AI summary by TwinMind

Hacking LightHouse Scores

Theo - t3․gg

44m 2s

Lighthouse Score Importance and Limitations

  • Product managers push for high Lighthouse scores, marketing/sales teams claim importance for web success
  • Developers use scores to evaluate frameworks, often used to shame Angular apps
  • Key question: Do these scores actually mean anything? Can they be hacked?
  • Sponsored content by Sentry, article by Salma on "How to hack your Google Lighthouse scores in 2024"

Field Data vs Lab Data Critical Distinction

  • Local performance testing on high-spec dev machines ($4,000 MacBook) not realistic for user experience
  • Lab data provides directional trends but doesn't paint full picture of reality
  • Field data from real users on various devices, connection speeds, browsers more accurate
  • Application performance monitoring tools like Sentry provide real-world user data
  • Philip Walton 2021 report: Almost 50% of pages scoring 100 on Lighthouse didn't meet core web vital thresholds based on HTTP archive data

Web Performance Beyond Speed Metrics

  • Google 2018 report: Bounce probability increases 32% if page load time exceeds 3 seconds, 123% at 10 seconds
  • Page load not single moment in time - multiple moments affect user perception
  • Speed influenced by business KPIs and sales targets
  • Physical limitations: humans and servers spread globally, infrastructure can only deliver limited bytes
  • Example: Ping.gg gets "super fast" user feedback but poor Lighthouse scores due to homepage video player
    • Mobile score: 44 performance
    • Desktop score: 95 performance
    • Site intentionally blocks mobile users and Firefox users (WebRTC bugs)

Lighthouse Performance Score Calculation

  • Weighted combination of core web vitals and speed metrics:
    • Total Blocking Time (TBT): 30%
    • Cumulative Layout Shift (CLS): 25%
    • Largest Contentful Paint (LCP): 25%
    • First Contentful Paint (FCP): 10%
    • Speed Index (SI): 10%

Total Blocking Time (TBT) Analysis

  • Measures time after FCP when main thread blocked preventing user input responses
  • Main thread considered blocked when JS task runs >50ms
  • Ensures page responds to physical user input (key presses, mouse clicks, touch, typing)
  • Frameworks like Qwik built specifically to solve this problem

Cumulative Layout Shift (CLS) Details

  • Measures unexpected visual shifts after content shown to user
  • React docs example: Logo switching based on localStorage causes complex implementation to prevent layout shift
  • Required blocking main render, checking localStorage, swapping before page display
  • Good CLS score: <0.1, poor score: >0.25

Largest Contentful Paint (LCP) Specifications

  • Marks when page's main content likely loaded and useful
  • Good LCP score: <2.5 seconds (higher threshold than expected)
  • Elements considered for LCP:
    • Image elements
    • Image elements inside SVG
    • Video elements (major issue for Ping.gg)
    • Elements with background images (URL function, not CSS gradients)
    • Block-level elements with text nodes
  • Excluded elements: zero opacity, full viewport coverage, placeholder/low entropy images

First Contentful Paint (FCP) and Speed Index (SI)

  • FCP: First point where users see something on screen
  • SI: Measures how quickly content visually displays during page load
  • Good FCP score: <1.8 seconds
  • Good SI score: <3.4 seconds
  • Demonstrates importance of loading states vs blank pages during server-side generation
  • Example: Next.js loading.tsx files provide immediate feedback vs waiting for server response

Lighthouse Scoring Methodology

  • Uses log-normal distribution derived from real website performance data (HTTP archive)
  • Site scored against real web performance data, not in isolation
  • Non-linear relationship: substantial improvements to low scores easier than improving high scores
  • Usability favored over raw speed in scoring algorithm

LCP Hacking Techniques

  • Show minimal LCP-qualifying content on load, delay real content until after Lighthouse test
  • Demo: Page with single H1 element, real content loads after 10 seconds
  • Lighthouse awards perfect 100 score even with single punctuation mark as LCP element
  • Approach works with arbitrary splash screens containing minimal viable LCP elements
  • Not recommended: terrible for SEO (robots can't discover main content), poor user experience
  • Rent Path 2021 case study: Improved Lighthouse score 17 points by increasing thumbnail image sizes
    • Made thumbnails larger than Google map tiles to control LCP element selection
    • Literally making page slower improved LCP score by controlling largest element

CLS Hacking Methods

  • Defer layout shifts until after page load event completion
  • Demo page: 0.143 CLS score despite immediate JS adding text elements
  • 5-second setTimeout delay results in perfect 0 CLS score with confetti celebration
  • CSS animations using transform don't trigger CLS (unlike DOM manipulation)
  • Hack won't work for field data or long-term user experience
  • Time span tests in Lighthouse correctly report non-green CLS (~0.16)

Speed Index - Unhackable Metric

  • Based on visual progress during page load timeline
  • No way to fake loading content faster than actual speed
  • Only optimization: load as much content as early as possible
  • Recommendations for improvement:
    • Deliver static HTML from CDN
    • Avoid/minimize images
    • Optimize image formats (Next.js image component example)
      • Serves WebP/AVIF based on browser capability vs original PNG/JPEG uploads
    • Minimize/eliminate CSS
    • Prevent JavaScript/external dependencies loading during test window

Total Blocking Time (TBT) Hacking

  • Measures main thread blocking after FCP preventing user input responses
  • Good TBT score: <200ms
  • JS-heavy SPAs prone to poor TBT scores
  • Example: Twitch.tv terrible scores due to client-side page generation
    • Mobile: Very poor performance
    • Desktop: Still poor due to extensive client-side processing
  • Possible hack: Defer all JavaScript until after Lighthouse test completion
  • Modern frameworks/hosting providers reducing TBT issues through better server-side processing

Lighthouse Limitations Demonstration

  • 2019 Manuel experiment: Terrible page scoring well in Lighthouse still works in 2024
  • Demo page: Input events disabled by CSS/JS for 5 seconds
  • Page technically unresponsive but scores 100 for performance and accessibility
  • Lighthouse can't detect every usability/accessibility issue
  • Cannot substitute for actual usability testing and common sense

Additional Hacking Strategies

  • Use fastest, highest-spec hardware for testing
  • Ensure fastest internet connection possible
  • Never use field data, only lab data with optimal conditions
  • Rerun tests with different conditions until desired results achieved
  • Potentially serve different page versions to testing IP addresses

Performance Monitoring Recommendations

  • Think of Lighthouse as "canary in coal mine"
  • Use application monitoring tools like Sentry for real production data
  • Focus on actual user experience over artificial scores
  • Field data more important than lab scores for business outcomes

TwinMind

Get summaries like this for your own meetings

TwinMind takes notes in the background during meetings, lectures, and calls, then writes the summary for you.

Free plan to get started

No bot joins your calls

Phone, desktop & Chrome apps

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome