AI Chatbot Trained on Your Data: How to Train It in 7 Steps

Created by

Do not index

Created time

Jul 25, 2025 07:35 AM

An AI chatbot trained on your data is a chatbot that answers from your own pages and files, and you train one in seven steps: list the questions, gather the content, clean it, structure it, load it into a tool, test it, and retrain it on a schedule.

TL;DR

"Training" a chatbot on your data means indexing your content so a language model can look it up at answer time. The model itself does not change.

Start from the questions customers ask. Gather only the pages and files that answer them.

Clean before you upload. Remove duplicates, outdated pages, and navigation text, and keep one fact in one place.

Structure pages so each section stands alone. The tool cuts content into chunks, and a chunk that needs the paragraph above it produces bad answers.

Test with 30 real questions before launch, and retrain every time the source content changes.

Key takeaways

Step	What you do	How to check it
1. List the questions	Collect 50 to 100 real customer questions	Every question has a named topic
2. Gather the content	Find the page or file that answers each question	No question is left without a source
3. Clean the content	Remove duplicates, old versions, and page clutter	Each fact appears once, with one value
4. Structure for chunking	Use descriptive headings and self-contained sections	Any section makes sense when read alone
5. Load it into a tool	Add URLs and files to a chatbot platform or your own retrieval code	The indexed page count matches what you sent
6. Test before launch	Run a fixed set of 30 questions and score the answers	The score is written down, with every failure traced to a cause
7. Retrain on a schedule	Refresh the index when content changes	A changed price shows up in the bot's answer

Summary graphic of the seven steps to train an AI chatbot on your own data: list questions, gather content, clean, structure for chunking, load into a tool, test, and retrain

What is an AI chatbot trained on your data?

An AI chatbot trained on your data is a chatbot that answers questions from content you supply, such as website pages, help articles, PDFs, and policy documents. A general language model writes the reply. Your content supplies the facts.

The word "training" is misleading here. In almost every business chatbot, nothing is trained in the machine-learning sense. The tool splits your content into passages, stores them in a searchable index, and retrieves the most relevant passages each time someone asks a question. This method is called retrieval-augmented generation, and what is RAG explains the full loop.

Approach	What happens to your data	Does the model change?	When content changes
Retrieval (what most tools do)	Content is split into chunks, indexed, and looked up at answer time	No	Re-index the changed pages
Fine-tuning	Example questions and answers adjust a copy of the model	Yes	Train again
Instructions only	Nothing is stored. You write rules for tone and scope	No	Edit the instructions

This matters for the work ahead. Because the bot can only retrieve what was indexed, the quality of the content decides the quality of the answers. The seven steps below are mostly content work.

If the goal is to give ChatGPT itself your files through a project, a custom GPT, or the OpenAI API, the guide to training ChatGPT on your own data compares those five methods with their costs.

Step 1: List the questions the chatbot must answer

The first step in training a chatbot on your own data is to list the questions it must answer. The question list decides which content you need, and it becomes the test set in step 6.

Export the last 50 to 100 questions from the places customers already ask: the shared inbox, live chat transcripts, the contact form, and sales calls.

Group them by topic, such as pricing, shipping, returns, setup, and account access.

Mark each question as in scope or out of scope. A question that needs account data, a refund decision, or legal advice is out of scope for a content-trained bot.

Count the questions in each topic. The largest topics get content first.

Keep the list in a spreadsheet with one row per question. Step 2 adds a column to it.

Step 2: Gather the content that answers those questions

Gathering content means finding the one page or file that answers each question on the list. Add a "source" column to the spreadsheet and fill it in for every in-scope question.

Content type	Include it?	Reason
Help center and FAQ pages	Yes	Already written as questions and answers
Pricing, shipping, and returns pages	Yes	These are the highest-volume topics for most businesses
Product and service pages	Yes	They hold specifications and feature details
Policy PDFs and manuals	Yes, if the text is selectable	Scanned pages may be read as images with no text
Past chat or email replies	Only after rewriting	They contain customer names and one-off exceptions
Blog posts and news	Rarely	Old posts carry old prices and retired features
Internal notes and drafts	No	Anyone who can chat with the bot can reach what it was given

A question with an empty source column is a content gap. Write a short answer for it now. A bot with no source for a question will either decline or guess.

There is no minimum amount of data. Retrieval works with a single page. Coverage of the question list matters more than the number of pages.

Step 3: Clean the content before you upload it

Cleaning content means removing anything that could make the chatbot retrieve the wrong passage. A retrieval system cannot tell which of two conflicting pages is current. It returns whichever passage matches the question best.

Run these five checks on every source.

Duplicates. Keep one version of each page. Remove print views, tag archives, and copies of the same PDF.

Conflicts. Search for every price, time limit, and plan name. Each one should have a single value across all sources.

Outdated content. Remove pages for retired products, expired offers, and old policy versions.

Page clutter. Check that menus, cookie banners, footers, and sidebars are not captured as content. Most platforms strip these from web pages. Spot-check a few pages after indexing.

Private data. Remove names, email addresses, order numbers, and internal comments.

Before and after comparison of chatbot training content: duplicates, old pricing pages, and page clutter removed so each fact appears once

Files need one more check. OpenAI's file search documentation lists the text formats its index accepts, including PDF, DOCX, PPTX, HTML, Markdown, JSON, and plain text. Spreadsheets with merged cells and PDFs made from scans are the usual problem files. Convert them to plain text or a simple table first.

Step 4: Structure the content so it chunks well

Structuring content for chunking means writing each section so it still makes sense after the tool cuts the page into pieces. A chunk is a passage of text that the index stores and retrieves as one unit.

Chunk sizes are set by the tool. OpenAI's vector stores, for example, split every file into 800-token chunks with a 400-token overlap between consecutive chunks by default. The allowed chunk size runs from 100 to 4,096 tokens. Most no-code platforms choose the size for you.

You control the writing. These rules make any chunk size work better.

Use headings that name the topic. "Refund policy for annual plans" retrieves better than "More information".

Keep one topic per section. A section that covers returns, exchanges, and warranty will match all three questions weakly.

Repeat the subject. Write "The Pro plan includes 5 seats" instead of "It includes 5 seats". A chunk that starts with "It" has lost its subject.

Keep an exception next to its rule. "Returns are accepted within 30 days, except for sale items" should stay in one sentence.

Turn tables into sentences where a row must stand alone. "Standard shipping to Canada takes 5 to 8 business days" survives chunking. A table cell that says "5 to 8" may not.

Write questions as headings in FAQ content. The customer's wording then sits next to the answer.

Worked example: sizing the content for one small business

This example shows the arithmetic for a hypothetical business with 60 website pages and 4 policy PDFs.

Item	Calculation	Result
Website text	60 pages at 3,000 cleaned characters each	180,000 characters
PDF text	4 files at 20,000 characters each	80,000 characters
Total content	180,000 + 80,000	260,000 characters
Size in SiteGPT pages	260,000 divided by 2,500 characters per page	104 pages, which is 10.4% of the 1,000 pages on Starter
Approximate tokens	260,000 divided by 4, assuming about 4 characters per token	65,000 tokens
Approximate chunks at OpenAI's defaults	65,000 divided by a 400-token step between chunks	About 163 chunks

The character counts and the 4-characters-per-token figure are assumptions. The result is still useful: a typical small business site becomes a few hundred chunks, and one bad page is a visible share of that index.

Step 5: Load the content into a tool

Loading the content means adding your URLs and files to a system that chunks, indexes, and retrieves them. There are two routes: a no-code chatbot platform, or retrieval code you write yourself.

How the tools were ranked

The three platforms below are ranked for one scenario: a small business that wants a website chatbot trained on its own pages and files, with no developer. Four criteria decided the order.

Whether you can test with your own data before paying.

How much content the entry paid plan indexes.

Whether content refresh is manual or automatic.

The entry paid price, billed monthly.

No tool is best at all four. Each one has a "best for" label.

Tool	Best for	Free option	Entry paid plan, billed monthly	Content limit on that plan	Monthly allowance on that plan
1. Chatbase	Testing with your own data at no cost	Free plan: 50 message credits a month, 1 MB of training content	Hobby, $40 a month	10 MB of training content	700 message credits
2. SiteGPT	A website chatbot fed from many source types	No free plan. 7-day trial, credit card required	Starter, $59 a month ($39 a month billed yearly)	1,000 pages of 2,500 characters each	4,000 GPT-4.1-mini messages or 400 GPT-4.1 messages
3. DocsBot	Large documentation sets on a low budget	Free plan: 50 source pages, 100 AI credits a month	Personal, $49 a month	5,000 source pages	5,000 AI credits

Bar chart of entry paid plan prices billed monthly: Chatbase Hobby at $40, DocsBot Personal at $49, and SiteGPT Starter at $59

1. Chatbase: best for testing with your own data at no cost

Chatbase ranks first for this scenario because its Free plan lets you index real content and ask real questions without paying. The Free plan includes 50 message credits a month and 1 MB of training content.

Paid entry point: Hobby costs $40 a month billed monthly. It includes 700 message credits, 10 MB of training content, and 2 members.

Automatic refresh: "Auto-resync agents" starts on Standard at $150 a month billed monthly.

Limit: Chatbase deletes agents on the Free plan after 14 days of inactivity. Removing the "Powered By Chatbase" label costs $99 a month.

The full breakdown is in the Chatbase review.

2. SiteGPT: best for a website chatbot fed from many source types

SiteGPT ranks second because it has no free plan, and it fits best when the content lives in more than one place. Sources include website URLs and sitemaps, uploaded files, YouTube, GitHub, cloud storage such as Google Drive and Notion, and help centers. When the content sits mainly in Notion, the guide to Notion chatbots compares connectors and sync schedules.

Paid entry point: Starter costs $59 a month billed monthly, or $39 a month billed yearly ($468 a year). It includes 1 chatbot, 1,000 pages, and 4,000 GPT-4.1-mini messages or 400 GPT-4.1 messages a month.

Automatic refresh: Starter is manual refresh only. Growth adds a monthly auto refresh at $129 a month billed monthly, or $79 a month billed yearly. Scale adds a weekly auto refresh and a daily auto scan.

Limits: The 7-day trial requires a credit card. Starter has no chat integrations and no API access. Removing SiteGPT branding is an add-on at $39 a month billed yearly.

Yearly billing on Starter costs $468 against $708 for twelve monthly payments. That saves $240 a year, or 34%.

Try SiteGPT free for 7 days

3. DocsBot: best for large documentation sets on a low budget

DocsBot ranks third for a general business website and first on content volume per dollar. Its Personal plan indexes 5,000 source pages for $49 a month billed monthly, with 5,000 AI credits a month.

Free option: The Free plan includes 1 bot, 50 source pages, and 100 AI credits a month, with no credit card.

Next plan up: Standard costs $149 a month billed monthly and includes 3 bots, 15,000 source pages, and 15,000 AI credits.

Limit: DocsBot states that AI credit usage starts at 1 credit per message and can be higher for larger models and AI actions. Unbranded chat widgets are listed on the Business plan at $499 a month billed monthly.

The full breakdown is in the DocsBot review.

What the entry plans cost per 1,000 messages

Dividing each entry plan's monthly price by its allowance gives a rough unit cost. Each vendor defines its own unit, so treat the figures as a rough guide.

Plan	Price, billed monthly	Allowance	Cost per 1,000 units
Chatbase Hobby	$40	700 message credits	$57.14
SiteGPT Starter	$59	4,000 GPT-4.1-mini messages	$14.75
DocsBot Personal	$49	5,000 AI credits	$9.80

On yearly billing, SiteGPT Starter works out to $9.75 per 1,000 GPT-4.1-mini messages ($39 divided by 4).

Building it yourself

Writing your own retrieval code fits a team with developers that wants answers inside its own product. You choose the chunk size, the index, and the model. You also build the chat interface, the refresh job, and the handoff to a person. Before you commit developer hours, check the first-year totals for each route in the cost to build a chatbot.

The costs and steps for the OpenAI route are in the guide to training ChatGPT on your own data. The broader build-or-buy trade-off is covered in custom AI chatbot.

How to choose a tool

If you need...	Pick...	Why
To test with your own content before paying anything	Chatbase Free or DocsBot Free	Both have a free plan. SiteGPT has a 7-day trial that needs a card
One bot fed from a website plus Google Drive, Notion, YouTube, or a help center	SiteGPT	Many source types feed one index
Thousands of documentation pages on the lowest paid plan	DocsBot Personal	5,000 source pages for $49 a month billed monthly
Content that refreshes without anyone clicking a button	SiteGPT Growth or Chatbase Standard	Auto refresh starts at $129 and $150 a month, both billed monthly
Answers inside your own app, with full control over chunking	Your own retrieval code	Platforms set the chunk size for you
A simple FAQ bot with a short list of questions	Any of the three on its entry plan	A short FAQ fits every content limit listed above

Decision tree for choosing a tool to train a chatbot on your own data, based on coding skills, free testing, where content lives, and how often it changes

For a bot that only answers a fixed FAQ list, the shorter walkthrough in how to build an AI FAQ chatbot is the better starting point. More platforms are compared in best no-code chatbot builders.

Step 6: Test the chatbot before launch

Testing a trained chatbot means running a fixed set of questions and scoring every answer against the source content. Use the question list from step 1.

Build a test set of 30 questions in three groups.

Group	Number	What a pass looks like
Answerable questions, in the customer's own words	20	The answer matches the source page
Questions the content does not cover	5	The bot says it does not know and offers a person
Edge cases: exceptions, old product names, two questions in one message	5	The answer includes the exception, or the bot asks for clarification

Score each answer as pass or fail. A bot that passes 24 of 30 scores 80%. Record the score so the next round has a baseline.

Then trace every failure to its cause.

What went wrong	Likely cause	Fix
The bot says it does not know, and the answer exists	The page was not indexed, or the wording differs from the customer's	Check the indexed sources. Add the customer's wording as a heading
The bot gives an old price or policy	An outdated page is still in the index	Remove the old page and re-index
The bot leaves out an exception	The exception sits in a different section from the rule	Move the exception into the same sentence as the rule
The bot answers a question it should decline	The instructions do not tell it to stay within the content	Add an "answer only from the provided content" rule
The bot mixes two products	Sections do not name the product they describe	Repeat the product name in each section

Fix the content first and the instructions second. Rerun all 30 questions after each change, because a fix for one answer can break another. The guide to chatbot hallucinations covers the guardrails for invented answers, and chatbot best practices covers when the bot should hand off to a person.

Step 7: Retrain the chatbot when the content changes

Retraining a retrieval chatbot means re-indexing the sources that changed. No model is retrained. The bot repeats whatever is in its index, so an index that is three months old gives three-month-old answers.

Retrain on every change that affects a customer. That includes prices, plans, policies, hours, and product launches.

Set a schedule for everything else. Monthly suits a site that changes rarely. Weekly suits a catalog or a help center that changes often.

Read the unanswered questions. Each week, review the questions the bot could not answer and write content for the common ones. To pull that list by asking Claude, connect the chatbot over MCP with the steps in building custom AI chatbots with ChatGPT and Claude.

Rerun the test set after each retrain. Compare the score with the last baseline.

Delete what is retired. Removing a page from the website does not always remove it from the index. Check the source list.

Ongoing tuning after launch, including how to use real transcripts as new training material, has its own guide in chatbot training.

Common mistakes when training a chatbot on your own data

The most common mistakes come from treating training as an upload task when it is a content task. Five show up often.

Uploading everything. More pages mean more conflicts. Index what answers the question list.

Skipping the cleanup. Two pages with two prices give the bot a coin flip.

Testing with your own wording. Staff use internal product names. Customers do not. Test with the customers' words.

Leaving out the "I do not know" rule. Without it, the model fills gaps from general knowledge.

Training once. The launch-day index goes stale with the first price change.

Frequently asked questions

Basics

Can you train a chatbot on your own data?

Yes. Most chatbot platforms let you add website URLs and files, and the bot then answers from that content. The platform indexes your content for retrieval. It does not change the underlying language model.

How do you train an AI chatbot on your own data without code?

Without code, use a chatbot platform. Add your website URL or upload files, write instructions for tone and scope, test the answers, and embed the widget on your site. Chatbase, SiteGPT, and DocsBot all work this way.

How does an AI chatbot trained on your data work?

The platform splits your content into chunks and stores them in an index. When a visitor asks a question, the system retrieves the closest chunks and passes them to a language model, which writes the answer from them.

Data

How much data is required to train a chatbot?

There is no minimum for a retrieval chatbot. One page is enough to start. The useful measure is coverage: every question on your list should have a source that answers it.

What data should you use to train a chatbot?

Use help center articles, FAQ pages, pricing and policy pages, product pages, and manuals with selectable text. Leave out drafts, internal notes, and anything with customer names or order details.

Do you need to chunk the data yourself?

No, on a no-code platform. The platform chunks the content for you. With your own code you set the chunk size. OpenAI's vector stores default to 800-token chunks with a 400-token overlap.

Is your data used to train the underlying AI model?

That depends on the vendor, so check each vendor's data policy before uploading. SiteGPT states on its home page that customer data is never used to train AI models.

Cost and time

How much does it cost to train a chatbot on your own data?

It costs $0 on the Chatbase or DocsBot free plans, within small limits. Entry paid plans, billed monthly, are $40 a month for Chatbase Hobby, $49 for DocsBot Personal, and $59 for SiteGPT Starter. SiteGPT Starter is $39 a month when billed yearly.

Can you train a chatbot on your own data for free?

Yes, within limits. Chatbase Free includes 50 message credits a month and 1 MB of training content. DocsBot Free includes 50 source pages and 100 AI credits a month. SiteGPT has no free plan. It has a 7-day trial that requires a credit card.

How long does it take to train a chatbot?

Indexing is the short part. The content work in steps 1 to 4 takes most of the time, and it depends on how many pages need cleaning. Plan the testing round as a separate block before launch.

Accuracy and updates

How do you keep a trained chatbot up to date?

Re-index the content whenever it changes. Some plans do this automatically. SiteGPT refreshes monthly on Growth and weekly on Scale, and Chatbase lists auto-resync from its Standard plan. On other plans, refresh manually.

Why does a trained chatbot still give wrong answers?

Wrong answers usually come from the content. Common causes are an outdated page still in the index, two pages that disagree, or an exception separated from its rule. Retrieval lowers the rate of invented answers. It does not remove them.

Can a chatbot be trained on a website URL?

Yes. Platforms such as Chatbase, SiteGPT, and DocsBot crawl a URL or sitemap and index the pages they find. Check the indexed page list afterward, because pages behind a login or built entirely in JavaScript can be missed.

Sources

All sources were checked on October 5, 2026.

Chatbase pricing

DocsBot pricing

SiteGPT pricing

SiteGPT home page

OpenAI API docs: Retrieval

OpenAI API docs: File search

AI Chatbot Trained on Your Data: How to Train One in 7 Steps