OCLC Connexion's metadata extractor tool has some exciting possibilities - even though I'd need to heavily edit any MARC records it spits out, what it spits out is usually better than starting from scratch. I decided to try it out on a few things today. I still haven't really done much with it, but I've at least got a bit of a feel for how it works.
I've only tried it on one website so far (http://www.sluggy.com), with so-so results. I decided to try it out on a few of DSL's databases. In most cases, the extractor spits out similar results whether I use our redirect URLs or copy and paste the URL from the Address bar. I tried this out with JSTOR, Academic Search Complete, Project Muse, and a few others and got pretty much the same results with either URL. One exception was Global Road Warrior. The record created when I used the redirect URL was just awful - no title in the 245 field, and the 546 field (Language Note) says this is in Ukranian. If I copy and paste from my browser's Address bar, on the other hand, the record is much nicer. The 245 and 546 fields are much more accurate, and the rest of the record is better. That's not to say the record wouldn't need to be edited - it would still need lots of editing to truly be useful - but it's better than before.
Of course, if I were going to catalog the databases DSL subscribes to, I would first do a search for them in OCLC. While OCLC would probably not include records for websites, it does include records for databases. For instance, there's a record for JSTOR (OCLC #36027621) that, at a glance, looks good enough to consider using. This record definitely looks better and less likely to require major editing than the record the metadata extractor spits out.