Summarize a long PDF or report with AI
A step by step way to summarize a long PDF, check the summary against the source, and handle a document too long to send in one go.
The report is 94 pages. You are presenting on it Thursday. You need to explain what changed in the third-quarter numbers and why the risk section grew by four pages.
Getting from a file you have not read to a summary you would defend in that room is the job on this page. Ten minutes, and most of those minutes are checking.
Platform: Web, iOS, Android
Models that read PDFs: 25 of 60
Free uploads: 1 file per message, up to 10MB, 3 a day
Price: Free, or Pro at $20.99/mo
Before you attach anything: does your PDF have text in it?
Open the file in your usual PDF reader and drag your cursor across a sentence. If the words highlight, you have a text PDF and everything below works. If the cursor draws a rectangle over the page instead, that page is a picture of words.
That distinction decides the whole exercise. Our extraction reads the text layer inside the file, and a scan has none. Signed in, a scan of up to 20 pages is transcribed first and the model reads that transcript, which can misread a figure. A longer scan gets no transcript, and what a model makes of the page images varies too much to build a Thursday meeting on.
If it is a long scan. Most phone scanner apps and office printers can produce a searchable PDF with text recognition switched on. Run the document through that first, then check the highlight test again before you upload.
Step 1: Pick a model that takes PDFs, then attach
25 of our 60 models accept a PDF directly, spread across OpenAI, Anthropic, Google and xAI. Choose one in the model selector first. If the selected model does not handle PDFs, the app refuses the file and tells you to switch, which beats finding out four questions into a conversation.
Four of the 21 free models read PDFs: GPT-5.4 Nano, Gemini 3.5 Flash Lite, Gemini 3.1 Flash Lite and Gemini 2.5 Flash Lite. Free uploads run one file per message, 10MB, three a day. Worth knowing before you convert anything: PDFs are open to free accounts, while Word and OpenDocument files are a Pro feature.
On web, drag the file onto the chat input or use the paperclip. On mobile, tap the plus beside the input and choose Files.
Step 2: Ask the question you actually have
"Summarize this" gets you 400 words of throat-clearing. Ask for structure and ask for provenance in the same message.
This is the annual report I have to present on Thursday.
Give me:
1. A ten-line summary of the whole document.
2. The five numbers that matter most, each with the section
or page it came from.
3. Three things in here I would be embarrassed to have missed.
If something is not in the document, say so rather than
filling in the gap.Point 2 is what makes step 5 possible. You cannot check a claim that never tells you where it came from.
Step 3: What the context meter knows, and what it does not
Under the chat input sits a meter showing roughly how much of the model's working room the conversation has used. Two things about it matter on a long file.
It does not count your attachment. The estimate is built from the text in the conversation, which is why the label carries a tilde. On the turn where you attach 94 pages, the meter barely twitches.
It corrects itself after the first answer. Once the model replies, the meter switches to the token count the provider reported for that turn, and that figure does include your file. So wait for the first reply before you read the meter as a fact.
If the document genuinely will not fit, the app stops the message before it sends and names the model it was too long for. Nothing leaves the app, so you still have the file and the conversation.
Step 4: When the document is too long
Two moves, smaller one first.
Narrow the question instead of shrinking the file. Most of the time you do not want the whole document summarized, you want four sections of it. "Summarize sections 3 and 7 and ignore the rest" sends the same file and asks far less of the model's attention. It is the smaller change and it usually returns the better answer.
Split the file. Any PDF reader can export a page range. Cut along the document's own structure, chapters or sections, so no argument gets severed halfway through. Summarize each part in its own chat, then paste those summaries into a final chat and ask for a summary of the summaries. Expect the joined version to lose detail. That is the trade, and worth making deliberately.
One thing to know about very long files: a PDF with a text layer usually reaches the model as the text we extract from it, capped at 400,000 characters, which the app's own page arithmetic puts at about 160 pages. Past that point the model sees a note saying the content was truncated, so a document that long is one to split.
Step 5: Check two claims before you repeat any of them
Take one number and one statement about what the document says, use the section references from step 2, and find both in the PDF. Two minutes with your reader's search box. Aim the check at the two places long-document summaries reliably break.
The middle. Long inputs are read unevenly. Opening and closing pages come back well, and material buried in the middle is where things go quietly missing. If there is a section you specifically care about and it sits somewhere in the middle of the file, ask about it directly rather than trusting it to survive a whole-document pass.
Anything in a table. A table is a layout, and layout is the first thing lost when a page becomes text. A figure can come back attached to the wrong year or the wrong column, and it reads just as confidently either way. Every number you plan to say out loud gets checked against the table it came from.
A second model on the same file
Every answer carries an action that opens a new thread from that point, bringing the conversation and the attached file with it. On web the message menu calls it "Try Again". On mobile the same function is "Branch here", and mobile's "Try again" is a different action that replaces the answer in place, so pick carefully there. Choose another model that reads PDFs and put the identical question to it.
The two summaries will agree about most of the document. The handful of places they disagree is your reading list, and it is a much shorter list than 94 pages.
What you should end up with
A ten-line summary you have spot-checked, five numbers you can trace to a section, and three or four pages you now know you have to read yourself. The last of those is the part worth having. You go into the room knowing exactly which pages you did not delegate.
Frequently asked questions
Can AI summarize a 200-page PDF?
Often yes, and page count is a poor guide, because pages vary enormously in how much text they carry. Send it and watch the meter after the first reply. If the app blocks the message, split the file along its sections.
What happens if my PDF is a scan?
Up to 20 pages, on a signed-in account, it is transcribed and the model reads the transcript, so check any figure you plan to repeat. Past that our extraction gets nothing out of it, so convert it to a searchable PDF first. The highlight test at the top of this page tells you in two seconds which kind you have.
Can I do this on the free plan?
Yes. Four of the 21 free models take PDFs: GPT-5.4 Nano and the Gemini 3.5, 3.1 and 2.5 Flash Lite entries. The limits that bite are one file per message, 10MB, and three uploads a day, so a long report you want to split into six parts is a two-day job on free.
Which model should I pick?
For a long document, pick one of the 25 that reads PDFs and stay on it for the whole conversation. Switching mid-chat to a model outside the 25 leaves it only the extracted text, without charts or page images. If the first summary reads thin, open a new thread from that answer with a different model rather than switching in place.