"Perceived load time" is not a panacea, and not a substitute for actual metrics. Do you actually go interview users on what they perceive? If one doesn't, and is instead just looking at the code _they_ wrote and thinking "yeah my code feels fast", that's biased, unscientific, and objectively an inferior quality assessment strategy than just looking at TTFB/TTI/etc and deriving conclusions from those hard numbers.

The only people I see obsess over these metrics have either their ego or pay tied to them.

> Do you actually go interview users on what they perceive?

Yes, regularly, because that's what effective teams do. I also limit bandwidth to 3G and record the sessions so that everyone can share in the learnings, which very rarely circle back to things like TTFB/TTI.

Interesting! How do you iterate / keep track of regressions?

Speedcurve.com - download the data and then use basic outlier detection on the time series as well as individual observations looking for anything out of the ordinary. Most of the time it was marketing uploading a 20MB PNG to the CMS, other times it's some 3rd party JavaScript (looking at your Criteo) doing something questionable.