A PDF draws text as numbered glyph shapes, and copying gives you letters only if the file also says which character each glyph stands for. Without that, a flawless page can paste as nothing. We built 15 one-page PDFs in Chinese, Japanese, Korean and Thai whose fonts carry no such map. Rendered by poppler, each was pixel-identical to its correctly mapped twin. Yet pdf.js 4.10.38, poppler's pdftotext 26.10.0 and pypdf 6.1.1 returned none of their 330 characters. Thai has a second problem: in our correctly mapped Thai file, the vowel ำ came out split in every extractor we ran.
Each PDF page is an inscription in glyph numbers
A content stream says, in effect, "draw glyph number such-and-such of this font here". Turning the number back into a character is a separate, optional job. The PDF 2.0 table of Type 0 font entries, as amended on the PDF Association's ISO 32000-2:2020 errata page (last modified 15 August 2025), describes ToUnicode as "(Optional) A stream containing a CMap file that maps character codes to Unicode values".
An extractor can look in three places: that map, the character table (cmap) inside the embedded font, and Adobe's published maps for fonts naming a standard CJK collection. Finding neither a ToUnicode nor a standard collection, pdf.js 4.10.38 emits each glyph's code as if it were a Unicode code point, as its source shows. Over our 15 mapless files, three variants of five test lines, pdf.js, pdftotext and pypdf scored 0 of 330.
Renumbered glyphs paste as blanks, original ones as stray letters
Our subset fonts renumbered their glyphs from 1. Read as code points, small numbers are invisible control characters (U+0001 to U+001F). The simplified Chinese line gave 18 characters, 15 of them control codes: a blank line that is somehow not empty.
With the original numbering, which in a large CJK font runs into the thousands, the same fallback lands on visible letters. Our Chinese sentence came out as 20 characters from Telugu, Greek, Sinhala, box-drawing and other blocks. It is the reverse of why some languages come out as empty boxes, where the characters exist and no font can draw them.
pdfminer.six deciphers the map inside the font
pdfminer.six 20250506 was the exception. For a CID font with Identity coding and no ToUnicode, its source reads the embedded TrueType font's own cmap. It recovered 85 of 110 characters with renumbered glyphs and 81 of 110 with original numbers. All the Chinese, Japanese and Korean came back, 63 of 63 both times; Thai managed 22 and 18 of 47, because shaped Thai glyph variants have no entry in the font's cmap.
With that cmap deleted, every extractor scored 0 of 110. Among the mapless files, switching extractor helped in that one class only, and only partly for Thai.
Big5 and Shift-JIS files stay illegible without Adobe's tables
Some CJK PDFs embed no font, naming a standard one such as STSong-Light and leaving the lookup to Adobe's CMaps, published in the cmap-resources repository under a BSD licence. Our nine such files held 146 characters in Unicode, GBK, Big5, Shift-JIS and Korean UHC encodings. pdf.js without those maps returned 0 of 146 and with them all 146, as did pdftotext and pdfminer.six. pypdf 6.1.1 managed 96, failing three encodings.
Our PDF text extractor loads Adobe's maps, hence its 146. Until 18 November 2026 it showed a green "Done" over all 40 of our test files, including 20 with no correct character, passed control codes such as U+0000 to the clipboard, and printed Chrome 157's vertical column one character per line. The revision of that date names 20 of our files as having no usable character map and withholds their junk; one mapless Thai file still passes with a warning. The check reads characters, so a mapless font whose numbers land on plausible letters can still slip through.
Thai sara am breaks into separate marks on the page
Our correctly mapped Thai line, น้ำใจ กำลังทำงาน กตัญญู ญี่ปุ่น ผู้ใหญ่ ปู่ ฐานข้อมูล, has 47 characters; every extractor returned 44. All three misses were ำ (sara am, U+0E33), which the Unicode 17.0.0 character data decomposes, for compatibility, into NIKHAHIT plus SARA AA. HarfBuzz 11.3.3 drew it that way, putting the tone mark of น้ำ between the halves. Noto Sans Thai has a ready-made ำ glyph, and shaping with the font's layout tables deleted gave the same split, so the shaper made it.
Extracted, the word reads NO NU, NIKHAHIT, MAI THO, SARA AA; NFKC normalisation cannot repair that, because the original puts the tone mark first. The SARA AA glyph also stands for a plain า, so no map that works code by code can rebuild ำ. /ActualText can, by attaching the intended text to a span of glyphs. With it, pdftotext returned 47 of 47, even where the font had no cmap. pdf.js, pypdf and pdfminer.six ignored it; the pdf.js source reads only the marked-content ID from such spans.
In Chrome 157's print of the line, which carried 13 /ActualText spans, pdftotext again returned 47 and the other three 39: the three ำ plus five glyphs that came out as U+0000. We make no claim to have found the split first; our counts are a cross-check.
Right-to-left columns can come back left column first
Vertical Chinese and Japanese columns run right to left. Our three Identity-V files, 51 characters in two columns each, kept every character in every extractor, and all but pdfminer.six kept the order; it read the left column first, leaving 26 of 51 in order. Chrome 157 placed its 20-character vertical line one glyph at a time; pdf.js marked every glyph as a line end, and pdftotext interleaved the two columns row by row, leaving 11 in order.
Scans paste nothing, mapless fonts paste invisible characters
No selectable text at all means no text layer: a scan, or text turned into outlines. A paste that looks blank yet has a length points to a mapless font with renumbered glyphs. Greek or Sinhala letters in a Chinese document can point to the same missing map with original numbering. Thai that looks right but holds ำ as two marks is the shaping split; a search for ำ misses those words.
When the character map is missing from the PDF, the text simply isn't in the file; only the source document has it. An extractor that reads the font's own table recovered our Chinese, Japanese and Korean while that table survived, but only part of the Thai. OCR reads the drawn shapes, which poppler drew intact in every file we built.
We made the 40 files with our own PDF writer, fontTools 4.60.1 and uharfbuzz 0.51.1, using Noto Sans SC, Noto Sans Thai and Nanum Pen (SIL Open Font License 1.1), plus one headless Chrome 157.0.8103.79 print. pdf.js (Apache-2.0) ran in Node 26.13.0; poppler is GPL-2.0 or GPL-3.0, pypdf BSD-3-Clause and pdfminer.six MIT. A correct character is one in the longest common subsequence of expected and extracted text, whitespace removed, out of 888. We built every case, so none of this says how often real PDFs fail.