ChatGPT or Perplexity for working with documents?

We make an extension for ChatGPT, so read this knowing where it comes from. There are document jobs where Perplexity is the better tool and this article says so.

Retrieval first versus chat first

Perplexity's default motion is a search. You ask, it runs queries, pulls back sources, writes an answer from them and attaches a footnote to each claim. Files you attach become another source inside that same pipeline. The product is organised around the question of where an answer came from.

ChatGPT's default motion is a conversation. You hand it something, it holds that in the thread, and you work on it: summarise, restructure, argue with it, rewrite it for a different reader. Browsing exists, but it is an option you invoke rather than the frame everything sits inside.

That difference shows up in small ways that matter. Perplexity is happy to tell you your sources do not cover something, because absence of a source is a normal outcome in a search product. ChatGPT tends to answer anyway, blending your document with everything it already knows, without flagging which half is which.

When the citing is the point

Some work is only useful if the provenance survives.

Desk research you will hand to someone else. Anything time sensitive, where the model's own training is stale by definition. Claims you would be embarrassed to repeat without a link. In those cases the footnote is not decoration, it is the deliverable, and Perplexity produces it by default rather than because you asked nicely.

When you already have the document

The contract is in your email. The transcript is on your disk. The report was sent to you last week. Nothing needs finding, and search adds nothing to a job that is really about reading closely.

This is where ChatGPT is stronger, and the reasons are ordinary rather than dramatic. It keeps a long thread of context over one file, so the fifteenth follow-up question still knows what the first established. It generates well, which matters when the output is a rewrite, a summary for a different audience, or a draft email built from what the document says. And it will write and run code over a spreadsheet, produce charts and check its own arithmetic, which is a category of document work Perplexity is not built for.

Spreadsheets are also exempt from ChatGPT's token cap, limited instead around about 50 MB, so a large dataset goes in where an equally long text document would not.

Where the honest numbers run out

ChatGPT's file limits are published. Text and document files are capped at 2 million tokens each, the hard size ceiling is 512 MB, and the Free plan allows 3 file uploads per day.

Perplexity does not publish an equivalent set in a stable, quotable form, and it has revised what each plan allows more than once. Any specific figure in a comparison article, including this one, is a liability rather than a help. Check your own account on the day you need to know.

The structural point stands without numbers: neither product is designed for you to hand over a book in one piece and get a faithful reading of all of it. Both work best on material that comfortably fits.

A rough division of labour

The interesting cases sit across both. "Is this clause unusual" needs your contract and the outside world. "Does this paper's finding still hold" needs the paper and everything published since. The practical pattern is to gather and cite in Perplexity, then bring the findings into ChatGPT with the document and do the thinking there.

What neither of them fixes

Scanned pages have no text. Retrieval from documents in ChatGPT is text only, except Enterprise, so a scan arrives empty. Run OCR first, whichever tool you use.

Verification stays yours. Both write confidently and both are sometimes wrong. Asking for the quoted sentence behind each claim catches more than reading the summary twice.

A vague question gets a vague answer. The largest quality lever in either product is still how you ask, and it costs nothing.

Choosing

Ask where your work starts. If it starts with a file, use ChatGPT and split anything too long. If it starts with a question and no source, use Perplexity and open the citations that matter. If it regularly starts with both, use each for the half it was built for rather than forcing one to cover the other badly.