How to Train a Custom GPT on Your Company Docs
CustomGPT is a no-code platform for building a retrieval-augmented chatbot from your own documents — PDFs, help center pages, Notion exports, and sitemaps. A typical company gets a working internal assistant live in two to four days, and most of that time is cleaning source documents, not clicking through settings.
The distinction that matters: CustomGPT does retrieval-augmented generation, not fine-tuning. Your documents are chunked, embedded, and searched at question time; the underlying model is not retrained on them. That is the right architecture for company knowledge, because it stays current and you can see which document an answer came from.
What sources can CustomGPT ingest?
Roughly everything you would expect, with different quality outcomes.
| Source | Setup effort | Answer quality |
|---|---|---|
| Help center sitemap | Low | High — already structured |
| PDF manuals | Medium | Medium — layout breaks chunking |
| Notion or Confluence | Medium | High |
| Google Drive folders | Medium | Varies wildly |
| Raw CSV | High | Poor — needs narrative |
The sitemap crawler is the strongest path for most SaaS companies. Point it at your docs domain and it handles the rest, including re-crawling on a schedule.
How do you set up a CustomGPT step by step?
Follow this sequence and you avoid the two classic rebuilds.
- Write the 100-word instruction block first. Who the bot serves, what it may answer, and what it must refuse. This block controls more output quality than any model setting.
- Add one source type only. Start with your help center sitemap. Get it answering well on that corpus before adding PDFs.
- Set chunk size deliberately. Default chunking splits by size, not by topic. For documentation with clear headings, enable heading-based splitting. For dense PDFs, keep chunk size smaller (around 500-800 tokens) and accept fewer retrieved passages.
- Add a refusal rule. “If the retrieved context does not contain the answer, say you don’t know and offer a support link.” Without this, the bot will confabulate from general knowledge.
- Test with 20 real questions. Pull them from your support inbox, not from your imagination. Log every wrong answer with the source document you expected.
- Fix the documents, not the bot. About 70% of wrong answers trace back to a source document that is outdated, contradictory, or badly formatted.
How much does CustomGPT cost?
CustomGPT prices per chatbot tier with a message allowance, not per seat in the simplest plans. Entry plans sit in the $99-$150 a month range, and higher tiers with more sources and messages run into the several-hundreds per month. Agencies that resell client chatbots need the higher tier because white-labeling is gated.
There is a free trial that lets you test with real documents, which is unusual in this category and worth using before you commit.
How do you keep answers accurate over time?
A RAG bot degrades quietly. Nobody complains on day one; six months later the bot quotes a price that changed twice.
Set a monthly review:
- Re-crawl sources on a schedule — weekly for help centers, monthly for manuals.
- Read the unanswered questions log. CustomGPT surfaces questions it could not answer. That list is your content backlog.
- Delete superseded documents. Two versions of the same policy produce contradictory answers and the retrieval picks randomly.
What does CustomGPT not do well?
Complex multi-step reasoning. Asking “which plan covers my situation given these five constraints” gets a synthesised answer that sounds confident and is often half right. Keep the bot for factual lookup.
Structured data queries. “How many tickets did we close in March” is a database question. A document bot cannot answer it and should not pretend to.
Permission-scoped answers. Source-level access control exists on higher tiers, but it is coarse. If your documents contain salary bands, do not put them in a bot that serves the whole company.
Who should build one?
Support leads with 200+ help articles and a support queue: yes, clearly. The bot deflects repetitive questions and answers consistently.
Anyone hoping to avoid writing documentation: no. A RAG bot amplifies the quality of your docs. If they are bad, it makes bad answers faster.
Key Takeaways
- CustomGPT builds a RAG chatbot from your documents without code; setup takes two to four days.
- Start with one source type — a help center sitemap — before adding PDFs or Drive folders.
- Write a refusal rule into the instructions or the bot will invent answers.
- Entry plans run about $99-$150 a month; a free trial with real documents is available.
- About 70% of wrong answers are document problems, not model problems — fix the source, not the prompt.
Full feature notes are on the CustomGPT page at iAIFeed.
