Fonts are the limit, and every converter has it
A PDF either embeds the fonts it uses or refers to one of fourteen standard faces that every reader already has: Helvetica, Times, Courier, Symbol and ZapfDingbats. Embedding an arbitrary font means parsing a TrueType or OpenType file, subsetting the glyphs, and writing the font program into the document — real work, done by a library that would have to be downloaded to your browser to do it.
This site does not ship that library, and it does not fetch anything from a third party, so the type here is one of the standard faces. Your document set in Calibri comes out in Helvetica at the same point sizes; Cambria comes out in Times. The letterforms are close enough that the document reads as itself, and the metrics are not identical, so lines break in slightly different places and a page may end one line earlier or later than Word would have ended it.
That is worth saying plainly, because "perfect formatting" is claimed by every free Word-to-PDF converter and delivered by almost none of them — the ones that do it properly are running Word or LibreOffice on a server, which means your document is uploaded to that server. This is the honest trade in the other direction: your file never leaves the device, and the type is the type a PDF reader already has.
What is actually read out of the .docx
A Word document is a ZIP. Inside it, word/document.xml holds the paragraphs, word/styles.xml holds the styles, and word/numbering.xml holds the list definitions. The package is inflated in the browser and the XML is parsed with the browser's own parser, which is why the whole conversion happens without a server.
Reading it properly means resolving styles rather than just reading runs. Word stores "this is a heading" once, as a style reference on the paragraph, and everything that makes it look like a heading — bold, 16 point, spacing above — lives in styles.xml, often inherited through several basedOn steps. A converter that ignores that produces a PDF in which every paragraph is 11 point regular, which is a surprisingly common failure. Here the chain is followed, and direct formatting on a run correctly overrides what the style said.
Lists are resolved the same way: the paragraph names a numbering definition, the definition says whether the level is a bullet or a decimal or a letter, and the counter is kept per list and per level so a numbered list arrives numbered rather than as a column of ones. The page size and margins come from the document's own section properties, so a document set up as A4 with narrow margins is converted as A4 with narrow margins.
Laying it out: real measurement, not guesswork
Every word is measured in the exact face and size it will be drawn in, and lines are broken at the point where the next word would cross the right margin. That sounds obvious and it is the step naive converters skip: they draw each paragraph as a single string, and anything longer than the column runs straight off the edge of the page and is simply not there when you open the file.
A line with mixed formatting is measured piece by piece, so a sentence with a bold name and an italic term in it wraps in the right place rather than roughly the right place. A single word wider than the column — a long URL, a reference number — is broken by character instead of being allowed to overflow. Justified paragraphs have their spaces stretched, but only on lines that were broken because they filled up, never on the last line of a paragraph, which is what stops three closing words being spread across the full width.
Pages are filled in order and a new one is started when the next line would cross the bottom margin. Explicit page breaks and "page break before" paragraph settings are honoured. Underlines are drawn as real lines under the words they belong to, at a thickness that scales with the type size. And because a document can contain characters the standard fonts have no glyph for — Greek, Cyrillic, CJK, emoji — those are counted, left out rather than replaced with something wrong, and reported to you as a number when the conversion finishes.
What this will not do, and when to use something else
Images are not carried over. Neither are headers and footers, footnotes and endnotes, text boxes, shapes, charts, embedded objects, comments, or tracked changes. Tables are walked so that their text is not lost — the cell contents appear in reading order at a small indent — but they are not rebuilt as ruled tables. Fields such as automatic page numbers and cross-references are not evaluated.
So the right way to think about this tool is: a document that is mostly words — a letter, an agreement, a report, a manuscript, an assignment, a CV without a graphic layout — converts well and converts privately. A brochure with images, a form with text boxes, or a document whose entire point is the visual arrangement will lose the parts that are not text, and you should use Word's own Save as PDF or LibreOffice's export instead.
It is also worth knowing that Word's own export is free and already on the machine, if the machine has Word. What it does not solve is the case this tool exists for: you are on a device without Word, or with a document you would rather not hand to an online converter, and you need a PDF of it now. Old .doc files from Word 97–2003 are a different, binary format and are not supported — open one in Word or LibreOffice and save it as .docx first.