The link you included shows that protobufs are at least 15% better for all users, and as much as 57% better for cases where the data is small. Doesn't that mean for 100% of users it will actually make a difference?
Your users might not care about the difference but it will be there.
Actually realizing that speed up for your users will take time away from delivering features.
Engineering is a trade off, always will be.
You don't optimize things for the cases when they are fast. (Unless the gain is a couple of orders of magnitude; certainly not for a 50% speedup.)
The 15% gain is the one that matters. On practice, it comes at the expense of a more complex (thus larger, negating some of it) and less reliable system. It is very rare that this trade-off is worth it.