Two flavors of a system with no history to work from
The cold-start problem is the central practical challenge in recommendation, and it comes in two distinct shapes that break the two families in opposite ways. A new user with no interaction history is invisible to collaborative filtering, there's no row of past behavior to compare against anyone else's, so similarity-based reasoning has nothing to work with. A new item with no interactions has the exact same problem from the other side: no one has clicked, watched, or bought it yet, so it has no presence in the interaction matrix collaborative filtering relies on, no matter how well-suited it would be to existing users.
Content-based filtering handles the new-item case reasonably well, since item attributes are known the moment the item exists, but it handles the new-user case just as badly as collaborative filtering does. Without any prior likes to build a taste profile from, a content-based system has nothing to match against either. So cold start isn't one problem with one fix, it's two separate failure modes, user-side and item-side, and a system needs a strategy for each.