Classic TCP congestion control treats packet loss as the signal to slow down. On a long link — say, across an ocean — loss happens for reasons that have nothing to do with congestion, so the connection slows down for no good reason. BBR instead estimates two numbers, the bottleneck bandwidth and the round-trip time, and paces packets to match. The result is steadier throughput on exactly the links where the old algorithms struggle.
On a modern Linux kernel it is two sysctl lines: net.ipv4.tcp_congestion_control=bbr and net.core.default_qdisc=fq. No kernel rebuild, no application changes. It is the rare optimization that is both free and boring, which is the best kind.