How to use AI with your own files and for research

Two of the most useful things a chatbot can do are reading your own files and researching a question for you. Both can save hours, and both can go quietly wrong. Here's how to ask about a PDF or spreadsheet, when a deep research mode is worth the wait, and how to compare sources so you end up with something you can trust.

Before you upload anything

Check what happens to the file first. Many chatbots let you upload documents, but the rules on storing them and using them for training differ by app, plan and setting. Don’t upload medical records, bank statements, contracts or work documents unless you know the app’s rules and your organisation allows it. Our AI privacy guide shows the settings for the main apps.

Asking about a PDF or document

Upload the file and ask a specific question, not just “summarise this”. Good questions to start with:

“List the five most important points, with the page number for each.”
“What does this say about cancellation fees? Quote the exact wording.”
“What questions does this document leave unanswered?”

Asking for page numbers and exact quotes makes the answer quick to check, because you can jump to that spot. For a long document, the chatbot may work from the parts it judges most relevant rather than reading every page closely, so it can miss things. Ask about one section at a time if it matters.

Scans and photos are harder. OpenAI’s help pages warn that ChatGPT may not reliably pull exact values from scanned files or image-based tables. If your PDF is a scan, check every number it gives you against the page.

Your sources first. Some tools, such as Google’s Gemini Notebook (called NotebookLM until July 2026), base their answers on the sources you add and point to the passage they used. That makes it less likely to mix in facts from elsewhere. It is still not proof that every answer is correct, so check the passages it points to. This is a similar idea to RAG.

Asking about a spreadsheet

For spreadsheets, some chatbots do the maths by writing and running small programs rather than guessing at the numbers. ChatGPT, for example, writes and runs Python code for many analysis tasks. That is usually more reliable than the model doing sums in its head, but it can still pick the wrong column or misread what a column means.

Tips that help:

Keep one table per sheet, with clear column names in the first row and no blank rows splitting the data.
Say what each column means if the names are short or coded.
Ask it to show the steps or the formula, then check one or two results yourself.
Ask for the answer as a table you can paste back into your spreadsheet.

A quick chat or a deep research mode?

	A normal chat with web search	A deep research mode
Good for	Quick facts, a few sources, follow-up questions	Questions that need many sources combined into a report
How long	Seconds	Several minutes or more
What you get	A short answer with some links	A long report with citations
Watch out	It may rely on one or two pages	A long report looks thorough even when parts are wrong

ChatGPT, Gemini, Claude and Perplexity each have a research mode (called Deep Research or Research), some only on paid plans or in some countries. Names, limits and plans change often, so check each app. ChatGPT’s version, for example, can search the public web and files you upload and includes citations so you can check the information.

Give a deep research mode a clear brief: what you want to know, why, which countries or years matter, and what kind of sources you trust most. Then read the report as a first draft, not a finished answer.

Comparing sources

Whichever mode you use, the answer is only as good as its sources. Before you rely on it:

Open the key sources. Make sure they exist and say what the AI claims.
Prefer the original. An official site, the study itself or the publisher’s own article beats a blog summarising it.
Look for agreement. If two independent sources say the same thing, you can be more confident. If they disagree, ask the chatbot to set out both views and the evidence for each.
Check the dates. A report can mix old and new figures in the same paragraph.

Our five-step check goes through this in more detail.

How we put this guide together

Claims about specific tools come from the makers’ own pages, linked next to each claim, including OpenAI’s guidance on data analysis and deep research. The other tips are general good practice. Tools are named as examples, not recommendations: features, names and limits change often, so check each tool’s own site. For tool picks, see the best AI tools for research. Got a tip that works for you? Tell us.