PGO operates at the cpu branch level. The main benefit, if i understand correctly, is to reduce branch mispredictions which lead to pipeline stalls.

Theres absolutely no reason to think that the specific workload of a system generally makes random/arbitrary/heuristic be the best way to pick a most likely branch for a given branching instruction.

For example, when scanning data you might check if the current value is equal to a test value, and branch on that to either code that handles the match, or else code that moves to the next entry to test that. This may be extremely likely to match (95%) as in a hash table, or very likely to not match (string search). A compiler can't tell which is true, and your workload literally has no impact on this unless it is pathological.

I don't follow. If I have an ingestion-heavy workload then my hash-map will certainly be stressed in a completely different way than if I have a read-only workload. The data that PGO collects during the former will be different than the data that it collects during the later, so, how is it not that the workload doesn't impact the PGO optimizations? Perhaps I misunderstood you.