Integrating a web of trust that doesn't have any roots at all isn't quite so easy. Unless you have access to the full trust graph, you can't build a chain between yourself and the counterparty to prove you own the identity/nym which you are asserting. So the question is how do you build a protocol that lets two parties who don't know each other find enough common connections and then how do you calculate the trustworthyness of those connections. When you have a short chain back to an organization that you axiomatically accept is competent and trustworthy it's easy (until your assumptions are violated, of course). If you don't have that, you need new protocols and bootstrapping mechanisms.
It's a problem easily solved: forget centralized certification and commoditize the actual trust itself. Like eBay feedback points. Every transaction completed to the satisfaction of both parties then credits each party with a new unit of "trust". You would make a public record of it, to establish a picture of how trust in major transaction hubs changes over time. At least then the relative trustworthyness of actors in the system is given an actual value, instead of the current situation of a qualitative notion of trust in the certificate authorities.
...and then people set up fake merchant identities and countless fake customer identities, and bump up their own trust ratings as high as they want before each scam. The best way to prevent this from happening is to elect third parties that are trusted by lots of people, to check identities before letting them into the system... oh wait, I just described the CA system...
The CA system isn't perfect, but those who succeed in breaking it are usually high profile (like Mike Hearn was saying: state-sponsored, etc), and they're not going to waste their time stealing the $42 I paid BitBrew for some coffee. But using self-signed certificates, anyone in the coffee shop with me can MitM me if they, perhaps, set up their system with the same SSID and I mistakenly connect through them when making my purchase.
Anything that is big enough to be worth stealing on a massive scale is usually done through direct merchant-customer interaction -- you don't usually buy a $17,000 car on the internet with your credit card... such large transactions usually use a second-level (or more-reliable) form of authentication beyond CAs. For "regular"-sized purchases, it makes complete sense to piggyback off of a "complete" system that is already in place, for which everyone who's ever bought anything on the internet is already capable of using.
One good thing about Bitcoin is that there is not something like a credit card number to steal. The worst someone can do is redirect a given transaction, but nothing like getting your CC number and draining your available credit, leaving you to spend hours on the phone with credit agencies trying to restore your "trust." Therefore, a passive eavesdropper who is able to decrypt the payment stream cannot directly benefit like they would with credit card transactions: with CC, they decrypt and steal databases of CC numbers and sell them on the black market (probably with BTC). But getting a database of past transactions executed in this way (via BTC) does not offer the attacker anything.
