TL;DR (June 2026): “Summarize this PDF” is three different jobs — report, paper, contract — each with its own prompt and its own trust model. Tool limits decide what is even possible: Claude.ai accepts 100 pages / 32MB per file in the web UI; Gemini accepts 1,000 pages / 50MB and reads each page as an image (best for scanned docs); ChatGPT caps any one file at 2M tokens but does text-only retrieval on Free/Plus/Pro. NotebookLM is for cross-reading many sources (50 free, 300 on Plus/Pro). The one rule that saves you on high-stakes docs: treat every page citation as a pointer to verify, never as a finding to trust.
The task
A 187-page McKinsey report just landed in your inbox at 4pm with a “thoughts before tomorrow’s strategy review?” note from your boss. Or it is a 38-page academic paper a colleague says you “really should read.” Or it is the 24-page SaaS contract that needs your sign-off by Friday and the only person on legal is on PTO. “Summarize this PDF” sounds like one task. It is actually three completely different jobs — different prompt, different tool, different trust model.
Where AI helps — and where it does not
AI is excellent at compression: turning 50 pages into 800 words that preserve the load-bearing arguments. It is also good at structural reading — picking out the methodology section of a paper, the indemnity clause of a contract, the data appendix of a report. Where it fails: invented quotes. When you ask for “exact page citations,” a stretched model will fabricate page numbers that sound plausible. Worse, it will paraphrase a claim that does not actually appear in the doc and attribute it to “page 27.” Treat the citation as a pointer to verify, not a finding to trust. The fix is structural: every summary should be auditable against the PDF, and your workflow should make the audit step cheap.
A common failure mode: the model summarizes the abstract (or the executive summary) and you mistake that for a summary of the full doc. Long PDFs need a chunked or whole-doc read; if the model finishes in 8 seconds for a 100-page report, it probably read the first 5 pages and the table of contents.
What to feed the AI
- The PDF itself, not a screenshot — upload the file so OCR is clean
- Page count and document type — knowing it is a 187-page report vs a 12-page memo changes the prompt
- Your real reading goal — decide-to-cite, decide-to-buy, decide-to-replicate, decide-to-sign
- The 2-3 questions you most want answered (the model summarizes around your questions, not the doc’s table of contents)
- Your domain context — “I am a Series B SaaS founder reading this from a procurement angle” produces very different output than “I am a corporate counsel reviewing risk”
- A trust constraint — “Every claim must include a page or section number; if you cannot cite, say ‘no citation’ rather than guessing”
- The output format you actually need — bullet brief, slide notes, a 200-word stand-up summary, or a 2-page memo
- The audience for your downstream output — your boss, your team, your future self
Three workflows, three prompts
1. Long reports (business / policy / industry research)
Read this {N-page} report. My goal: {decision the summary supports}.
My background: {role + what I already know about this topic}.
Return:
1) One-sentence core conclusion in the report's own terms.
2) Each major section as one bullet (≤3 sentences). Mark sections that change my decision vs sections that are background.
3) The 3 most consequential data points, each with page number. Mark any that the report does not show source data for.
4) 2 claims I should push back on — quote 8-15 words and give the page number.
5) 3 questions a skeptic would ask the authors. The questions you can answer from the doc, answer in one sentence; the rest mark "needs follow-up."
6) The single sentence I should walk into tomorrow's meeting with.
Trust constraint: if you cannot cite a page for a claim, write "uncited" rather than guessing.
Attachment: {upload}2. Academic papers
You are a senior peer reviewer in {field}. I am deciding whether to {cite / replicate / build on} this paper.
Return:
1) The research question (1 sentence).
2) 3 distinctive method features (data, model, study design) and why each matters.
3) Main result with the effect size and how much I should trust it (sample, replication, confound risk).
4) 2 likely weaknesses in the experimental design.
5) Relation to the 3 most relevant prior works in the field.
6) 2 concrete follow-up experiments worth pursuing if I work in this area.
7) The headline finding I would put in a 1-sentence Slack message to a colleague.
Trust constraint: do not invent prior-work citations. If a claim has no support in the paper, say so.
Attachment: {upload}3. Contracts / legal documents
You are a paralegal familiar with {US / EU / Chinese} commercial contracts. I am the {buyer / vendor / partner} signing this.
Return:
1) Each party's obligations as bullets, with section references.
2) The 5 clauses I should be most cautious of (indemnity, exit, IP assignment, auto-renewal, jurisdiction, liability cap). For each: quote the exact 10-25 word clause text and cite the section number.
3) 3 clauses I should negotiate to soften — name the specific edit (e.g., "cap indemnity at 12 months of fees, not unlimited").
4) 3 sentences I can send to my counterparty proposing those changes.
5) The single risk a lawyer would flag in a 30-second skim.
Important: this is not legal advice. The lawyer reviews the final version.
Attachment: {upload}Shorter variant — single-question lookup
Read this PDF. Answer only this question: {one specific question}.
Give the answer + the page number + 12-25 quoted words of supporting text.
If the PDF does not actually answer this, say "not found" — do not infer.
Attachment: {upload}Tool fit (limits as of June 2026)
Pick the tool by the document’s hard limits first, then by quality. The page/size caps below are the consumer web/app limits, which are tighter than the raw model context — they are what actually stops your upload.
| PDF type | Tool of choice | Hard limit (web/app) | Why |
|---|---|---|---|
| Long report (under 100 pages) | Claude (Sonnet 4.6 / Opus 4.7) | 100 pages, 32MB/file | 1M-token context holds the whole doc; strong at preserving load-bearing arguments |
| 100–1,000 page report | Gemini 3.1 Pro | 1,000 pages, 50MB/file | Reads each page natively; highest single-file page ceiling |
| Academic paper with formulas | ChatGPT (GPT-5.5) | 2M tokens, 512MB/file | Python tool re-derives equations and replicates small experiments |
| Contract / legal | Claude (Opus 4.7) | 100 pages, 32MB/file | Best long-quote retention; least likely to silently paraphrase clauses |
| Chinese / Japanese / Korean PDF | Kimi / Tongyi / Gemini 3.1 Pro | varies by tool | Stronger CJK OCR and domain terminology |
| Scanned / image-only PDF | Gemini 3.1 Pro | 1,000 pages, 50MB/file | Native vision reads pages as images; no separate OCR step |
| Cross-read many sources | NotebookLM | 50 sources (Free), 300 (Plus/Pro) | Indexes a whole project for grounded cross-citation |
Two limits people trip over: ChatGPT does text-only retrieval on Free/Plus/Pro (visual retrieval is Enterprise-only), so charts and scanned pages get lost — send those to Gemini. And Claude.ai’s 100-page web cap is far below its 1M-token model context, so a 250-page report needs Gemini, the Claude Files API, or a split-and-combine pass.
Sample output
A useful report summary headline: “The McKinsey deck argues SMB SaaS will consolidate into 4 vertical winners by 2028, but the supporting data ends at 2024 — the trend extrapolation is the load-bearing claim and is uncited.” That single sentence does more than 800 words of dot-pointing.
A useful contract risk flag: “Section 9.4 (Indemnity, p.11): ‘Customer shall indemnify Vendor against any claims arising from Customer’s use of the Service, without limitation.’ ‘Without limitation’ is the part that makes this uncapped — propose ‘capped at 12 months of fees paid’ as an edit.”
A useful paper finding: “Main claim: prompt chaining improves accuracy by 14pp on the benchmark. Trust: medium. The benchmark is the authors’ own (n=200), no third-party replication. Confound: chain length and total token budget are co-varied, so you cannot tell if it is the structure or just more compute.”
How to refine
- If the summary feels generic: “Rewrite focused only on what changes my decision to
[sign/cite/buy]. Drop anything that is just background restatement.” - If citations look suspicious: “Quote the exact 12-25 words behind each claim. If you cannot locate a quote in the doc, mark ‘uncited’ and do not paraphrase.”
- If the model misses page 100+: “Confirm: did you read all
[N]pages? Name 3 specific findings from the last third of the document. If you did not read past page[X], say so.” - If the legal tone is over-hedged: “Cut every ‘may,’ ‘might,’ ‘could potentially.’ If a clause is risky, say so. If a clause is fine, say so. Hedging on every clause is not helpful.”
- If the academic critique feels shallow: “Apply a hostile peer reviewer’s standard. Name the experiment you would run to break this paper’s main claim.”
Common mistakes
- Trusting a fast read: a 100-page summary returned in 6 seconds is the abstract, not the doc; force a verification question on a late-section finding.
- Letting AI paraphrase legal text: paraphrased clauses miss the loaded words (“without limitation,” “sole discretion,” “perpetual”); always quote contract text verbatim.
- One prompt for all PDFs: the same prompt makes a contract review read like a report exec summary, and you miss the actual risk.
- No domain context in the prompt: same paper read by “a senior reviewer in NLP” vs “a curious generalist” gives very different output; specify which.
- Skipping the OCR step on scanned PDFs: image-only PDFs without OCR produce hallucinated summaries because the model is reading the file name and metadata.
- No page-by-page audit on high-stakes docs: for contracts and policy work, treat the summary as a guide to where to read, not a substitute for reading.
- Forgetting the doc’s age: a 2019 industry report has 2019 assumptions; the summary should flag the data freshness, not present it as current.
- Multi-doc cross-reading without source tagging: if you summarize 5 PDFs at once and the output does not tag which claim came from which source, you cannot audit.
FAQ
- What is the real per-file page/size limit (June 2026)?: Claude.ai web UI rejects any PDF over 100 pages or 32MB; the Claude Files API goes to 500MB. Gemini accepts up to 1,000 pages / 50MB and reads each page as an image. ChatGPT caps a single file at 2M tokens / 512MB but does text-only extraction. For anything past those caps, split the PDF and run a final combine pass, or send it to NotebookLM as one of up to 50 (Free) / 300 (Plus/Pro) sources.
- Should I always convert PDFs to markdown first?: Helpful for messy PDFs (scanned, multi-column, lots of tables). Skip for clean modern PDFs — current models handle them natively. The exception: financial filings with deeply nested tables; markdown almost always helps there.
- How do I check the model did not hallucinate the summary?: Pick one specific claim with a page citation, open the PDF to that page, and verify the quoted 12-25 words exist. If they don’t, ask the model to redo with stricter quote retention. Two mismatched citations and you switch tools.
- Why did ChatGPT miss the charts in my report?: On Free, Plus, and Pro, ChatGPT does text-only retrieval — it extracts the text layer and ignores chart images and scanned pages (visual retrieval is Enterprise-only). For chart-heavy or scanned PDFs, use Gemini 3.1 Pro, which reads every page as an image.
- Is the contract clause review safe to send to my counterparty?: The clause analysis is research; the email draft to your counterparty is a starting point. A lawyer must review the actual ask before you send it. This article is not legal advice.
- Can the model spot what is missing from a doc? — Only if you ask. Try: “What standard sections are missing from this contract compared to a typical
[SaaS / NDA / employment]agreement?” The model is more useful as a checklist than as a free-form critic.