Understanding the Inconsistencies in PageSpeed Insights Performance Metrics
Achieving consistently high performance scores for your website is a common goal for developers and site owners alike. Recently, I dedicated several weeks to optimizing my custom frontend, with the aim of reaching a perfect 4/4 (100/100) score on both desktop and mobile in Google’s PageSpeed Insights (PSI). While desktop performance has been solid, and I’m nearing the same on mobile, I’ve encountered perplexing discrepancies: PSI results fluctuate dramatically over short periods, sometimes dropping as low as 79/100—even with the exact same page source and testing conditions.
The Journey Towards Optimization
My efforts involved a meticulous approach:
- Code Optimization: Inline critical CSS, compress HTTP responses, and prioritize essential assets.
- Resource Management: Preload key fonts and hero images, alongside a layered JavaScript preload strategy—loading minimal external scripts before “window.load,” then sequentially loading consent scripts, tracking, and UI dependencies based on user interaction.
- Script Handling: I suspect that third-party scripts—particularly CookieYes for cookies and Google Analytics for tracking—may be impacting performance metrics, although the exact connection remains elusive.
Through these steps, I’ve consistently achieved perfect scores in Lighthouse (with network emulation throttled to Slow 4G and CPU to low-tier), but PageSpeed Insights results remain unreliable and inconsistent.
Diagnosing the Variability
Despite most indicators pointing to optimized loading, PSI scores vary significantly within minutes. For example, one test yields a “Performance” score of 79/100, while another, just five minutes later with the same source code and cache state, reports 99/100. This inconsistency is baffling, especially since:
- The “window.load” event fires within 0.5 seconds in all cases.
- According to performance logs (using
performance.now()), the First Contentful Paint (FCP) and Largest Contentful Paint (LCP) occur rapidly—FCP ~2.3s, LCP ~4.2s—well before the measured PSI window.
Furthermore, PSI’s “Best Practices” logs indicate no significant errors that could explain these fluctuations.
Key Observations
- The website is fully cached via LiteSpeed Cache, with cache cleared before testing.
- Minor differences in results are not due to changes in source code or server response.
- The disparity hints at external factors influencing how PSI measures performance
