Do People Finish Long PDFs? What Reading Data Says About Length

By Oleh Tsyupa, Founder of PDFTrackr · Published 2026-09-03 · Updated 2026-09-03

8 min read

Past twenty pages, the median document is finished in 14.6% of sessions. Between two and nineteen pages it runs between 32% and 44.4%.

Those are document-level medians: each document with at least five sessions gets one vote, which is the basis to read first because a handful of busy documents otherwise carry the corpus. On the session-weighted basis the same bands read 31.3%, 54.8%, 28.7% and 11% — and the 5–9 pages figure there is one customer, which is the subject of its own section below. A session counts as finished when the furthest page it rendered is at least the document's page count, so twelve pages into a twelve-page paper is finishing it and twelve pages into a sixty-page one is not.

Based on 3,745 recorded sessions across 247 documents and 277 share links, 22 Sep 2025 – 25 Aug 2026, extracted 26 Aug 2026 — multi-page documents only, because a one-page document is finished by being opened. That restriction drops 153 sessions from the 3,898 recorded: 83 on one-page documents, and 70 more whose document has no page count on record. A session counts as finished when the furthest page it rendered is at least the document's page count. Every recorded session inside the restriction is counted, automated ones included, so the group that renders nothing at all stays in the denominator.

If you want this measured on your own document rather than on ours, share it once as a tracked link — that is enough to produce the same numbers for your reader, and it is free. Or open the live demo to see what the per-page record looks like first.

The short version, and what it is not

This page answers one question: does the number of pages change whether a reader reaches the end? It deliberately does not re-argue what finishing means or how often it happens in general — that is our page on whether anyone finishes your whitepaper, which owns the definition and works the same dataset from the other direction. Every figure on this page is the same extract as that one, so the two cannot drift apart.

It is also not about how long a reading session lasts. How long people actually spend reading a PDF carries the durations and the fixed-page drop-off curve. Time and completion are different axes: a reader can spend eight minutes on the first four pages of a fifty-page report and finish nothing.

Finishing, band by band

Read the median-document column first. It gives each document with at least five sessions one vote, so a single heavily-read file cannot decide the answer. The session-weighted column counts every recorded session and is published beside it because the disclosure rule requires it, not because it is the better answer.

Completion by document length, on both bases, from one extract. The median-document column is the one to read. 3,745 recorded sessions across 247 documents and 277 share links, 22 Sep 2025 – 25 Aug 2026, extracted 26 Aug 2026 — multi-page documents only, because a one-page document is finished by being opened. That restriction drops 153 sessions from the 3,898 recorded: 83 on one-page documents, and 70 more whose document has no page count on record. A session counts as finished when the furthest page it rendered is at least the document's page count. Every recorded session inside the restriction is counted, automated ones included, so the group that renders nothing at all stays in the denominator.
Document lengthMedian document finishesSession-weightedDocuments with 5+ sessionsWhat it is safe to say
2–4 pages37.5%31.3%17Short does not mean finished — on the median document, roughly three sessions in five still stop early
5–9 pages44.4%54.8%21The strongest band on both bases, but the session-weighted figure is one sender
10–19 pages32%28.7%22Statistically hard to separate from the shortest band
20 pages or more14.6%11%40The cliff — finishing roughly halves against every shorter band
Automated opensKept in the denominator here, deliberatelyKept in the denominator here, deliberatelyPDFTrackr classifies automated opens and excludes them from a sender's view counts; this dataset keeps them so the group that renders nothing stays visible

The pattern to take from that table is the shape rather than any single number. Three of the four bands sit inside a range you could not tell apart without a much larger corpus. The fourth is not close to them. If length mattered smoothly, the middle two bands would sit between the first and the last. The ten-to-nineteen band does — 32% against the two-to-four band's 37.5% on the document basis, and 28.7% against 31.3% session-weighted. The five-to-nine band does not: it finishes above the two-to-four band on both bases, a rise before the fall rather than a gradient, and the next section shows how much of that rise is one sender.

The band that is one customer

The 5–9 pages band is the one an incautious reading turns into a headline, so here is what is inside it. Session-weighted it finishes at 54.8%, well clear of every other band — and 508 of its 865 sessions, 58.7% of the band, come from a single sender with 4 documents.

Drop that one sender and re-derive. The band falls to 37.8% on 357 sessions — still the highest band, but its lead over the two-to-four band narrows from 23.5pp to 6.5pp. On the document-median basis the same removal moves it from 44.4% to 41.4%, which is barely a move at all. That difference between the two bases is the entire reason both are published.

So the sentence “five to nine pages is the length people finish” is not supported by this data. What the leave-one-out check moves is the size of the gap rather than its direction: without that account the band is still the highest on both bases, but a 23.5pp session-weighted lead becomes 6.5pp, which is not a page-count rule. We are writing that down rather than quietly leading with the bigger number, because the check is cheap and the alternative is publishing one customer's behaviour as a finding about readers in general. The cliff past twenty pages survives the same check — it is spread across 40 documents, more than any other band.

Why the cliff sits where it does — carefully

We can measure where completion falls. We cannot measure why, and this section is explicitly the speculative part of the page.

The mechanical reading is that finishing is defined against the document's own length, so a long document simply asks for more. A reader who gives any document the same eight minutes finishes a six-page brief and does not finish a forty-page report, without their behaviour changing at all. On that reading the cliff is arithmetic rather than psychology, and the interesting question is not “did they finish” but “did they reach the part that mattered”.

The compositional reading is that documents of different lengths are different kinds of document. Two-to-four pages is a one-pager, a quote, a covering note. Twenty-plus is a report, a prospectus, an annual review — often circulated for reference rather than for reading end to end, and often skimmed to one section on purpose. A low completion rate on a reference document is not a failure; it may be the document working correctly.

Both readings are consistent with the data and this page does not choose between them. What follows practically is the same either way: on a long document, completion is the wrong headline metric, and which pages were reached is the right one.

What to do with a long document

Three things follow that do not require believing either explanation.

Front-load, and stop measuring the end. If the median document past twenty pages is finished in 14.6% of sessions, then anything you need the reader to see cannot live on page twenty-eight. Put the ask, the number and the recommendation early, and treat the rest as the evidence a minority will check.

Split, rather than shorten. The data does not say a shorter document is finished more often — inside two to nineteen pages it barely moves. It says a document over twenty pages is finished much less often. Splitting a thirty-page report into a six-page summary and a twenty-four-page appendix is a different move from cutting it to twenty-two pages, and only the first one crosses the cliff.

Change the question you ask of the data. Per-page reading tells you which section held attention and where people stopped, which is actionable on a long document in a way a single completion percentage never is. That is what per-page reading data on your own document is for.

Measuring your own

Our corpus is documents whose owners chose to share them as tracked links, which is already a selection, and it is weighted towards business documents rather than long-form reading generally. A figure measured on somebody else's corpus would land somewhere else on the same axis. The useful move is to measure your own, and it takes one link.

Share the document as a tracked link instead of attaching it, send it as you normally would, and the record comes back page by page — the mechanics, and what the free plan includes, are set out on the pillar page. What comes back is: what rendered, in what order, for how long, and where the reader stopped. Automated opens are classified when the session closes and excluded from the counts, which matters on a long document because a mail-security scanner that fetches page one and leaves would otherwise sit in your data as somebody who bounced. How large that effect is on our own corpus is measured on our page on whether PDF view counts are real.

When Pro becomes the right choice

For this question, free is the whole answer and it is not a teaser: page-by-page reading, the drop-off page, the automated-open filter and twelve months of history are all on the plan that costs nothing. One long document shared as one link produces every number on this page for your own reader.

Pro earns its nine dollars later and elsewhere. Once you are running enough documents that fifty files or fifty active links stops being enough, once a long report needs replacing behind a link people already have without breaking the URL or losing its reading history, or once you want twenty-four months of history rather than twelve because the review cycle is annual. None of that changes what you learn about length; it changes how many documents you can ask it of.

Frequently asked questions

Do people finish long PDFs?

Usually not, and the fall is sharper than a smooth relationship with length would predict. On our own hosted corpus, giving each document with at least five sessions one vote, documents of twenty pages or more are finished in 14.6% of sessions, against 32.0% for ten to nineteen pages and 37.5% for two to four. Finishing is measured against each document's own page count, all time to 25 Aug 2026, extracted 26 Aug 2026.

Does making a document shorter make people finish it?

Not reliably, on this data. Between two and nineteen pages the median document's completion moves around between roughly a third and a half with no clean relationship to page count, so trimming a fifteen-page document to nine is unlikely to change much. The visible effect is at the twenty-page boundary, which is why splitting a long document into a short summary and a long appendix is a different intervention from trimming it.

What is the ideal length for a PDF?

This data does not support one. Five to nine pages completes best on both bases, but the size of that lead is one customer: 508 of the band's 865 sessions come from a single sender with four documents, and removing them takes the band from 54.8% to 37.8% session-weighted — still the highest band, with its lead over the shortest one down from 23.5pp to 6.5pp. A 6.5pp gap on 357 sessions is not a page-count rule, and publishing it as an ideal length would be one customer's behaviour dressed up as a finding about readers.

Can you measure completion on a PDF you emailed as an attachment?

No, and neither can anybody else. An attached file is a copy on someone else's disk; it does not report which pages were opened, and no document tracker changes that, ours included, though PDFTrackr does classify and exclude automated opens from the sessions it can see. Completion can only be measured on a hosted copy — a link the reader opens in a viewer — which is both the limit of this data and the reason it exists at all.

Do bots and scanners inflate these numbers?

They are inside this particular dataset on purpose, and they push the completion figures down rather than up: every recorded session is counted, automated ones included, so that the group which renders no page at all stays in the denominator. That is a methodological choice for this page. In the product itself, automated opens — corporate mail-security scanners, link previewers — are classified at session close and excluded from a sender's view counts, and the dashboard shows what the filter removed.

Sources

  1. Nielsen Norman Group — How Little Do Users Read? (reading behaviour on screen, not documents; the closest published prior on the general question)
  2. Slate — How people read online: why you won't finish this article (scroll-depth measurement on articles rather than PDFs)
  3. freshspectrum — How long does it take to read a report? (reading-rate estimates, the arithmetic behind time against page count)
  4. Microsoft Learn — Safe Links in Microsoft Defender for Office 365: URLs scanned before message delivery, the mechanism behind a session that renders nothing (accessed 2026-08-12)

Find out where your own readers stop

Share one long document as a tracked link and read the drop-off page by page. Free: 50 files, 50 links, 12 months of history, no card.

Create a free tracked link

Keep reading: how long people actually read a PDF, does anyone finish your whitepaper, and how free PDF tracking works.

Oleh Tsyupa

Founder, PDFTrackr

Has analysed over 3,000 tracked document-viewing sessions on PDFTrackr.