We publish the same resume designs three ways: as Google Docs templates, as LibreOffice Writer templates, and, for one design, as atsresume, a LaTeX class on CTAN. On the page they look nearly the same. OpenResume, the parser we test with, reads the PDF's text layer: which strings were written, at what position, and where one line ends and the next begins.
So I ran each export through the harness we run on every template: OpenResume's open-source parser pinned to commit 4f8255a (pdf.js 3.7.107), a field-by-field comparison against values written down before the run, and pdfminer.six for the raw text. The parser is one open-source project, not Workday or Greenhouse. If you'd rather query numbers than read about them, the raw Google Docs results are on Kaggle as two CSVs, one row per layout and one per font.
Results
US Letter, OpenResume fields read correctly:
| Template | Google Docs | LibreOffice 26.8, font installed | LibreOffice 26.8, font missing | Writer download |
|---|---|---|---|---|
| Classic (Georgia, education first) | 38/38 | 38/38 | 30/38 | Classic resume template |
| Plain (Arial) | 33/33 | 33/33 | 33/33 | Plain resume template |
| Student (Georgia) | 35/35 | 35/35 | 31/35 | Student resume template |
"Font missing" means LibreOffice 26.8 on macOS, run headless so it could see only the fonts bundled with LibreOffice. Writer replaced Georgia with Linux Libertine G and Arial with Liberation Sans.
With the fonts installed, LibreOffice and Google Docs gave the parser the same fields and the same character counts per font: 1,320 regular, 231 bold and 147 italic Georgia characters in Classic, in both files.
The matching cover letters for Writer have no fields for the parser, so they get the raw-text check only: every sentence present, in order, on one page. All four files (Plain and Student, Letter and A4) passed.
The interesting rows are the ones where something went wrong.
Look at glyphs, not the page
The first thing I check in any PDF is whether a single glyph stands for more than one character:
from pdfminer.high_level import extract_pages
from pdfminer.layout import LTChar
def multi_char_glyphs(path):
found = []
def walk(node):
if isinstance(node, LTChar):
if len(node.get_text()) > 1:
found.append((node.get_text(), node.fontname.split("+")[-1]))
elif hasattr(node, "__iter__"):
for child in node:
walk(child)
for page in extract_pages(path):
walk(page)
return found
A ligature isn't a problem by itself. When the glyph for "fi" maps back to the two letters f and i, extraction is fine. With DejaVu Serif, Classic had five fi and fl glyphs and still read 38 of 38. The trouble starts when the mapping is wrong.
Fallback font 1: Linux Libertine G
With Linux Libertine G, Classic had ten ligature glyphs, and nine of them carried one character too many. The glyph drawn for "fi" in "first" carries the text fir, and then the real r follows. Here's what pdf.js returns:
{"str":"firr","x":296.55,"y":385.55,"hasEOL":false}
{"str":"st-year economics students","x":306.35,"y":385.55,"hasEOL":false}
The parser read the bullet as ...policy memos forfirrst-year economics students. The missing space before it is the same glued-word bug we saw with Google Docs fonts. The doubled letter is new. The same PDF gave fluuent, Theeory, Drafteed, and a control character in the middle of the email address. pdfminer reads the same doubled letters, so this is in the file, not in one reader. Classic dropped to 30 of 38 and Student to 31 of 35 (Pipettiing, aft er).
Plain was fine. With Liberation Sans the PDF had no ligature glyphs at all, and it read 33 of 33.
Fallback font 2: same font, older LibreOffice
The other bad row isn't in the table. On Linux with LibreOffice 7.4 and no Georgia installed, fontconfig picked DejaVu Serif, and Classic read 21 of 38.
Re-exporting with DejaVu Serif on LibreOffice 26.8 read 38 of 38, so DejaVu Serif can produce a PDF that reads correctly. I haven't isolated what caused the 7.4 failure. The text items:
LibreOffice 7.4
{"str":"New York, NY","x":493.7,"y":581.3,"hasEOL":false}
{"str":"Investment Banking Analyst Intern","x":43.3,"y":567.9,"hasEOL":false}
LibreOffice 26.8
{"str":"New York, NY","x":493.75,"y":581.35,"hasEOL":false}
{"str":"","x":43.3,"y":567.9,"hasEOL":true}
{"str":"Investment Banking Analyst Intern","x":43.3,"y":567.9,"hasEOL":false}
The y coordinate drops by about 13.4pt in both files, but only 26.8 has the empty end-of-line item. OpenResume starts a new line only at hasEOL, so with 7.4 the city ran into the next line: the job title came back as New York, NYInvestment Banking Analyst Intern, and the parser found six jobs instead of three. I've only tested these two versions, so I can't say where in between it changed.
LaTeX: you write the text layer yourself
With Writer or Google Docs you get whatever text layer the exporter writes. In LaTeX it's set in the class file. The parts of atsresume.cls that matter here, trimmed:
% pdfLaTeX
\DisableLigatures[f]{encoding = *} % microtype: fi and fl stay two glyphs
\input{glyphtounicode}
\pdfgentounicode=1 % every glyph maps back to Unicode
% LuaLaTeX and XeLaTeX
\setmainfont{texgyrepagella}[..., Ligatures = {TeX, NoCommon}]
% no word split across lines
\hyphenpenalty=10000
\exhyphenpenalty=10000
% places and dates: pushed right with glue, not a table
\unskip\nobreak\hfil\penalty50\hskip1em\hbox{}\nobreak\hfil
The example resume uses the same person and text as Classic. Compiled with pdfLaTeX, LuaLaTeX and XeLaTeX, on Letter and A4, all six PDFs read 38 of 38 with zero ligature glyphs. The full check (18 PDFs including edge cases, under TeX Live 2026 and the TeX Live packages in Debian 12 and Ubuntu 22.04) passed with all three.
Two differences showed up anyway. TeX turns a typed apostrophe into a curly one, so "professor's" comes out as professor’s. That was the one change our expected values needed. And the extractor matters: pdftotext -layout kept every place and date beside its line, while plain pdftotext guessed at columns and moved "Providence, RI" and "Sep 2023 – Dec 2024" elsewhere in their entries, usually after the bullets. Same PDF, different reading order, depending on the tool.
Google Docs, for comparison
Google Docs gives you the least control over the text layer. In our tests, 16 of 32 fonts saved letter pairs as ligatures, and a text box that Google Docs exported as a picture took the name and contact details out of the text entirely (34 of 46 fields). I went through how the Google Docs side is tested in an earlier post.
What I'd check on any export
- Run
pdffontson the PDF. If the font in the PDF isn't the one you picked, find out what replaced it first. - Run
multi_char_glyphs()and look at what each glyph maps to. Compare each glyph's extracted text with the letters it represents. In the Libertine export, the glyph forfireturnedfir. - Dump pdf.js text items for one entry and check that each line ends with a
hasEOLitem.
Limits: one parser, one sample per template, US Letter. "Font missing" was simulated with a headless run, and the LibreOffice 7.4 result is one export on one machine.
I run ReadableResume. More on how the templates are tested against resume parsers.
I used AI to help draft this post.
Top comments (0)