Guide

How to produce a tagged PDF from LaTeX

LaTeX's tagging project reached usable in 2024. What works now, what still needs post-processing, and the maths problem nobody has solved.

About 15 minutes · involved

The state of it

For twenty years the honest answer to “can LaTeX make an accessible PDF” was no. That changed with the LaTeX Project’s tagged-PDF work, which has been usable since the 2024 format releases and continues to develop.

What follows is what works now. Check the current latex-lab documentation before relying on any specific detail, because this is moving.

The minimum that works

\DocumentMetadata{
  lang        = en-GB,
  pdfversion  = 2.0,
  pdfstandard = ua-2,
  testphase   = {phase-III, math},
}
\documentclass{article}

\usepackage{hyperref}
\hypersetup{
  pdftitle            = {A Structured Title, Not the Filename},
  pdfdisplaydoctitle  = true,
}

\begin{document}
\section{Introduction}
...
\end{document}

Three things about this preamble matter.

\DocumentMetadata must come before \documentclass. It is not a package; it configures the format before the class loads. Putting it after silently disables tagging.

testphase selects how much of the tagging code is active. Higher phases tag more and are less settled. phase-III handles sectioning, lists, tables, links and floats. Add math for equation tagging, which is the newest and least stable part.

pdfdisplaydoctitle = true is not optional. Without it a reader announces the filename regardless of pdftitle. See the title guide.

Compile with LuaTeX. lualatex document.tex. The other engines do not support the tagging callbacks.

What comes out correct

• Sectioning. \section, \subsection and friends become H1…H6 with correct levels, provided you do not skip levels in the source. Skipping in LaTeX produces skipping in the PDF — condition 14-003.
• Lists. itemize, enumerate and description become proper L/LI/LBody.
• Links. hyperref’s links get /Link structure elements, not just annotations.
• Tables. tabular becomes a Table with TR and TD. Header rows need to be declared — see below.
• Language. lang in \DocumentMetadata sets /Lang, and babel’s language switches emit per-run language, which is more than most word processors manage.

What still needs work

Table headers

tabular has no concept of a header row; it is a grid of cells. The tagging code cannot infer which row is a header, so you must say:

\begin{tblr}{
  row{1} = {font=\bfseries},
  cell{1}{1-3} = {cmd=\TableHeader},
}

Or use \rowstyle from tabularray. Without it every cell is a TD and the table fails 15-003. This is the most common LaTeX tagging failure.

Floats

A figure or table float appears wherever LaTeX put it, which is frequently not where it is discussed. Structurally it is tagged in the order the float was placed, so a figure pushed to the top of the next page is read there.

There is no automatic fix, because the correct reading position is a judgement about your document. Either constrain float placement with [h] and \FloatBarrier, or fix the order in post-processing.

Alt text on graphics

\includegraphics produces an image with no alt text. Provide it:

\includegraphics[alt={Revenue rose from 4.2 to 6.8 million between 2021 and 2025}]{revenue.pdf}

The alt key needs a recent graphicx. For a decorative rule or ornament, use artifact instead so it is excluded from the structure rather than given empty alt text.

For a chart, the alt text should carry the finding, not describe the picture — see writing alt text. A data-heavy figure usually also wants the underlying numbers in a table somewhere, because no sentence substitutes for twelve data points.

The maths problem

This is where honesty is more useful than optimism.

PDF has no native representation for mathematical semantics. The available routes are:

Associated MathML — PDF 2.0 allows a structure element to carry an associated file containing MathML, and this is the technically correct answer. Reader support is uneven: some assistive technologies use it, many ignore it, and PDF/UA-1 does not standardise it at all. If your target is PDF/UA-1 rather than 2.0, this route is outside the standard you are conforming to.

/ActualText on the equation — a linear string, like x squared plus y squared equals z squared. Universally supported and semantically flat. A reader cannot explore the expression, navigate its structure, or re-hear a sub-term, which is precisely what mathematical reading requires.

Alt text — a description. Adequate for a decorative equation, inadequate for anything a reader needs to follow.

The math testphase generates structure and associated MathML for LaTeX’s own maths. It is the best available option, and you should still expect a reader with a screen reader and a maths document to be doing real work that the file does not do for them.

If the document is a teaching text or anything where the mathematics is the content, consider producing an HTML version alongside the PDF. HTML with MathML has genuinely good support; PDF does not. Publishing both is not a defeat, it is the correct engineering answer.

Hebrew and Arabic in LaTeX

LuaTeX with babel’s bidi support produces correctly ordered text in the output — that is, the stored characters are in logical order, which is right. Do check it rather than assume it, because some older packages, and bidi under XeTeX in particular, reverse text at the source level and produce exactly the visual-order failure described in TG-RTL-001.

The two-keystroke check applies: compile, open, select all, copy, paste into a plain-text editor, and look at where the Hebrew final letters fall.

Verify the output

The tagging code is young and your document is probably doing something it has not seen. Do not skip this.

lualatex document.tex
verapdf --flavour ua1 --format text document.pdf

Then the human check: select all, copy, paste, and read the order. Floats and multi-column layouts are where LaTeX’s output most often surprises, and no validator reports either.

Related failure conditions

• 01-005 Content is neither marked as Artifact nor tagged as real content Every text-showing, image-painting and path-painting operator on a page must sit either inside a marked-content sequence that maps to the structure tree, or inside an artifact sequence. Content that is in neither is orphaned.
• 06-001 Document does not contain an XMP metadata stream The catalog has no /Metadata entry pointing at an XMP packet. PDF/UA-1 requires one — it is where the conformance claim and the document title live.
• 11-001 Natural language for text in page content cannot be determined No /Lang is in effect for some text: neither on the catalog, nor on an enclosing structure element, nor on a marked-content sequence.
• 14-003 Numbered heading levels in descending sequence are skipped The heading hierarchy jumps a level going down — H1 straight to H3. Going back up any number of levels is fine; going down must be one level at a time.
• 13-001 Graphics objects are not tagged with a Figure tag An image or vector drawing that carries meaning is not inside a Figure element. Either it is untagged, or it has been swept into a paragraph.

Last updated 2026-09-10 .