Although it's a complete pain to standardize the results, you can have benchmarks as part of your CI (It's a bit easier if you can use some kind of normalised measurement like cache misses per loop iteration, but rdpmc generally isn't allowed - or even available from the OS).

Latency from the user perspective is a bit harder, although I guess for a webpage you could run it headless into a video recording then time how long to go from a blank screen to showing information?