Quotes have ~always been an exact match on the tokenized query text, not a substring match on the corpus text. No synonyms, reordering, gaps, etc, but the matches -- and failures -- are sometimes not obvious at first blush.
If you search for "don't stop me now", for instance, that "don't" tokenizes to "don t", so it will match the tokenized strings "don't", "don t", "don-t", "don, t", etc ... but not "dont", because that's outside tokenization.
On the other hand, snippets mostly are substring matches of the query text, so if you see a result to a literal query that doesn't have a snippet, you know it's probably one of the weird matches.
If you use quotes around a phrase, it will reorder terms and make substitutions with synonyms in addition to straight up ignoring the quoted phrase no katter how many times you add +. If you then fiddle with settings (randomly not available depending on star alignment and device) to change it to 'verbatim' it will still reorder and split up tokens in the phrase.
term -term
For example:
steve -steve returns 0 results
test -test returns 4,780,000,000 in my search
starting with google/youtube videos"Exact tokenized matching" can look like "exact string matching" a lot of the time. Until you hit some of the edge cases it's like kerning: https://xkcd.com/1015/