I'll start by saying that I'm a massive GWT fan, but yes there are many pitfalls, but most if not all we were able to overcome:
Problem: Long compile times, as your project grows so does the amount of time it takes to compile it. I've heard of reports of 20 minute compiles, but mine are on average about 1 minute.
Solution: Split your code into seperate modules, and tell ant to only build it when it's changed. Also while developing, you can massively speed up compile times by only building for one browser. You can do this by putting this into your .gwt.xml file:
<set-property name="user.agent" value="gecko1_8" />
Where gecko1_8 is Firefox 2+, ie6 is IE, etc.
Problem: Hosted mode is very slow (on OS X at least) and does not come close to matching the 'live' changes you get when you edit things like JSPs or Rails pages and hit refresh in your browser.
Solution: You can give the hosted mode more memory (I generally got for 512M) but it's still slow, I've found once you get good enough with GWT you stop using this. You make a large chunk of changes, then compile for just one browser (generally 20s worth of compile) and then just hit refresh in your browser.
Update: With GWT 2.0+ this is no longer an issue, because you use the new 'Development Mode'. It basically means you can run code directly in your browser of choice, so no loss of speed, plus you can firebug/inspect it, etc.
http://code.google.com/p/google-web-toolkit/wiki/UsingOOPHM
Problem: GWT code is java, and has a different mentality to laying out a HTML page, which makes taking a HTML design and turning it into GWT harder
Solution: Again you get used to this, but unfortunately converting a HTML design to a GWT design is always going to be slower than doing something like converting a HTML design to a JSP page.
Problem: GWT takes a bit of getting your head around, and is not yet mainstream. Meaning that most developers that join your team or maintain your code will have to learn it from scratch
Solution: It remains to be seen if GWT will take off, but if you're a company in control of who you hire, then you can always choose people that either know GWT or want to learn it.
Problem: GWT is a sledgehammer compared to something like jquery or just plain javascript. It takes a lot more setup to get it happening than just including a JS file.
Solution: Use libraries like jquery for smaller, simple tasks that are suited to those. Use GWT when you want to build something truly complex in AJAX, or where you need to pass your data back and forth via the RPC mechanism.
Problem: Sometimes in order to populate your GWT page, you need to make a server call when the page first loads. It can be annoying for the user to sit there and watch a loading symbol while you fetch the data you need.
Solution: In the case of a JSP page, your page was already rendered by the server before becoming HTML, so you can actually make all your GWT calls then, and pre-load them onto the page, for an instant load. See here for details:
Speed up Page Loading by pre-serializing your GWT calls
I've never had any problems CSS styling my widgets, out of the box, custom or otherwise, so I don't know what you mean by that being a pitfall?
As for performance, I've always found that once compiled GWT code is fast, and AJAX calls are nearly always smaller than doing a whole page refresh, but that's not really unique to GWT, though the native RPC packets that you get if you use a JAVA back end are pretty compact.