The same pages, half the bytes

A PDF that weighs more than it needs to costs something every time: when it is stored, when it is sent by mail, when it is downloaded on a telephone, when it is kept for ten years. The team measured the documents on this site before and after the engine learned to carry less that nothing reads, and not one of them lost a single mark.

The documents, before

7,686,790 bytes

The same documents, now

3,249,799 bytes

Saved

57.7 %

What did not change

The saving comes from what a document carries rather than from what it shows: an embedded typeface now describes only the characters the document actually draws, instead of every character the typeface knows how to draw. A three-page letter no longer carries an alphabet it never uses.

Every document, measured

Sizes in bytes, biggest saving first. Every figure in the right-hand column is the size of the file this site serves, so any row can be checked by downloading the document behind it.

Document Before Now Saved
Cutting out what you draw 107,155 36,391 66.0 %
Text and shapes filled with a gradient 108,742 38,009 65.0 %
Shrunk headings and a turned label 108,630 38,026 65.0 %
Breaking a paragraph up, and aligning it 108,987 38,284 64.9 %
Coloured cell borders and fills 109,060 38,350 64.8 %
Gradients in a line and in a circle 109,198 38,458 64.8 %
Transparency and blend modes 109,346 38,691 64.6 %
Text poured into columns 109,635 38,993 64.4 %
An invoice laid on a letterhead 218,913 77,869 64.4 %
A cell's margin, border, inner space and background 109,685 39,067 64.4 %
Paragraph indents and spacing 109,754 39,188 64.3 %
Break points after a slash or a hyphen 109,804 39,250 64.3 %
Soft hyphens and no-break spaces 109,923 39,394 64.2 %
Stamps fitted to a box 110,127 39,656 64.0 %
Grey, screen and printing colours 110,363 39,815 63.9 %
Changing style in the middle of a paragraph 110,563 40,057 63.8 %
Links and marks on placed text 110,681 40,181 63.7 %
Lists made from a hanging indent 110,883 40,462 63.5 %
One pattern, repeated across whatever it fills 111,351 40,764 63.4 %
One drawing, placed many times 111,086 40,758 63.3 %
Breaking an overlong word 111,252 40,892 63.2 %
A self-paginating table 223,298 82,269 63.2 %
Fonts kept inside the file 111,279 41,000 63.2 %
Content shown and hidden by layer 111,577 41,120 63.1 %
The same numbers, through spaces that say what they mean 111,666 41,191 63.1 %
Internal and external links 112,095 41,491 63.0 %
Columns measured from their cells 111,935 41,485 62.9 %
A contents page with dot leaders 111,913 41,580 62.8 %
Tabulations and their stops 112,426 42,022 62.6 %
Cells merged down the rows 112,649 42,396 62.4 %
Overlong text in a narrow cell 112,900 42,653 62.2 %
Colour the document defines 113,403 42,897 62.2 %
A sheet that measures what it says 113,086 42,783 62.2 %
Custom page labels 113,543 43,049 62.1 %
A chart of colours, and a drawing that names places in it 113,810 43,318 61.9 %
One hatch, laid down in as many colours as the page asks for 113,667 43,278 61.9 %
An interactive form 114,391 43,746 61.8 %
One rectangle, four masks 114,195 43,774 61.7 %
Functions, the shape a colour follows 114,356 44,092 61.4 %
A multi-page invoice table 114,981 44,608 61.2 %
Several plates at once, each at a strength of its own 114,968 44,672 61.1 %
A boarding pass 114,813 44,643 61.1 %
A picture fitted into a box 116,023 45,312 60.9 %
A sheet as it goes to the press 115,238 45,011 60.9 %
A tagged, accessible document 116,151 45,753 60.6 %
A handbook that says how it wants to be read 116,738 46,471 60.2 %
A sheet of shelf labels 116,867 46,733 60.0 %
The device a file's colours are meant for 119,232 49,002 58.9 %
A slideshow that runs itself 119,354 49,089 58.9 %
An archivable PDF/A-3 form 120,803 50,301 58.4 %
A Factur-X electronic invoice 120,878 50,373 58.3 %
An archival sheet carrying fields of its own 122,159 51,905 57.5 %
A register whose every part says where it came from 124,045 53,704 56.7 %
A landscape certificate 166,624 96,346 42.2 %
A delivery note 167,495 97,393 41.9 %
A tagged list whose column is a picture 169,424 98,908 41.6 %
A payslip 169,101 99,096 41.4 %
A commercial invoice 169,872 99,848 41.2 %
Pictures inside table cells 172,697 102,470 40.7 %
A two-page services contract 172,715 102,673 40.6 %
A gallery of picture formats 183,781 113,285 38.4 %
A minimal PDF 14,752 14,752 0.0 %
Images on a page 9,332 9,332 0.0 %
The evaluation watermark 81,420 81,420 0.0 %
All the documents 7,686,790 3,249,799 57.7 %

Three documents at the foot of the table did not move at all, and could not have: they carry no font. One draws shapes and no words, one is pictures alone, and the last is an evaluation copy, whose text is drawn line by line rather than written with a font.

See the examples themselves