Best AI Chatbots Trained on Your Own Data (2026): Limits Compared

Every tool here promises a chatbot "trained on your data". The differences sit in the small print: which sources it reads, how much text each plan holds, how often it re-reads your pages, and whether visitors see where an answer came from. This roundup puts those limits side by side, in each vendor's own unit.

It is for small business owners and support leads who want a custom ChatGPT-style assistant that answers from their own pages, PDFs, Notion or Google Drive. You will be able to pick a tool by the shape of your data and estimate the monthly bill.

PepoChat is our product; we include it where it fits and say where other tools are the better choice. Facts below were checked on vendor pages on 11 November 2026.

What does "trained on your own data" mean for a chatbot?

A chatbot "trained on your own data" almost never retrains the AI model; it stores your content in a searchable index and looks up the most relevant passages before it writes each answer.

Retrieval-augmented generation (RAG) is the method behind this: the tool splits your pages and files into short passages, indexes them, retrieves the best matches for each question and asks the model to answer from those passages only. Our RAG explainer for non-engineers walks through both phases, and what a knowledge base chatbot is covers what it cannot do.

Three consequences follow. A plan's "training" cap is a storage cap on extracted text. An edit reaches the bot only when the source is re-read, so automatic re-sync matters. And a scanned PDF with no text layer is invisible unless the tool reads it with a vision model.

Quick comparison

PepoChat, CustomGPT.ai and DocsBot lead this list for different data shapes: a small business site with files and wikis, a very large cited library, and a documentation set with help desk tickets.

Prices and features from each vendor's own pages, checked 11 November 2026, in US dollars on monthly billing unless marked.

Training limits compared: what each chatbot can read and how much

The eight tools differ most on two things: whether they read cloud wikis, help desk content and video, and how they measure size, in pages, characters, words, megabytes or documents.

Prices and limits from each vendor's pricing page and docs, checked 11 November 2026.

To compare the units, remember that a megabyte of plain English text is roughly a million characters. SiteGPT's Starter page allowance is about 2.5 million characters, DocsBot's Personal allowance about 25 million, and Chatbase's 40 MB about 40 million. CustomGPT.ai counts documents and words, and Botpress counts storage, so neither converts cleanly.

How we chose

We ranked the eight tools by what each is best for, using five criteria.

1. What it reads: sites, files, Notion and Drive, help desk content, video and audio.
2. How much it holds on the free and first paid plan, in the vendor's unit.
3. How it stays current: automatic re-sync, and from which plan.
4. Whether visitors see sources behind an answer.
5. What happens after the answer: handoff, actions and price per answer.

We read each vendor's pricing page, help center and docs on 11 November 2026 and compared them on the criteria above. We did not run paid trials.

1. PepoChat: best for small businesses that want answers, handoff and actions from free

PepoChat is an AI support agent for websites, best for small businesses that want an agent trained on their site, files, Notion and Google Drive that also hands off to a team inbox and takes actions, starting free.

Pricing: Free is $0 with 100 AI replies a month, 5 knowledge sources and 20 MB of text, no card. Starter is $39 a month ($390 a year) for 1,000 replies, unlimited sources and 200 MB; Growth is $149 for 5,000 replies and 1 GB; Scale $499 for 17,000 and 5 GB. Each answer is one reply; extras cost $39 per 1,000.

Strengths: It reads a 200-page crawl, a sitemap, up to 20 links, PDFs including scanned pages, DOCX, Markdown, CSV, JSON, images, Q&A pairs, its own Help Center, and Notion and Google Drive, synced daily. A crawl, sitemap or connected app is one source, and paid plans re-fetch pages weekly. Free includes the team inbox and human handoff; every plan gets 22 prebuilt actions (Shopify orders, Stripe, Zendesk tickets) and Cal.com or PepoSmart booking; paid plans add voice and phone. Install with one script tag or the WordPress plugin.

Limits: An import reads 200 pages, so a bigger site takes several imports with path filters, still one source. Free refreshes website pages only when you click Sync website pages. No YouTube or audio (upload transcripts), no model picker, no source links under answers, and the "Powered by PepoChat" line shows on every plan.

Verdict: The broadest free plan here for answers backed by a person. Start free, then compare plans.

2. CustomGPT.ai: best for very large document libraries with cited answers

CustomGPT.ai is a retrieval platform built around cited answers, best for associations, schools and knowledge-heavy companies with thousands of documents, videos or recordings that need every answer linked to its source.

Pricing: CustomGPT.ai's pricing page lists Standard at $99 a month ($89 yearly) for 1,000 queries and Premium at $499 ($449) for 5,000. There is no free plan; the 7-day trial needs a card and charges when it ends. A standard message uses one query; lead capture or web search add more.

Strengths: Standard holds 5,000 documents per agent and 60 million stored words; Premium holds 20,000 and 300 million. It reads websites, sitemaps, PDFs and "1,400+ document formats", YouTube, audio files and podcasts, and connects Google Drive, Notion, Confluence and the Zendesk and HubSpot knowledge bases. Citations are included on every plan.

Limits: Automatic sync starts on Premium at $499. The documented route to a person is a "Talk to a Human" button linking to your email, phone or booking page, and $99 for 1,000 queries is steep for a small site.

Verdict: The pick when the library is huge, mixed in format and must be cited; our Chatbase vs CustomGPT vs SiteGPT comparison goes deeper on its meter.

3. DocsBot: best for documentation and help desk ticket libraries

DocsBot is an AI chatbot builder for documentation and support content, best for software teams that want cited answers drawn from docs, cloud drives and solved help desk tickets, with a free plan to test on.

Pricing: DocsBot's pricing page lists Free at $0 with 1 bot, 50 source pages and 100 AI credits, Personal at $49 a month ($492 a year) with 5k pages and 5k credits, Standard at $149 with 15k pages, and Business at $499 with 100k pages. A page is "about 5k characters of cleaned text".

Strengths: DocsBot lists 28+ source types, among them websites, sitemaps, RSS, YouTube, documents, Notion, Confluence, Google Drive, SharePoint, Zendesk Guide and solved tickets from Help Scout, Freshdesk, Zoho Desk and Zendesk. Its Display Sources setting shows "the cited source list" after an answer. Refresh is monthly on Personal, weekly from Standard.

Limits: Free takes only basic source types and never refreshes. Notion (pages, not databases) and other cloud sources need Personal; tickets need Standard. Credits multiply with larger models, up to 50x, and the bot "stops responding" at the monthly limit.

Verdict: The best fit when your answers already live in docs and closed tickets and you want a source list under every reply.

4. SiteGPT: best for big websites that want source links

SiteGPT is a website chatbot trained on your pages, files and connected apps, best for content-heavy sites that want large page allowances and a source link under each answer.

Pricing: SiteGPT's pricing page lists Starter at $59 a month or $468 a year, Growth at $129 or $948, and Scale at $429 or $3,108. Starter includes 4,000 GPT-4.1-mini or 400 GPT-4.1 messages, 1,000 pages and 10 files. No free plan; the 7-day trial needs a card.

Strengths: A page is "2,500 cleaned characters", and the allowance grows to 10,000 pages on Growth and 50,000 on Scale. It learns from links, sitemaps, YouTube, CSV, TXT, PDF, DOCX, PPTX and Markdown files, and connected Notion, Google Drive, Dropbox, OneDrive, Box and GitHub. SiteGPT's docs say "Answers can link to the pages they came from."

Limits: Starter has only manual refresh; Growth syncs monthly and Scale weekly. Ten files is tight for a manual library. The docs say it "cannot check an order, a booking, stock levels, or a customer's account".

Verdict: Strong for a large public website on a mid-range budget, if you can live without actions.

5. Chatling: best for a large text allowance on a small budget

Chatling is an AI agent builder with a live chat inbox, best for budget teams that want millions of characters of knowledge, a 1,000-page crawler and optional source links for $25 a month.

Pricing: Free gives 200 AI credits and 500,000 knowledge-base characters. Starter is $25 a month ($250 a year) for 1,750 credits and 10 million characters, Growth $75 for 30 million and Scale $295 for 100 million. Extra characters cost $2 a month per million.

Strengths: Chatling's data source list covers a crawler that "can extract up to 1,000 pages", sitemaps, URL lists, PDF, DOCX and TXT documents, text, FAQs, and Zendesk, Zoho and Notion imports. The Show Sources for AI Response option lists the sources behind an answer. Auto-sync runs weekly or monthly on Starter and daily from Growth.

Limits: Google Drive and YouTube are not on the source list. Credits per reply run from 0.5 to 25 by model, and Free's two base models cost 5 credits each, about 40 replies a month. Free uploads stop at 1 MB, with no auto-sync.

Verdict: The cheapest paid plan here for a big text knowledge base, if you watch the model you pick.

6. Wonderchat: best free plan with citations and Google Drive

Wonderchat is an AI support agent with a help desk and shared inbox, best for small teams that want a free plan, source citations, Google Drive and training on past support tickets.

Pricing: Wonderchat's pricing page lists Free with 50 AI replies a month, 100 webpages per agent and 10K characters of file training. Starter is $39 a month ($32.50 yearly) with 200 webpages and 500K characters, Basic $149 with 1,000 webpages and 3 million characters, and Scale $499 with 2,000 webpages and 10 million.

Strengths: Free already includes a shared inbox and live chat handover to your team. Google Drive is on every plan, SharePoint on Scale, and the FAQ says "You can train Wonderchat on your past tickets and customer conversations." The citation setting "Cite URL sources when answering" adds read-more links or names the PDF and page.

Limits: Page caps are small, so a 600-page site needs Basic. Retraining is manual on Free, monthly on Starter and Basic, weekly only on Scale. The agent stops answering when replies run out unless you top up.

Verdict: A good free start for a small site that wants cited answers and a team inbox.

7. Chatbase: best when model choice and channels matter more than volume

Chatbase is an AI agent platform, best for teams that want one trained agent on the web, WhatsApp, Instagram and Messenger with a choice of AI models, and whose knowledge fits in 40 MB of text.

Pricing: Hobby is $40 a month ($32 yearly) for 700 message credits, Basic $150 ($120) for 4,000 and Business $500 ($400) for 15,000. There is no free plan, only a 7-day trial. Every plan holds 40 MB of extracted text per agent, and extra storage costs $25 per 10 MB a month, up to 80 MB.

Strengths: Chatbase's data sources guide covers website crawls, sitemaps and individual links, PDF, DOC, DOCX and TXT files up to 20 MB each, text, Q&A pairs, Notion pages, closed Zendesk or Salesforce tickets and Zendesk macros. Changes take effect within a minute.

Limits: "Scanned PDFs with no text layer cannot be trained on." Weekly auto-resync starts on Basic. Google Drive and YouTube are not in the guide, and visitor-facing citations are not documented.

Verdict: Choose it for channels and models; for big libraries the 40 MB cap binds first.

8. Botpress: best for builders who want control of retrieval

Botpress is a platform for building AI agents visually or in code, best for teams with a builder who wants to control knowledge bases, storage and re-crawl schedules in custom flows.

Pricing: Free includes 100 conversations, 100 MB of vector storage and 100 MB of files, hard-capped. Plus is $150 a month billed annually for 250 conversations, 1 GB of vector storage and 10 GB of files, with packs of 100 conversations at $65; Team is $750 billed annually for 1,500. Storage add-ons cost $40 a month.

Strengths: Botpress's knowledge base guide covers whole websites or specific pages, documents in PDF, HTML, TXT, DOC, DOCX and Markdown up to 50 MB each, Tables you can query in plain language, rich text, live web search, and Notion and Google Drive through integrations. You can schedule re-crawls, for example daily.

Limits: Storage is counted in megabytes of vectors, so page capacity is hard to predict, and each page is indexed once by default. Citations are stored for the builder in a variable, not shown to visitors unless you build it.

Verdict: Right when you have a developer; overkill for an FAQ bot.

What about Intercom Fin and Tidio Lyro?

If your team already works in Intercom or Tidio, their own agents are the shorter route, so they are not ranked. Fin costs $0.99 per outcome, runs on Intercom and on help desks such as Salesforce, HubSpot, Freshworks, Zoho and Gorgias with a 50-outcome monthly minimum, and trains on your knowledge sources and data connectors. Tidio's Lyro gives every account 50 free conversations once, then starts at $32.50 a month for 50 Lyro conversations.

Which chatbot fits the shape of your data?

Pick by where your answers live today: the right tool for a 60-page website is rarely the right one for 10,000 PDFs or a help desk full of solved tickets.

Mostly a website

Choose a crawler with a sitemap import and automatic re-sync on a plan you can afford: PepoChat re-fetches weekly from $39, Chatling from $25, SiteGPT from $129. Clean the pages first, as our guide to training a chatbot on your website and PDFs explains.

Mostly PDFs and manuals

Check file types, per-file size and scanned pages. PepoChat reads scanned PDFs with a vision model; Chatbase cannot train on a PDF without a text layer. Split long manuals by topic, as our guide to formatting docs for an AI chatbot shows.

A Notion or Google Drive wiki

Connect the wiki rather than exporting it, so edits flow through. PepoChat, CustomGPT.ai, SiteGPT, DocsBot (from Personal) and Botpress read both; Chatling and Chatbase read only Notion, Wonderchat only Drive. Our Notion and Google Drive chatbot guide covers sharing settings.

Help desk articles and tickets

DocsBot (Standard), Chatbase (Business) and Wonderchat train on solved tickets. A public help center is just a website to any crawler, and PepoChat's built-in Help Center feeds the agent directly.

A very large document set

Past tens of thousands of documents, go to CustomGPT.ai (60 million words on Standard, 300 million on Premium) or DocsBot (100k pages on Business), and trial each on your real files.

It must also do things

If visitors ask "where is my order", "book me in" or "can I talk to someone", answers alone will not do. PepoChat adds handoff, 22 prebuilt actions and booking on every plan; SiteGPT says it cannot check orders or bookings.

What does it cost to train a chatbot on 600 pages and 30 PDFs?

For a 600-page website, 30 PDF manuals and about 1,000 questions a month, the cheapest plan that holds everything costs from $25 to about $345 a month.

Assume 5,000 characters of clean text a page (3 million) plus 2 million in the manuals: about 5 million characters, or 5 MB. Each question gets one answer on a 1-credit model where credits apply. These are list-price estimates on monthly billing, not quotes.

Prices from each vendor's own pages, checked 11 November 2026.

The arithmetic. Chatling: 1,000 replies use 1,000 of Starter's 1,750 credits. PepoChat: Free holds 5 MB but not 31 sources (one site plus 30 files), so Starter covers the content and exactly 1,000 replies; the 600 pages take three imports of 200, still one source. DocsBot: 5 million characters is about 1,000 source pages, inside Personal's 5k. Chatbase: 1,000 credits exceed Hobby's 700, so one $40 block of 1,000 credits makes $80. CustomGPT.ai: 630 documents fit Standard, and 1,000 messages use its 1,000 queries. SiteGPT: about 2,000 pages and 30 files exceed Starter's 1,000 pages and 10 files. Wonderchat: 600 webpages exceed Starter's 200. Botpress: if 1,000 questions arrive as 500 conversations, Plus covers 250 and three $65 packs add 300, so $150 plus $195.

What to do next

• Write down where your answers live (site pages, PDFs, Notion, Drive, tickets) and their rough size, then shortlist the two tools whose limits fit.
• Run 30 real customer questions through each shortlisted bot and check every answer against the page it came from; our hallucination guide has the test method.
• Leaving Chatbase because of its 40 MB cap or its credits? Our best Chatbase alternatives ranks nine options by the reason you are leaving.
• Create a free PepoChat workspace, import your site as one knowledge source, connect Notion or Google Drive, and see what it answers before you pay anything.