Google's plan to make Core Web Vitals an SEO factor, with the deadline set for May 2021, has teams scrambling to get their metrics in order. The acronyms in play are CLS, LCP, and FID, and tooling for measuring and diagnosing problems keeps appearing. Fixing what the tooling finds is a different matter. Poor CLS, for instance, usually traces back to content loaded after the fact — ads being the usual suspect — that pushes the layout around. The honest remedies are to stop injecting such content or reserve its space in advance so the page moves less.
When the hero image becomes the problem
LCP is a harder case. Say a large hero image is slow to paint and drags the LCP score down. Chris Castillo's approach is to defer the hero background image entirely until the user interacts with the page.
Castillo tested the idea lightly and found that some visitors did not notice the delay. He is upfront about the tradeoff: the background image does not load until an interaction occurs, which means a fallback has to stand in until it does. He asked a handful of friends to open the page on their phones and report anything odd. None flagged a problem. His observation was that every friend who tested it either already had a finger on the screen or touched it right away while the page was loading, so the image arrived too quickly to notice.
Although this accomplishes the goal, it’s not without a cost. The background image will not load until the user interacts with the screen, so something needs to be used as a fallback until the image can be loaded. I asked a few friends to load the page on their phones and tell me if they found anything strange about the page, and none of them noticed anything “off”. What I observed is that the few friends I asked to test this all had their fingers on the screen or quickly touched the screen when the page was loading, so it happened so quickly they didn’t notice.
Measuring users versus pleasing a machine
The technique itself is documented reasonably well, but its purpose is to get a machine to hand out a better test score. That looks like the opening of an odd phase in web performance work: the metrics have moved toward user-centric measurement, yet people are building tricks to game those numbers with methods that, if anything, slightly degrade the user experience.




