Aaah, I see. So it well could be that XeTeX typesets double and tripple I's and relies on Unicode-internal decomposition of Roman numerals when searching for them. They'll then match the I's in the text. S.