Multi-Column PDF Reading Order: Why Text Gets Scrambled

Document Automation

Multi-Column PDF Reading Order: Why Text Gets Scrambled

A multi-column PDF scrambles text because the file stores characters as positioned fragments on a page, not as a stream of paragraphs in reading order. Software has to guess which fragment follows which, and on a two-column page a wrong guess splices the end of a left-column line onto the start of the matching right-column line. HiDocument analyses PDFs, Word documents and images, and like any tool that reads PDFs it is only as reliable as the text order it receives.

Reading order is the sequence in which a human would read a page: down the first column, then down the second, with sidebars and footnotes handled separately. A native PDF rarely records that sequence explicitly, which is why contracts with side-by-side terms, term sheets, and scanned journals are the files most likely to produce a confident but wrong summary.

Why does a multi-column layout break text extraction?

Extraction breaks because a PDF is a drawing instruction set, not a document model. Each text run is placed at x and y coordinates, and the order in which runs appear in the file is whatever order the authoring program wrote them in.

Most extractors fall back on a geometric heuristic: sort by vertical position, then by horizontal position. That works for single-column pages. On two columns, it reads one full line across both columns before moving down, so you get a sentence made of half of column one and half of column two.

  • Line-by-line sorting merges neighboring columns into one garbled sentence.
  • Block detection works better but fails when columns have uneven widths or a heading spans both.
  • Tagged PDFs can carry an explicit reading order, but many contracts exported from word processors or scanners are untagged.
  • Scanned pages add OCR on top, so the engine must find the columns before it can order them. Our guide to why OCR fails on scanned PDFs covers that earlier stage.

What does scrambled reading order look like in a real document?

It looks like a clause that almost makes sense. The grammar is intact inside each fragment, so a model reading the output will often smooth over the join and produce a fluent summary of text that the document never actually says.

Consider a two-column license where the left column ends a sentence with "the Licensee shall not" and the right column, one line level with it, begins "Either party may terminate on thirty days notice." A line-by-line extractor produces "the Licensee shall not Either party may terminate on thirty days notice", and the negation is now attached to the wrong idea.

Common symptoms to check for:

  1. Sentences that change topic in the middle of a line.
  2. Section numbers that appear out of sequence, such as 4.1, 5.1, 4.2, 5.2.
  3. Table cells whose labels and values are separated by unrelated text.
  4. Footnotes or page headers that show up inside a paragraph.
  5. A summary that mentions only half of the obligations you can see on the page.

Which document types are most at risk?

The riskiest documents are those designed for printing in narrow columns or with mixed regions. The table below ranks common business documents by how often their layout, not their content, causes extraction trouble.

Document typeLayout featureTypical failureRisk
Standard form contract, single columnOne text flowHeaders and footers repeated in bodyLow
Two-column terms and conditionsParallel columnsLines merged across columnsHigh
Term sheet with side-by-side termsLabel column plus value columnLabels detached from valuesHigh
Insurance policy with sidebarsCallout boxes and margin notesSidebar text inserted mid-paragraphMedium
Scanned legacy agreement with annotationsSkewed pages, handwritingHandwritten notes misread or droppedHigh
Bilingual agreementTwo languages side by sideLanguages interleavedHigh

How can you check reading order before you trust a summary?

You check it by comparing a small sample of extracted text against the page itself. The test takes a few minutes and catches the failure that matters most, which is a confident summary of text that was never written.

  1. Open the original PDF and pick the densest page, usually the one with the most columns or the key commercial terms.
  2. Select a paragraph in the PDF viewer and paste it somewhere neutral. If the pasted text is already scrambled, the file itself is the problem, not the analysis tool.
  3. Ask the analysis tool to quote the exact wording of one clause, then compare it to the page word for word.
  4. Check that section numbers run in order in the extracted text.
  5. Repeat on one page from the middle and one from the end, because layouts often change after the first page.

If you review many files at once, the manual check does not scale well. That is the point where a tool that lets you ask questions of the document and read exact quotations back saves effort: you can use the document grader to get a quick view of a file, then verify the specific passages that drive your decision against the source page. A tool can narrow where you look, but it does not replace reading the operative clauses.

What fixes work when the order is wrong?

The most reliable fix is to give the analyzer a cleaner source. A single-column version of the same document nearly always extracts better than a two-column one, so request an editable copy from the sender whenever you can.

  • Ask for the Word original. Word files hold paragraphs in order, so the column layout stops mattering. HiDocument accepts Word documents as well as PDFs.
  • Export each column as its own page region. Cropping a two-column page into two single-column images is tedious, but it forces a correct order for the pages that matter.
  • Re-save from a tagged source. Exporting from a word processor with accessibility tags on often records the reading order explicitly.
  • Split mixed pages. Separate a table or a sidebar from the running text so the extractor treats each as its own region.
  • Verify by quotation. Whatever the fix, confirm the key clauses by quoting them against the page.

For table-heavy pages, the issues overlap with the ones described in how AI tools extract tables from PDF documents, and for very long files the page-count effects are covered in why a 300-page PDF behaves differently from a 3-page one.

What about handwritten annotations in the margins?

Handwritten margin notes are the hardest case, because they sit outside the printed column structure and often overlap it. Printed-text OCR is built for regular type, so handwriting is frequently skipped, misread, or merged into the nearest printed line.

Treat any handwritten change to a contract as unverified until a person has read it on the page. A struck-out clause or a handwritten initial can change the meaning of a printed paragraph, and extracted text will usually not tell you that the change exists. If a signed copy carries annotations, record them in a separate note and keep the scanned image with the analysis.

Is it worth paying for a tool when a layout is the problem?

The most common objection is that a better tool should not need help from you. That is fair, and layout-aware extraction does keep improving, but no tool can promise a correct reading order on every file, and you should be wary of any that does. The practical answer is to budget a few minutes of verification per file rather than to expect perfection.

Cost matters too. The free plan includes 10 analyses a month with a 5 MB file cap, which is enough to test your own two-column files and see how they behave. Pro is $12 a month with a 10 MB file cap and a 10-file bulk limit. HiDocument is a document analysis tool, not a law firm, and its output is not legal advice.

People also ask

Can a model fix reading order by itself?

Sometimes. A language model can often infer that two fragments do not belong together, but it can also smooth the join into a plausible sentence. Verify the wording of any operative clause against the page.

Does OCR change the reading order problem?

Yes, because OCR adds a layout-detection step before text order is decided. A skewed or low-resolution scan makes column detection less reliable.

Are Word files safer than PDFs?

Generally yes for reading order, since paragraphs are stored in sequence. Text boxes and floating tables can still sit outside the main flow, so spot-check those as well.

What is the next step?

Take one two-column file you already work with, run it through the check above, and note whether the clause quotations match the page. You can create a free HiDocument account, upload the file under the 5 MB cap, and ask the document chat to quote a specific clause. If the quote matches the page, you have a baseline for that file type. If it does not, ask the sender for the Word original. See plan limits and pricing if you need the 10 MB cap or bulk mode.

Frequently Asked Questions

Why does text from a two-column PDF come out mixed up?

A PDF stores text as positioned fragments rather than ordered paragraphs, so software must guess the reading order. Simple extractors sort by line and read straight across both columns, which merges the end of a left-column line with the start of a right-column line. Layout-aware extraction reduces this but does not eliminate it.

How can I tell if a PDF has a reading order problem?

Select a paragraph in a PDF viewer and paste it into a plain text editor. If sentences change topic mid-line or section numbers run out of sequence, the order is wrong. Then ask your analysis tool to quote one clause and compare it to the page word for word.

Is a Word document better than a PDF for AI analysis?

For reading order, usually yes, because a Word file stores paragraphs in sequence and column layout does not affect it. Floating text boxes and tables can still sit outside the main flow, so check those separately. HiDocument analyses both Word documents and PDFs.

Can AI read handwritten notes in a contract margin?

Not reliably. Printed-text OCR is designed for regular type, so handwriting is often skipped, misread, or merged into a nearby printed line. Treat any handwritten change as unverified until a person has read it on the page, and keep the scanned image alongside the analysis.

Does a confident AI summary mean the text was extracted correctly?

No. Fluent wording says nothing about whether the source text arrived in the right order, because a model can smooth over a bad join and still sound certain. Confirm important clauses by comparing exact quotations against the original page before you rely on a summary.

Ready to analyze your own documents?

Upload any PDF, Word doc, or image — get 10 types of AI analysis instantly. Free to start, no credit card required.

Try HiDocument Free →

Related Articles