How we tested

For every domain we query Google PageSpeed Insights (which serves CrUX). We take the 28-day 75th-percentile of real Chrome users: LCP, INP, CLS. A site "passes" Core Web Vitals only if p75 LCP ≤ 2.5s AND INP ≤ 200ms AND CLS ≤ 0.1. This is the metric Google uses for ranking; no client-side script can change it without genuinely making the page faster for real users.

GET https://www.googleapis.com/pagespeedonline/v5/runPagespeed
  ?url=https://www.<dealer>/&strategy=mobile&category=performance&key=<key>
# read loadingExperience.metrics.* (p75); fall back to originLoadingExperience

We run Lighthouse (mobile, simulated throttling) and take the median of N runs to cut variance. Lab scores are reproducible but synthetic: they reward deferring or hiding work, even when the user's real experience is unchanged. We report lab beside field precisely to show the gap.

npx lighthouse <url> --only-categories=performance \
  --throttling-method=simulate --chrome-flags="--headless=new"  # ×N, take median

To test a speed layer in isolation we load the same page twice — once with the layer's scripts allowed, once with them blocked at the network — 12 runs, median-of-3. If the layer worked, blocking it would make the page measurably worse. On the DealerInspire dealer we tested, blocking Auto Genius left LCP unchanged and TTI slightly faster: the layer relocates main-thread work, it doesn't remove it.

lighthouse <url> --block-url-patterns="*autogenius*" ...   # blocked
lighthouse <url> ...                                       # allowed

To measure our own edge rewrite honestly, we serve BOTH the original and the optimized HTML from the same localhost origin, so each pays the identical cross-origin sub-resource cost. The delta is then purely the rewrite. (Comparing an optimized localhost page against the live origin would unfairly penalize the optimized side — a trap we hit and corrected.)

control  = origin HTML + <base href> only
optimized = + critical CSS + non-blocking sheets + deferral
# both served from localhost, Lighthouse median of 5