A tenth of a percent is enough
A multiplexed TCP stream treats a lost packet as a pause for every request that shares the connection. At 10% loss the congestion window collapses and the page feels frozen. At 0.1% the graphs still look clean and a single large transfer already waits on retransmission.
I stopped using ICMP loss as a proxy for that wait. Retransmits on the long-lived flow, and the gap between application writes, are the numbers that match what a person notices. Raising the send buffer does not put the dropped packet back on the path.