Two-column PDFs: diagnosing and fixing reading order
Last reviewed
A two-column paper that converts badly does not look slightly wrong. It looks like two documents shuffled together: a clause from the left column, then a clause from the right, then back again, for forty pages.
This is the single most alarming conversion failure, and also one of the most predictable. It has one cause, that cause is measurable, and you can usually tell within ten seconds of reading the output whether it happened.
8 minute read
Why it happens at all
A PDF stores glyphs at coordinates. It does not record that the page has two columns, and it does not record that the left one should be read to the bottom before the right one starts. Reading order has to be reconstructed, and the only evidence available is where the text sits.
If a converter reads a page in pure top-to-bottom order, every visual line that spans the page picks up text from both columns at once. The result is exactly the interleaving above. So before anything else happens, the page has to be tested for a column split, and the text divided into groups that are read in sequence rather than in parallel.
What the gutter test measures
The test is geometric. The page width is divided into a hundred vertical bins, every text run marks the bins it overlaps as occupied, and the widest surviving run of empty bins is a candidate gutter.
A candidate has to clear several thresholds before the page is treated as two columns, and each threshold exists to reject a specific false positive:
| Gutter width | At least 3 bins, and at least one body-text height across |
|---|---|
| Gutter position | Midpoint between 30% and 70% of the page width |
| Column balance | Each side holds at least 20% of the page's text runs |
| Straddling text | No more than 8% of runs cross the gutter |
| Page density | At least 24 text runs on the page |
Each threshold rejects a real mistake
Read that list as a set of refusals rather than a set of requirements and the design becomes clearer.
- The width rule refuses to treat word spacing in centred text as a column break. A gutter narrower than a single character almost never is one.
- The position rule refuses to split a page down a wide left margin, which would put the margin notes in one column and the entire document in the other.
- The balance rule refuses a split where one side holds four words. That is a pull quote or a page number, not a column.
- The straddling rule is the important one. A banner title, a full-width figure or a rule across the page crosses the gutter. A few of those are normal and they stay with the left column so their vertical order survives. Many of them mean the page is not really two columns at all, and the split is abandoned.
- The density rule refuses to analyse a title page, where there is not enough text for the geometry to mean anything.
The layouts that defeat it
Every one of these is common in real academic and technical publishing, and each fails in a way you can predict from the rules above.
| Three columns | Only one gutter is found; two columns get merged |
|---|---|
| Unequal columns | The gutter midpoint falls outside the middle 40% |
| Full-width figure mid-page | Too much straddling text; the split is abandoned |
| A tight gutter | Narrower than the body height, so it reads as word spacing |
| First page of a paper | Single-column abstract over two-column body |
| Footnote rules across the page | Straddling runs push past the 8% ceiling |
Diagnosing it in ten seconds
Read the first four lines of the converted Markdown against the first four lines of the left column of the PDF. That is the entire diagnostic.
If they match, the split worked and the rest of the document almost certainly did too, because the test runs per page against the same layout. If line one matches and line two comes from the right column, detection failed and the whole document is affected.
There is a subtler failure worth knowing about: the split can succeed on most pages and fail on the one page carrying a full-width table. That page alone will be interleaved, and it is easy to miss because the document either side of it reads perfectly. If a paper contains one wide figure or table, check the pages around it specifically.
Fixing a document that interleaved
Repairing interleaved text by hand is not realistic past about two pages, and it is not the right approach anyway. Change the input instead.
- Crop the PDF to one column and convert twice, once per side, then concatenate. Most PDF tools can crop a page range in one operation, and this is by far the most reliable fix.
- Convert page ranges separately if only some pages are affected. The good pages keep their good conversion.
- Look for the publisher's single-column version. Many journals publish an author manuscript or an accepted version that is single-column, and it converts cleanly with no work at all.
- Look for the source. A paper written in a word processor or a typesetting system has an original that states its structure instead of implying it, and converting that is always better.
What to check even when it worked
A successful column split still leaves two things worth verifying, because both are silent.
Footnotes sit at the bottom of a column and are often set smaller than the body. They convert as ordinary paragraphs at the point where the column ends, which means a footnote can land in the middle of a sentence that continues on the next page. Search the output for the small numeric markers and move the notes to where you want them.
Figure captions behave the same way. A caption under a figure in the left column appears in the text stream at that position, not attached to anything. In a document you intend to keep, it is worth demoting captions to their own lines and marking them, so a later reader can tell a caption from an assertion.