In the original experiments, participants who generated a word from a cue, such as producing a rhyme or a synonym, remembered it better than participants who simply read the same word. Bertsch and colleagues later aggregated 86 studies and 445 effect sizes and found an average advantage of about 0.40 standard deviations for generated over read material, with the effect moderated by factors like test type and retention interval.
The effect is a close cousin of retrieval practice: both show that memory favors what the mind produces over what it receives. Generation applies at encoding, when you first meet the material, which makes it especially relevant to note-taking. A paraphrase in your own words is generated; a pasted quotation is not.
For a highlighting workflow, the implication is to add a generative step on top of capture. Completing a claim from memory before rereading it, titling a highlight yourself, or writing the one-sentence takeaway of a section are all small acts of generation that convert captured text into produced understanding.