Suppose your quiz scores over six weeks are 78, 84, 80, 86, 79 and 83. They bounce between 78 and 86 even though you studied the same way every week. That bounce is the noise: about 4 points either side of an average near 82. Now you try a new method and score 85. Is the method working? The 85 sits inside the band you already reach by luck, so the data cannot tell you yet.
If you had scored 95 instead, that would be 13 points above the average, more than three times the usual swing. Luck alone rarely does that, so the change is probably signal. The rule of thumb: a change well beyond the normal wobble is likely real; a change inside it is not proven, though it is not disproven either.
More data sharpens the picture. One quiz at 85 proves little, but five quizzes in a row between 85 and 88 would be hard to explain as luck, because random bounces go both ways and cancel when you average them. One contrast to keep in mind: a difference caused by who was measured, such as surveying only students who chose the new method, is not noise at all. That is sampling bias, and collecting more of the same biased data does not fix it.
Think of it like this. Listening for a friend in a noisy cafeteria works the same way. The chatter is noise, always there and never saying anything. A whisper is lost in it; a shout cuts through. The quieter the room, or the more times your friend repeats the message, the easier it is to hear. A small effect in data likewise needs low noise or many measurements to show up.
Core idea. Signal vs noise is the judgment of separating a meaningful pattern from random fluctuation in the same data.