Three seconds to 200 milliseconds, and almost none of that came from making the code faster.
The endpoint was waiting on other people's systems. One page assembled itself from live calls out: carrier rate lookups per shipping option, a provider check, a lookup after that. Each one sequential, each one fast enough on its own, and the tails added up. Your p99 becomes the worst p99 among your vendors, and no amount of query tuning touches that.
Two changes did the work.
First, stop serializing calls that do not depend on each other. Five sequential 300ms calls are a second and a half. The same five in parallel are 300ms plus noise.
Second, and this is the one that actually matters: decide a freshness window per field, not per request. Not everything on a page has to be true as of this millisecond. A shipping rate can be minutes old. A catalog attribute can be hours old. Price and stock usually cannot. Once you write that down explicitly, most of the fan-out leaves the request path and becomes a background refresh.
The latency number is the visible part. The real win is that the endpoint stops failing when a vendor has a bad afternoon.
I think your API has at least one field you would admit does not need to be live...