Drop a thumbtack 50 times. It lands point-up 27 times and on its side 23 times, so your experimental probability of point-up is 27/50 = 0.54. No reasoning about its shape could have produced that number, because a thumbtack is not symmetric and its two ways of landing need not be equally likely. Data is the only route.
Drop it 50 more times and you might see 31 point-up, so the estimate becomes 58 out of 100, or 0.58. Neither number is wrong; both are estimates, and estimates from small batches swing more. With 10 drops, 7 point-up (0.7) would surprise no one; with 1,000 drops, the fraction usually moves only in the second decimal place. That is why an honest estimate always says how many trials it came from.
Rolling a fair die 60 times and getting 12 sixes gives an experimental probability of 12/60 = 0.2, while the theoretical probability is 1/6, about 0.167. These are different quantities: one comes from the die's symmetry, the other from these particular rolls, and a gap of this size is normal in 60 rolls, not proof of a loaded die. When a question gives counts of what happened, compute the experimental probability; when it gives a fair object and no data, reason it out theoretically.
Think of it like this. Judging a restaurant by its reviews instead of its menu is the experimental approach: you go by what actually happened to the people who ate there. Three reviews can mislead you; three hundred give a fair picture. The limit: reviews describe past meals, and if the kitchen changes, as when a coin gets bent, the old data stops applying.
Core idea. Experimental probability estimates a chance from what actually happened, successes divided by trials; it works for any process, symmetric or not, but it is only an estimate, and it becomes more reliable as the number of trials grows.