A reorganisation is the moment a chain discards blocks it had already accepted and adopts a different branch. It is a normal event at shallow depth and an emergency at depth.
Two valid branches can exist at once
Block producers work in parallel and cannot see each other instantly. If two find a valid block at nearly the same moment, part of the network sees one first and part sees the other.
Both branches are legitimate under the rules. Nodes simply build on whichever they received first, and the network is briefly split without anyone doing anything wrong.
The split resolves when the next block extends one branch, because nodes follow the branch with the greatest accumulated work or the strongest set of votes.
The discarded branch does not vanish quietly
Blocks on the losing side are orphaned. The work that produced them earns nothing, and any reward they contained is never spendable.
Transactions inside those blocks return to the pending pool if they are not already in the winning branch, so they are not lost but they are no longer confirmed.
Anything that depended on their ordering is recomputed against the new sequence, which can change outcomes in applications where position within a block matters.
Depth is what turns a nuisance into an attack
A one-block reorganisation happens routinely and costs nothing beyond a brief wait. A reorganisation several blocks deep requires outproducing the entire network over that span.
An attacker who can do this can spend coins, wait for a service to release goods or credit, then publish a branch in which the payment never happened.
The cost of that effort is what confirmation requirements are priced against, which is why exchanges wait longer for larger deposits on the same chain.
Software bugs cause them as well
If part of the network upgrades to rules the rest does not share, the two groups accept different blocks and follow separate branches for as long as the disagreement lasts.
Reconciling that requires operators to converge on one version, and the branch abandoned in the process can be far deeper than any adversary could have produced.
Why designs increasingly avoid the possibility
Consensus mechanisms that finalise blocks by explicit vote remove reorganisations beyond a checkpoint, because reversing a finalised block means validators contradicting signatures they have already published.
The protocol can detect those contradictions and destroy the stake behind them, which converts an open-ended probability into a defined and very large cost.