I did not know this until today... vision models don't hunt for the white space between letters like old OCR does. They just chop the whole image into a fixed grid of equal squares (patches) and treat each square like a word in a sentence.
I found it interesting, that Git itself has built in similarity notion... when it packs objects, it groups files by path+size, runs delta cmpression to find which are close.
I am with you here.. there is actually name for why companies do this. They are not pushing app because it is better, but because browser tab cannot lock you in.
This will help if you want to dig deeper - https://vectree.io/c/how-ocr-works-traditional-pipelines-vs-...