On one hand, I mostly agree. On the other, if the sample is correctly created (aka: both “really random” and “really representative”) then it should be enough. The additional problem is that polls are known to be poorly representative, because a lot of people just troll their way through them, giving bullshit answers that are undetectable and pollute the end results. Finally, truly random and truly representative and really hard to achieve, so that’s an additional source of errors.