open the list of pages to see what it read, then ask the question your customers ask most and add what’s missing.

Train your AI chatbot in the live editor ↓

Training the AI agent in your helpdesk

If your team already answers customers in a helpdesk, its AI agent learns mainly from your help center. The steps above still apply, with one difference: your help center articles are the training. Fix an answer by fixing the article, then check how the agent picks up the change.

Intercom’s Fin can learn from your Help Center pages and internal articles

Intercom Fin re-syncs weekly, Zendesk’s web crawler is set to “Never” by default, and Freddy relearns a source at most once a day.

When you’ll outgrow it

A ready-made chatbot is built to answer questions. When the bot must carry out multi-step work, like walking a customer through a refund with checks along the way, you need to design that flow yourself. Some tools now cover part of it: Chatbase includes procedures, Tidio Lyro calls your apps through Actions, and Intercom Fin’s Procedures handle claims like a damaged order. Check yours first.

2. Workflow builder: design how the bot handles each situation

A workflow builder starts the same way: you give it your website and files, and the vendor runs the knowledge base. Training goes further. You decide which situations the bot must handle, and design a playbook for each: the steps to follow, the tools to call, and when to hand over to a person. That takes a clear picture of your support cases, but no code.

We trained Botpress and Voiceflow with the same store’s website and files, on free plans and in preview only. Microsoft Copilot Studio and Google’s Conversational Agents (formerly Dialogflow CX) work the same way, and the steps below apply to them too.

Step 1: Give it your website and files

Enter your website, pick the bot’s job, and upload what the website doesn’t cover, such as a returns FAQ. Then check that the pages you need got in, as in the first approach. In Voiceflow, our broken-item question got only a pointer to the policy while the homepage was the bot’s only page. Once we added the damaged-items page, the answer gave the full steps.

The same question answered before and after adding the policy page

In Copilot Studio, give each knowledge source a detailed name and description: Microsoft’s docs say the description helps the agent decide which source to use.

Step 2: Read what setup wrote about your business

These tools write the bot’s instructions for you from your website: the instructions in Botpress, the global prompt in Voiceflow. Read them before anything else, because that text shapes every answer.

Botpress auto-written instructions
Botpress auto-written instructions

Voiceflow’s assistant asks for your approval at each step, which is your chance to correct it: it named our store after the subdomain in its web address and described the whole business by the size of one mug. Botpress’s auto-filled guardrails, on the other hand, already told the bot not to invent policies, prices or completed actions.

Step 3: Write a playbook for each situation

A playbook is a plain-English brief for one situation, written the way you’d explain a task to a new teammate: what to check, what to ask, what to do, and when to hand over. You can describe it yourself or generate it from your content. Botpress wrote a damaged-delivery playbook from our pages on its own, and answered our broken-item question correctly from it.

Voiceflow's auto-written refund playbook
Voiceflow’s auto-written refund playbook

Use fixed steps only where a conversation must follow a script exactly. Botpress’s docs suggest this, and Google’s console offers the same split between generative playbooks and fixed flows. The skills for writing chatbot instructions carry over to playbooks.

Step 4: Connect the tools each playbook needs

A playbook that checks an order or opens a refund needs a connection to the system that holds that data. Connect only what a playbook uses, and decide where the bot hands the conversation to a person.

Step 5: Test every playbook

Run each playbook through the cases it should handle and the ones it shouldn’t. Voiceflow’s test builder runs scripted turns or a simulated customer, with success criteria you set. Google’s console has a simulator for the same job.

What it costs

  • Botpress: Free includes 100 conversations in total and $10 of AI usage. Plus costs $189 a month for 250 conversations, and $65 for each extra 100. Our whole training session used $0.06 of the free credit; preview chats didn’t count as conversations.
  • Voiceflow: the Free plan is a one-time credit grant, not meant for production. Pro is $60 a month and Business $150.

When to move on

Every playbook needs to stay current as your policies change, and the bill grows with conversations. Move on when the pieces must live inside your own app. be paying for a canvas you leave empty. Move on when the pieces must live inside your own app.

3. RAG pipeline you assemble: connect the parts yourself

RAG (retrieval-augmented generation) is the lookup-then-answer process the first two approaches run for you. Here, you build it from parts: a store for your content, an embedding model that indexes it, a chat model that answers, a chat window, and a way to keep the content current.

You need a working grasp of how these parts fit together, well-structured content, and the patience to troubleshoot when they don’t. Visual tools like n8n and Dify need no code, whereas the API route, such as OpenAI’s file search or Amazon Bedrock Knowledge Bases for teams on AWS, does.

Our n8n build

We used n8n’s RAG Starter template, which has one workflow to upload files and one to answer questions, on an n8n Cloud trial with an OpenAI key, four short files from the same store and seven questions.

Turning it into a working bot took these steps:

  1. Import the template and select your OpenAI credential in each model node.
  2. Pick the chat and embedding models – the template recommends none.
  3. Give the upload and answer workflows the same memory key, or the chat searches an empty store.
  4. Upload your files and check the upload step actually ran (in our build, it didn’t at first).
  5. Add a system message telling the bot to search the files first and to say when it doesn’t know.
  6. Ask real customer questions, starting a fresh chat for each.
OpenAI credential added in n8n
platform.openai.com → API keys → Create new secret key → paste into n8n

It took around 60 minutes and 8 fixes to the first correct answer, and no code. One fix meant decoding an error message from the model’s API. As imported, the bot never searched the files and made up its first answer until the system message fixed it.

Chatbot answers from model training and after knowledge base fix

Keeping it current

Updates were the weak spot. When we changed a file, both the old and the new version stayed in the store, and the bot gave the new answer only after we re-ran the upload by hand. The template’s store also lives in memory, which n8n marks “for experimental use only”: it’s lost if n8n restarts.

Builders on Reddit describe the same work: one who builds RAG systems says most of it is data hygiene, and another needed separate workflows just to push changed records into the store.

What it costs

  • n8n: Cloud Starter costs $24 a month. The self-hosted Community Edition is free.
  • Our test day: $0.0035 in model fees, with four short files.
  • A real website, our estimate: about $0.52 per 1,000 answers on GPT-6 Luna, $1.10 on GPT-5.6 Luna and $5.20 on GPT-6.1 Terra. This assumes about 3,700 input and 300 output tokens per answer, more than our test used, because a real website’s pages are longer, so more text is looked up per answer.
  • OpenAI’s file search, if you take the API route: $2.50 per 1,000 searches, plus storage.
  • Indexing a 200-page website: under a cent, at about 1,000 tokens a page, by our estimate.

The model bill stays small. The real cost is your time: building a chat window for your website, a handoff to a person and an update routine all come on top of our build, and the work continues at every content change.

When to move on

This route strains on updates, on a store that forgets everything on restart, and on every part you build yourself. Move on when you need a permanent store, your own hosting, or a change to the model itself.

4. Your own stack or model: run every part yourself

Everything the RAG pipeline asks, plus developer skills and your own infrastructure. Nobody hands you a working sandbox: you pick the components, host them, and keep them running. This approach splits into two different jobs. Running your own RAG stack still uses your ordinary content, like your website and documents. Fine-tuning changes the model itself, and needs a training dataset built for that purpose plus GPU hardware.

Your own RAG stack

A framework such as LlamaIndex splits the job into five stages: load your data, index it, store the index, query it, and evaluate the answers. You choose and run each part. The database for your content can start free: Pinecone’s Starter plan includes 2 GB of storage. On cost, forum users report keeping their setup under $25 a month at about a million queries, but that’s one person’s setup, not a benchmark.

Own RAG workflow creation

This makes sense when your data must stay on your own servers, or when you need control the hosted approaches don’t give you.

Fine-tuning an open model

OpenAI is winding down self-serve fine-tuning, so fine-tuning now means an open model you train yourself.

Whether fine-tuning can teach a model new facts is disputed. OpenAI treats knowledge a model lacks as a retrieval job and uses fine-tuning for consistent behavior. Unsloth, which makes a fine-tuning tool, says fine-tuning can add knowledge too.

For a business bot, the practical line is: fine-tune for how it answers, and keep the facts in retrieval. Even where fine-tuning adds knowledge, a changed policy means another training run, while a retrieval bot needs one updated document.

Fine-tuning takes three things:

  • A dataset, usually question-and-answer pairs. A support bot’s comes from your own tickets and docs; people-focused bots, like HR or recruiting assistants, can start from bought datasets, such as the employee data Coresignal sells for AI training among other uses.
  • A training tool, such as Hugging Face’s TRL or Unsloth Studio, which needs no code. Unsloth’s LoRA method trains about 1% of the model’s weights.
  • Hardware: small models run free on Colab, or on a GPU with as little as 3 GB of memory, per Unsloth. For bigger runs, Cherry Servers lists an NVIDIA A100 80 GB server from $2.362 an hour.

Who should consider it

Developer teams that need full control over where data lives, or a model that answers in a set format no instruction can hold. There’s no reliable public figure for a whole project’s cost or time; the bill is built from unit prices for the database, GPU hours and people. Lemon.io’s guide to types of AI engineers lists integrators, who build RAG on existing models, at $90–$180 an hour, and machine learning engineers, who train models, at $120–$250+.

How to pick the right way to train your chatbot

Find the row that describes you:

If this is youStart hereMove up when
You answer customers yourself, with no helpdeskReady-made AI chatbot: Elfsight AI Chatbot, Chatbase, Tidio LyroThe bot must run multi-step tasks in other systems
Your team already works in a helpdeskIts own AI agent: Intercom Fin, Zendesk AI agents, Freddy AI Agent, HubSpot’s customer agentThe bot must run multi-step tasks its procedures can’t cover
You know which situations the bot must handle, step by stepWorkflow builder: Botpress, VoiceflowThe pieces must live in your own app or stack
You have someone technical and want to own each partRAG pipeline you assemble: n8n, Dify, OpenAI file searchYou need a permanent store, your own hosting or a model change
Your data must stay on your servers, or answers must follow a set formatYour own stack or model—

You move up because of what the bot has to do, not how many pages it reads: more content rarely needs a bigger setup, more behavior might. Whichever approach you choose, the same checks come before launch: read what setup wrote and which pages got in, ask your customers’ real questions in new chats, and find the refresh setting. In our tests, those checks caught a typo that trained a 404 page, a misread business description, and content syncs that ran weekly or never.

Frequently asked questions

Can I train ChatGPT with my own data?

Not for long through custom GPTs: OpenAI is retiring custom GPTs and its builder tools by the end of 2026. For a chatbot that answers visitors on your website, use a tool built for that job, from a ready-made chatbot to your own stack.

How much content do I need to train a chatbot?

Less than you might expect, if it covers what people actually ask. A 12-page store website and four short files were enough in our tests, once every question we asked had an answer somewhere in them.

Does a chatbot learn from its conversations on its own?

Not a ready-made one like Elfsight AI Chatbot. It answers only from the content you provide and doesn’t update its knowledge from individual user chats. Instead, it shows you the questions it couldn’t answer, so you decide what to add and what to keep out.

Do I need to fine-tune a model to get my brand's tone of voice?

Usually not. In the tools we tried, tone is a setting: Elfsight drafts one from your website, and Botpress offers presets such as Friendly and Professional. OpenAI’s guide suggests fine-tuning only when a model’s format or tone stays inconsistent, and says many of its largest deployments used only instructions and retrieval.

How do I test a chatbot before it goes live?

Ask it your customers’ real questions, each in its own chat, and compare every answer with your pages. After each fix, ask again in a fresh chat. Testing needn’t cost anything: Elfsight’s first 100 test messages in the editor are free, and Botpress didn’t count preview chats as conversations.

Where to start training your chatbot

On every approach we tried, a wrong or missing answer came down to a fact the bot didn’t have or an instruction it hadn’t been given. To train your own AI chatbot, start small: list the questions customers ask most, paste your website into a ready-made chatbot’s free plan, ask each question in a new chat, and add content for every miss.

Once it answers those questions well, it’s ready for real customers: our step-by-step tutorial shows how to add the AI chatbot to your website.

Key references

Article by
AI Content Specialist
Kristina covers AI topics at Elfsight and Beamtrace: she writes about AI chatbots, LLM visibility, and how AI is reshaping search and customer experience – with practical takes for business and website owners. You can read more about her work at kristinaty.me.