Your friend says she can tell two brands of cola apart by taste. You give her 10 blind cups and she names 9 correctly. If she were only guessing, each cup would be a coin flip, and the chance of 9 or 10 correct is 102410+1=102411≈0.011, about 1 time in 93. That number is the p-value. Results this good are rare for a guesser, which is evidence she really can taste the difference.
Why "at least as extreme" and not "exactly 9"? When there are many possible outcomes, every exact one is unlikely: even exactly 50 heads in 100 fair flips has probability only about 0.08. So the p-value adds up every outcome as far as or farther than yours from what the null predicts. If the alternative allows either direction, as in "the coin is unfair", both tails count; if it names one direction, as in "she does better than guessing", only that tail does.
The p-value is P(data this extreme∣H0), which is not the same as P(H0∣data). Reversing a condition changes the probability: nearly every professional basketball player is tall, but very few tall people are professional basketball players. A p-value of 0.011 does not mean a 1.1% chance she is guessing; that would also depend on how common real taste-testers are. Nor does a tiny p-value mean a big effect: with huge samples, trivial differences produce tiny p-values.
Think of it like this. A p-value is the referee's reaction to a suspiciously lucky run of dice: "if these dice are fair, a run this good or better turns up about 1 game in 500." It rates how odd the data are under the innocent story. It does not say how likely the innocent story is; that also depends on how many players actually cheat.
Core idea. Assume the null is true, then ask how often chance alone would give data at least as extreme as yours. That tail probability measures surprise: small means hard to explain by chance. It does not directly tell you how likely the null is or how big the effect is.