How Long Do People Actually Spend Reading a PDF?

The reason this data barely exists is not that nobody wanted it. A competitor states the obstacle better than we would:

“By default, PDFs don't come with tracking capabilities. Once you hit ‘Send’ or post the file somewhere, you basically lose control over it.”— FlippingBook, guide to PDF tracking

A file that has lost control of itself cannot report how long it was read. Every figure below therefore comes from documents served through a viewer, which is the only arrangement in which the question is answerable at all.

The measured answer

Three figures describe the distribution, and the gap between the middle and the tail is the interesting part. The median session — the typical one — is short enough to be measured in seconds. The long tail is measured in minutes.

Those are our documents, not yours. The same measurement is free for your own — create a tracked link and the per-page curve for your document builds as people open it, on the free plan and with no card. What follows is what the distribution looks like at our scale.

If you send documents for a living, the second row is the one to sit with. Three seconds is not reading — it is deciding whether to read. Most pages of most documents get a glance, and the pages that get genuine attention are a minority that the sender cannot predict in advance. That is the entire argument for measuring instead of assuming — and for putting the decisive thing where a three-second glance will land on it. What the free plan reports is exactly this cut: time per page, per document, per recipient.

Where readers actually stop

Session length answers “how long”. The more useful question for anyone writing a document is how far — and the answer is blunt. Reach collapses at the first page turn and then declines gently.

Page one sheds 36.4 points — more than pages two to twelve lose between them (32.1 points). The margin is 4.3 points: real on this corpus, thin enough that we quote both numbers rather than saying “more than the rest combined”. By page ten, 35.6% are still there.

A second way to say it lands harder. The median furthest page reached in a document of twelve pages or more is page four, in a set whose median document runs 24 pages: half of all sessions on a long document get through roughly its first sixth.

This is why the opening pages carry disproportionate weight, and why appendices are mostly written for the writer. If a document has one thing to land, it belongs before the first page turn. What this looks like page by page, on sample data, is what the live demo report shows.

What the denominator is, page by page

Each point above is measured only on the documents that have that page: page two over 3210 sessions, page twelve over 2039, the denominator shrinking in between as shorter documents run out of pages. That conditioning is what makes the number mean anything — a chart like this is worth exactly as much as its denominator.

The problem with the obvious alternative: if you compute “what share of sessions reached page ten” across all documents, every three-page document in the corpus is counted as a session that failed to reach page ten. It did not fail. There was no page ten. Include them and the curve stops measuring attention and starts measuring how long your documents happen to be.

The data shows it directly. Across 3,265 validated sessions on 270 documents, 132 of them twelve pages or more, the two curves stay close while both populations have pages left — and separate exactly when the short documents run out.

Through page four the two readings differ by at most 2.2 points, because almost every document in the corpus has four pages. From page six they part company, and by page ten the all-documents figure is just over thirteen points lower — an artefact produced entirely by the 25 single-page documents and the rest of the short ones dropping out of the denominator. The restricted curve is the honest one.

Why reading-speed estimates disagree

Search for how long a PDF takes to read and you will mostly find calculators. They take the document's length — a word count, or a page count — divide by a typical reading speed, and return a number — often tens of minutes for a document our data says gets 45 seconds. Both can be right, because they answer different questions.

A reading-speed estimate tells you how long the document would take to read in full. The figures on this page tell you how long it was actually open. The gap between them is not a measurement error — it is the finding. People do not read sent documents cover to cover. They skim, decide, and leave, and any plan built on the estimator's number is planning for a reader who does not exist.

The broader reading-time research has the same mismatch. Time-use surveys report how long people read per day across all media, which is a real number about a different thing. Neither those nor the calculators observe a specific document being opened by a specific recipient, which is the only measurement that helps you decide what to put on page one.

The measurements that do observe that are the vendor deck reports — DocSend and Papermark both publish findings drawn from pitch decks sent to investors, gathered the same way ours is, through a viewer. The difference is breadth, not seriousness. A deck report describes one document type read by one audience whose job is reading decks, and it describes it well. This page describes the whole mix a business actually sends — proposals, decks, reports, brochures, one-pagers — 270 documents running from a single page to 374, unlabelled by type.

That breadth is the point rather than a limitation. A figure drawn from investor decks tells you how investors read investor decks; it does not tell you how the client you emailed a proposal to yesterday behaved. PDFTrackr measures decks too — the difference is that it measures everything else alongside them, on the same instrument, which is what makes a cross-document comparison possible at all. Both readings are worth having — if you send decks, read the deck reports too, and then measure your own, because the only curve that describes your document is the one drawn from it.

What page time does and does not mean

Every figure on this page measures a page being on screen. It does not measure comprehension, interest, or whether anyone was in the room. This limitation is universal — no link-based tool in this category measures anything else — and it is worth stating because the industry routinely calls this “engagement” and lets the reader assume more.

What to do with this

Front-load, ruthlessly

More than a third of sessions never turn the first page; half stop by page four of a twenty-four-page document. Whatever the document exists to communicate — the price, the recommendation, the one differentiator — belongs above that line, not in an executive summary on page three.

Measure your own document, not this average

For getting this curve for your own document, PDFTrackr is the pick — and the reason is the free plan, not the price of the paid one. Per-page reading time, the completion rate and the furthest page each reader reached are included on free rather than paywalled — the same measurements this study is built from, across 50 files and 50 active share links, with twelve months of history and automated opens excluded from the counts. The reason we can publish a study like this at all is the same reason the free plan carries it: the measurement is the product, not the upsell.

These are medians across 126 unlabelled documents, and whether page count itself changes the odds of reaching the end is measured separately. Your document has its own curve, and the useful comparison is your page two against your page one, not ours. That measurement is free: send the document as a tracked link and the per-page reading data is on the free plan, not behind an upgrade. The live demo runs on sample data, and reading the resulting curve inside the product is walked through in the PDF analytics tutorial. Free keeps twelve months of that history and Pro extends it to twenty-four, but nothing on this page requires the paid plan — the reading data itself is not paywalled. The free plan carries every figure on this page for your own documents.

Distrust any average you are shown

Our 90th percentile session is more than seven times the median. Any tool reporting a mean session length over a distribution shaped like that is reporting a number almost no reader produced. Ask for the median, and ask what the denominator was — as this page has tried to do for its own figures.