[0]: https://dropbox.tech/security/zxcvbn-realistic-password-stre...
That XKCD comic came out August 2011, and this article was released April 2012, eight months later. At the time, it made sense to use that as an example of a "highly entropic password" in a blogpost targeting a general technical audience.
It has now been 156.5 months since that comic was released, and 164.5 since zxcvbn was originally trained on its 2011 dataset. "Correct-horse-battery-staple" and its variations are widely used strings and is no longer an example of a strong password.
It's worth being careful about our definition of entropy. The same highly-entropic source generates "hunter2" just as often as it generates "aaaaaaa" just as often as it generates "jpnj6i3".
A "good password" is one that is unlikely to be guessed by an offline cracker. But there are countless strategies an attacker might choose, most of which involve taking a list of passwords and applying rules to them. This is why a highly-entropic source is important (e.g. of all the 32-character passwords, 'weak' ones like "aaaaa..." make a negligible percentage of them) as well as uniqueness ("mP7t6e8TAH..." is a weak password the moment it's leaked in a breach.)
You can't know what strategy the attacker will pick beforehand-- the idea is that it's negligibly likely that an attacker chooses a strategy that cracks your password in a few guesses.
Zxcvbn intentionally takes compromises to be a performant best-effort estimation of how many guesses an attacker would take for your password. (It would improve its estimation of guesses by actually trying to guess your password with `hashcat` and the antipublic combolist, or whatever people are using nowadays, but then it would take eons to provide an estimation rather than milliseconds.)
Zxcvbn is imperfect by design. It's a tradeoff it makes for being fast and small.