It ultimately depends on your threat model. But assuming a state actor has access to NetFlow data, an attack could work like this:

* State actor determines that an IP belonging to a VPN company had a session on example.com around t1-t2

* You -> VPN server at t1

* VPN server -> example.com at t1+latency

* More traces from both sides until around t2 as you browse the site

By correlating multiple samples, and accounting for latency between you and the VPN server and delay introduced by the VPN itself, they would be able to get decent confidence that it was you.