Well managed routing nodes and sufficient inbound liquidity on the payee node. Paying large nodes in the network has high probability of succeeding, that's not so much the case for smaller nodes which are sometimes less maintained: not enough channels, not enough inbound.
Also, there's more work to be done when in comes to routing algorithms.

It's probably the error messaging from LN itself being very vague. A large part of this is due to the desire to preserve privacy, which is great, but it makes it very hard for us to dig into most of our payment failures and do something about it.

The Lightning report we recently released has some great data and graphs on our failure reasons and rates: https://blog.river.com/the-lightning-network-in-2023/