With ML methods, they do benefit from taking data from millions to billions.
You can see that yourself with Google by searching for something obscure (eg. "how to unblock a drain jammed with cat fur") and getting three friends to click the 2nd search result. Come back in a week, and suddenly the 2nd result is now the first result! (ignore the videos and answer box)
It turns out just three new data points is enough to influence ranking for that query. It even affected related queries if you change the word "jammed" for "gunked".
You can see that yourself with Google by searching for something obscure (eg. "how to unblock a drain jammed with cat fur") and getting three friends to click the 2nd search result. Come back in a week, and suddenly the 2nd result is now the first result! (ignore the videos and answer box)
It turns out just three new data points is enough to influence ranking for that query. It even affected related queries if you change the word "jammed" for "gunked".