A tree of prefixes: lookup time depends on the key length, not on how many keys you store.
A trie stores keys in a tree where each edge is one character and each node represents a prefix. The root is the empty prefix, and a node is marked terminal exactly when it completes a stored key.
Because the structure is the set of prefixes, searching is a walk down one path: one edge per character of the key, and no comparisons against any other key. Lookup cost is O(m) for a key of length m, independent of n.
Trie search cost does not grow with the number of stored keys, so a trie suits very large or unbounded key sets: dictionaries, autocomplete, and IP routing tables all use them.
The cost is space. A trie stores one node per distinct prefix, so a dictionary of n words of length m can need O(nm) nodes, while a hash table needs O(n). Tries are also awkward with keys that are not sequences — arbitrary numbers or objects need an encoding step first.
Storing car, cat and cart creates one shared path for c-a, then a fork. The branch point is a real cost and a real benefit: the set of all stored keys sharing a prefix is available for free, which is what makes prefix queries and autocomplete cheap.
It also makes a useful distinction visible. Searching ca in a set containing car, cat and cart walks the whole path successfully but finds no terminal node, so the answer is correctly not a member — a prefix is not a key.