On this page
A narrator starts a sentence in the left column, reads a sentence from the right, and returns to the first argument halfway through a thought. Every word sounds clear. The document makes little sense. The first thing to inspect is the text sequence delivered to the narrator.
This guide gives you a page you can inspect, a known intended order, and actual extraction output. Use it to separate a file-structure problem from a recognition problem and from a playback problem. Then repeat the relevant checks on the hardest page of your own document.
Start with three different questions
Swipe horizontally to compare all columns →
| Check | What to do | What a failure suggests |
|---|---|---|
| Are the words available? | Copy one ordinary sentence into a plain-text editor. | An image-only page, inaccessible text, or an extraction problem. |
| Are they in the right order? | Copy two complete paragraphs and the transition to the next column. | The extraction or document structure may not preserve the intended sequence. |
| Does playback follow that text? | Listen to the same passage while following the source. | If the transcript is correct, inspect the reader's segmentation, position, or pronunciation. |
Being able to select a few words answers only the first question. A whole-page image can also have a previously generated OCR layer, so successful selection does not prove that the text came directly from an authoring document. Save the pasted passage; it is useful evidence when you compare two readers or report a problem.
The reproducible example: same page, different sequence
We created two original one-page PDFs. Each has four paragraphs in a left column, labeled L1 through L4, and four in a right column, labeled R1 through R4. The intended prose sequence is the complete left column followed by the complete right column. The labels are visible in the files, so you do not need to infer the intended order from a screenshot.
In the first source, paragraph elements alternate between left and right. In the second, all left-column paragraphs precede all right-column paragraphs. Both were exported by Chromium with PDF tags enabled. PDF.js reported a structure tree on each page. We then requested the text items and preserved the order returned by the API.
Swipe horizontally to compare all columns →
| Fixture | Observed paragraph sequence | Matches intended prose order? |
|---|---|---|
| Interleaved source PDF | L1 → R1 → L2 → R2 → L3 → R3 → L4 → R4 | No |
| Corrected source PDF | L1 → L2 → L3 → L4 → R1 → R2 → R3 → R4 | Yes, for these eight paragraphs |
Both files produced 22 nonempty text items in this run; both contained the eight prose paragraphs. The number of extracted items therefore did not distinguish the useful transcript from the interleaved one. Compare the first raw output with the corrected raw output and the intended paragraph transcript.
This result is deliberately narrow. It shows what PDF.js 5.6.205 returned for these fixtures, without layout reconstruction. A reader that uses other ordering logic may behave differently. We did not traverse the accessibility tags to produce these transcripts and did not test NVDA, VoiceOver, or a commercial reader. Presence of tags alone does not certify their correctness.
L1 R1 L2 R2 L3 R3 L4 R4L1 L2 L3 L4 R1 R2 R3 R4Why a correct-looking page can have an unusable order
W3C describes logical order in tagged PDFs primarily through the tag order, with content order inside individual tags also relevant. Its two-column example warns that incorrect tagging can cause text to be read across columns. It recommends structuring the source correctly before export, then checking complex layouts after conversion. See W3C technique PDF3.
For a reader, the practical distinction is between a visible placement and an intended sequence. Our fixtures place both columns correctly, but their source paragraph order differs. A method that simply follows the extracted sequence inherits that difference. A method that reconstructs a sequence has another job to perform, and its proposed order still needs checking.
A five-minute inspection for your own PDF
- Choose three pages: a normal prose page, the densest layout, and a page boundary where an argument continues.
- Write the intended transition: the last sentence of the left column and the first sentence of the right are enough to expose a common failure.
- Copy the passage: preserve the raw output before removing line breaks, captions, or repeated headers.
- Listen to that exact passage: record where narration first differs from the intended sequence.
- Repeat in a second viewer if needed: use the same file and passage. A difference helps localize the problem; it does not prove that either reader will handle the entire document correctly.
Name the observation precisely: “page 8 reads the figure caption between the two halves of a sentence” is actionable. “The voice is bad” is not enough to distinguish order from pronunciation. Include the page number, expected transition, actual transition, reader version, and whether a copied transcript shows the same issue.
Choose a repair that addresses the observed failure
If you control the original document
Correct the source structure and export again. Our corrected fixture does exactly this: it changes the paragraph order in the source while keeping the two-column appearance. Retest the exported file rather than assuming the source edit guarantees the final result. This also makes the correction repeatable when the document is updated.
If you have only the PDF
Adobe documents numbered reading-order regions and tools for moving them, splitting incorrectly combined regions, and handling headings, captions, and figures. In Acrobat Pro, its current guide starts from All tools → Prepare for accessibility → Reading Order. Save a separate copy before editing. Complex structures may need the Tags panel rather than a simple region move. Follow the Adobe Reading Order guide for the document you have.
If you need to listen now
Prepare a checked listening copy one section or column at a time. Include the section heading, preserve the paragraph order, and keep page references beside the text. Verify the beginning and end of every copied region so that manual selection does not silently drop a line. This can be a useful personal workaround; it is not a repaired accessible PDF to redistribute.
If the page has no usable text
Use the scanned-PDF OCR workflow. Recognition recovers words from images. It does not automatically establish which columns, captions, or notes belong together. Inspect the recovered words and their sequence as separate checks.
Handle notes, captions, and tables deliberately
Swipe horizontally to compare all columns →
| Content | Question to answer | Useful listening treatment |
|---|---|---|
| Footnote | Does the argument need it immediately? | Keep the reference and make the note reachable; test return to the main sentence. |
| Figure caption | Which paragraph introduces the figure? | Keep the caption with its figure and inspect the visual when it carries the evidence. |
| Table | Do values depend on row and column headings? | Read labels with values; avoid a flattened list that loses those relationships. |
| Repeated header | Is this page furniture or part of the argument? | Omit it only from a checked listening copy, preserving useful section headings. |
| Equation | Would spoken symbols preserve the relationship? | Return to the original notation and surrounding explanation. |
These are editorial decisions about a particular document, not a rule to delete every bracket, number, or short line. A superscript can mark a footnote, a power, or a reference. An automatic cleanup that treats them identically can change the meaning while making narration sound smoother.
Check the listening copy before a long audio export
Replay the first complete paragraph, every type of layout transition, and the last paragraph of the chosen range. Check at least one name, one number, and any sentence containing a negation. If the content supports a decision where exact wording matters, verify the relevant passages directly rather than treating this sample check as validation of the whole file.
You can use iReadAll PDF Voice Reader to keep the page available while listening. The same inspection applies to any reader. A successful run on our eight-paragraph fixture is a useful diagnostic result, not evidence that another document's formulas, tables, or notes will be interpreted correctly.
Samples & worksheets
One original page with labeled paragraphs in two visual columns.
The same prose and layout, authored column by column.
The eight prose paragraphs in their intended order.
Unedited text from the interleaved fixture, including page furniture.
Machine-readable extraction results for the diagnostic PDFs and test book.
Questions & answers
Why does my PDF reader jump between columns?+
The text sequence delivered to narration may alternate between visually separate columns. Copy the problematic passage to inspect its order, then compare it with playback. The file structure and the reader's extraction strategy both matter.
Does a tagged PDF always have the correct reading order?+
No. Both original PDFs in our demonstration contain a structure tree, but PDF.js returned different paragraph sequences. Tags need to be correct, and an application may use a different extraction path. Tag presence alone is not an accessibility certification.
Will OCR fix a PDF that already has text?+
Not necessarily. OCR can recover words from an image, but a sequencing problem can remain. Diagnose the existing text before replacing it, and inspect any OCR output against the page.
Did you test commercial PDF readers with these samples?+
No. We measured raw PDF.js 5.6.205 extraction from original teaching fixtures and published the output. Results for specific reading applications, screen readers, and other PDFs must be tested separately.
How this guide was researched
We authored two one-page PDFs containing the same original prose and visual two-column layout, changing the source document order. Chromium 149.0.7827.55 generated both with tagging enabled. We extracted text using PDF.js 5.6.205 getTextContent without reordering its items. The downloadable raw transcripts and JSON report record the observed sequences. This is a constructed diagnostic demonstration, not a comparison of commercial readers, a screen-reader test, or an accessibility certification.
Hands-on work recorded Sep 9, 2026. The method and sample scope above define what was tested.
Evidence checked Sep 9, 2026. Product features can change; dated observations describe the version examined.
Published by iReadAll. AI tools assist research and editing; the team is responsible for verification. Read our editorial policy.
Sources & references
- W3C PDF3: Ensuring correct tab and reading orderW3C · Checked 2026-09-09
- Adobe: Reading Order tool for PDFsAdobe · Published 2026-02-26 · Checked 2026-09-09
- PDF.js PDFPageProxy API: text content and structure treeMozilla · Checked 2026-09-09
- iReadAll original PDF extraction report and reproducible fixture resultsiReadAll · Published 2026-09-09 · Checked 2026-09-09
Cite this guide
iReadAll Product & Editorial Team. “PDF Reading Order Problems: How to Diagnose and Fix Them.” iReadAll, Oct 5, 2026. https://ireadall.ai/blog/pdf-reading-order-problems
