The null hypothesis is the starting position a test assumes until the data argues otherwise.
It is usually the dull explanation: no difference, no effect, no relationship.
Making it specific is what lets a test work out how surprising the data is.
The null is written as an exact statement, such as that the 2 means are equal.
The alternative can be 2-sided, meaning a difference in either direction, or 1-sided, meaning larger only.
Choosing a 1-sided test after seeing which way the data points doubles the real false alarm rate.
The result is either to reject the null or to fail to reject it, never to accept it.
Failing to reject means the data is consistent with the null, which is weaker than showing the null is true.
The null does not have to be zero; it can be a value from past results, such as a 30 minute average.
A badly chosen null makes a test pointless, such as testing whether a conversion rate is exactly 0 when it plainly is not.
Every statistical test has a null built in, so knowing it tells you what a significant result actually means.
For a test of whether 2 categorical variables are related, the null is that they are not.
In practice, writing the null and the alternative in plain words before collecting data prevents confusion later.
Teams that skip this step often argue afterwards about what the test was meant to show.
The common misconception is that failing to reject the null proves it, when absence of evidence is not evidence of absence.