Which text that a browser does not show at first the output holds.

The choice answers one question: can a reader get to the text? It does not depend on how the page hides it.

aria-hidden is not hidden for this option. It takes an element away from assistive technology only: a browser still shows the element, so a reader sees its text, and every choice keeps it. One rule is older than this option and does not change with it: for an inline <svg aria-hidden="true"> no choice writes the title, the description or the label of the graphic.

Every choice also keeps what the converter never treated as hidden: inert, a closed <details> or <dialog>, <datalist> and <option> text, and an element hidden by a class name, a style sheet rule, opacity, content-visibility, a zero size or an off-screen position. The converter reads the inline style attribute only. It does no layout and reads no style sheet, so it cannot tell that a rule in a style sheet hides an element.

A kept element converts like the same element without the attribute or style. Kept <template> and <noscript> content converts where it is written, as if the two tags were not there: a row in a <template> inside a <table> is a row of that table, and the content of a declarative shadow root comes before the other children of its host element. Slots are not resolved. All keeps <noscript> content with every preprocessing preset. A <template> or <noscript> in the document head holds metadata (<link>, <meta>, <style>), not text for a reader, and no choice keeps it. A document with no <head> tag has a head too: it starts at the doctype, at the <html> tag or at the first metadata element, and it ends where the body starts.