Git tracks revisions then. Commits have a tree object that describes the codebase at the time of the creation of the commit. You don't need to transverse any previous commits to recreate that tree. Commits also have other commits objects listed as their parent commits, but they are not needed to examine the current state.
Git doesn't use deltas in the way that Mercurial does.
But from a more practical standpoint, tracking changesets rather than just revisions means recording ancestry ("commits also have other commits objects listed as their parent commits") such that you can freely create merge revisions and the system will Do The Right Thing with minimal hand-holding.
The internal repository format isn't really relevant here, it's about the way of thinking that the system is designed to support.
Let's put it this way: the key feature of DVCSes is that they store explicit relationships between a particular commit and its parents. This makes it trivial to reconstruct actual changesets as per my definition, which in turn makes merge machinery much more reliable.
Now, if we look at Subversion, we see that given a particular revision there's no way to reliably determine its' "parent", because the fact that r6 is followed by r7 doesn't mean that changes in r7 were based off of r6.
If we look at any DVCS, each commit would have an explicit reference to its' parent(s), so there's always a non-ambiguous way to reconstruct the entire "chain" of relationships between commits.