The PDF problem

A PDF knows where every mark of ink goes on a sheet of A4 and, strictly, nothing else. Which glyph follows which, where a sentence ends, what is a heading and what is a page number — none of that is in the file as fact. It is in the file as geometry. Every tool that reads PDFs aloud, reflows them, or streams them one word at a time is not reading the document. It is reconstructing one.

This is not a design failure; it is the design. PDF was built to do one thing perfectly — make a document look identical on every printer and screen on earth — and it does. The cost of perfect fidelity is that the format stores the result of typesetting rather than the text that went in: positioned glyph runs, coordinates, fonts. The paragraph, as a thing, was compiled away. Three decades later the format meant for print has become the default container for papers, reports, contracts, statements and syllabi, and every reader who is not a laser printer inherits the bill.

The bill itemises like this. Fixed layout means nothing reflows: on a phone, an A4 page becomes a pan-and-zoom exercise. Two-column layouts store their text in whatever order the producing software happened to emit, which may be column by column, may be interleaved, and may be neither. Headers, footers and page numbers sit in the text stream like anyone else. Hyphenated line-breaks split words in half as literal fact. And a scanned PDF — a photograph of paper wearing a .pdf extension — contains no text at all.

Every class of reading tool hits this wall in its own way. E-readers render PDF pages as tiny pictures of paper, and their reflow modes mangle exactly the documents that most need them. Phone reader modes usually refuse PDFs outright. Text-to-speech tools read the guessed order aloud, which turns a mis-ordered column into confident nonsense — worse than an error you can see. And single-word streams like ours depend on the order utterly, because the reader never sees the page and cannot notice something has gone wrong. The struggling is universal because the missing information is universal.

PDF accessibility, on paper and in practice

The format’s own answer exists and has for a long time: tagged PDF. Tags embed the missing structure — this is a heading, this is a paragraph, read the columns in this order, this image means the following — and a properly tagged file works well with screen readers and reflows respectably. There is a standard for it, PDF/UA, and public-sector accessibility rules in several jurisdictions nominally require it. PDF accessibility, as a specification, is largely a solved problem.

In practice, the tags are almost never there. Export paths that could produce them are optional and off, most authors do not know the feature exists, and nothing visibly breaks when it is skipped — the file looks identical, because looking identical is the one thing PDF always delivers. So assistive software meets an untagged wall and guesses, the same way every reflow engine guesses. The accessibility failure of the world’s PDFs is not a hard research problem. It is an unfunded chore, distributed across everyone who ever clicked Export.

The consequence lands on predictable people. A student with a screen reader gets the scanned course pack; the employee with low vision gets the fixed nine-point two-column report; anybody on a phone gets the A4 page designed for nobody’s pocket. Formats are policy, whether or not anyone decided them, and PDF’s defaults quietly decide who reads with ease and who reads with effort — the thread running through the whole of assistive reading technology.

The file looks identical, because looking identical is the one thing PDF always delivers.

The export button, stated plainly

What on-device parsing can recover

Our app takes PDFs, so this is our fight too. The parsing runs on your machine with pdf.js, the open-source engine Firefox uses, and nothing is uploaded — the point of the whole architecture, argued in offline by design. What pdf.js gives us honestly: every text run on the page, with its position, size and font. From that we rebuild lines out of fragments, rejoin hyphenated words, drop repeating headers and footers, and infer column order from geometry. On a clean, born-digital, single-column document this works well — well enough that you would not guess the paragraph had ever been compiled away.

Now the failure modes, ours as much as anyone’s. Column inference is a heuristic, and unusual layouts — sidebars, pull quotes, marginal notes, forms — can shuffle the order in ways we cannot always detect, let alone repair. Mathematical notation and tables lose their two-dimensional meaning when forced into a one-dimensional stream, and single-word presentation makes that worse, not better: a table read out cell by cell is noise. And a scanned page yields nothing, because there is nothing there to yield: we do not run OCR, and we would rather say so than pretend.

Because the reconstruction can fail silently, it is worth knowing the symptoms — they are the same whichever tool is doing the guessing, and spotting one early can save an hour of confidently misread report.

• Sentences that end nowhere and restart mid-thought — the classic mark of interleaved columns being read across instead of down.
• Running headers or page numbers surfacing inside paragraphs every few hundred words, on a suspiciously regular rhythm.
• Words split by hyphens that make no sense, or fused pairs where a line-break vanished without leaving a space.
• Captions, footnotes or sidebar text arriving at the wrong moment, mid-argument, in a different register from the prose around it.

Triage for the reader stuck with PDFs

Practical advice, in the order we would apply it. First, establish which kind of PDF you are holding: try to select some text. If nothing selects, it is a scan, and no reflow tool or reader will help until it has been through OCR — modern OCR is good, widely available, and turns the photograph back into text, imperfectly but usably.

If text selects, try to escape the format before fighting it. A remarkable fraction of PDFs are secondary: the paper has an HTML version at the publisher or a preprint server, the report has a web edition, the manual has an online copy. Ten seconds of searching often beats any amount of parsing, because the HTML version still contains the paragraphs as paragraphs. The same document as an EPUB is a different world again — that comparison is the subject of e-book apps and their typography.

If the PDF is the only copy, born-digital and single-column, almost any tool’s reflow will do, ours included — this is the case the heuristics handle. If it is multi-column or heavily designed, be suspicious of every reflow: spot-check the output against the page for a paragraph or two before trusting an hour of reading to it. And for the dense, hostile, table-ridden cases, the honest answer is sometimes the oldest one — a big screen, the actual layout, and your own eyes doing the navigation the file assumed they would.

The larger lesson generalises past this one format. PDF is what happens when presentation is stored and meaning is discarded; every reading tool since has been paying to reverse that decision, one guessed column at a time. The formats that store meaning — HTML, EPUB, even plain text — cost nothing extra to produce and stay readable by machines, by assistive software, and by whatever reading tools exist in twenty years. If you produce documents, the choice sits in your export dialog. Choose the one that keeps the paragraphs.