Tan Countergambit

OK, so I was looking through opening lines for because why not, and I was analyzing the Marshall Defense on the engine. After 1. d4 d5 2. c4 Nf6 3. cxd5 I was peremptorily going to make the obvious move Nxd5, which leads to the Marshall's theoretical refutation when I noticed something. The engine was not giving White a large advantage of +1 and such, but a normal opening advantage. The three top lines didn't even include Nxd5, but instead included c6 and e6. To me, e6 looked dumb because I doubt black has good compensation, but following 3... c6. Then, following the natural move 4. dxc6 Nxc6, suddenly Black is ahead in development and attacking White's pawn. White can respond with 5. e3 or 5. Nf3, both met with 5... e5!. It leads to an endgame where White doesn't have castling rights (not a big deal) and Black is much more active (more of a deal) but is down a pawn. This gambit is known as the Tan gambit, but is very rarely played as compared to the Marshall. Why? Especially because Black can get an attacking queenless middlegame, and this is regarded by computers as far better than the natural continuation, has nobody decided to play this line? Is there some refutation I'm missing?

Even assuming the computer assessment is accurate (some slightly old engines are known be weirdly optimistic about early pawn sacrifices, only to realize their mistake later on) this simply shows that 2...Nf6 is so bad that black should already consider sacrificing a pawn for insufficient compensation, not that this gambit is in any way interesting.

As to why taking back with the knight is more common, I would assume most people playing 2...Nf6?! are low rated players that to not realize 3.cxd5 leads to trouble, not people that have made an in depth study of Nxd5 vs c6 and decided for the former.

Players who play the Marshall Defense likely don't know theory, and if White plays properly as post #2 says, 5. Nf3 e5 (trying to get a queen trade) 6. dxe5 Qxd1+ 7. Kxd1 Ng4 8. Ke1 Bc5 9. e3 Ngxe5, black is always a pawn down for basically nothing.

8...g6 is definitely an improvement, though.

So far played only in correspondence games, and Black is scoring reasonably.

In correspondence you get an opening book right? At least it's that way on lichess. Both players play most of the opening perfectly and so +0.6 is decent.