Lab · 7
Can AI achieve a perfect Google PageSpeed Insights score?
Every experiment so far has been about what AI can build. This one turned Google's own performance grader on the site itself — can AI find every real issue, fix it, and get a genuinely re-measured score all the way to perfect, on both mobile and desktop?
Hypothesis
AI can achieve a perfect Google PageSpeed Insights score on mobile and desktop.
Finding
Not perfect, but close to perfect.
Tested by fixing and deploying to the real, live production site, then running one fresh (not cached) Lighthouse analysis per device — not a synthetic benchmark, not averaged across repeated runs.
Method
Audit
Ran a genuine PageSpeed Insights analysis against the live site and read every flagged issue.
Fix
Built five targeted fixes tied to specific flagged problems — no speculative changes.
Deploy
Pushed all five fixes live and verified each one directly in the live page's DOM.
Re-measure
Forced a fresh, non-cached Lighthouse run for both mobile and desktop and recorded the real scores.
Tech
The experiment
Four real, verified problems, five shipped fixes, and two more fixes that were investigated and deliberately turned down.
AI handled
- Reading the audit and separating real, fixable problems from diagnostics that don’t move the score
- Building and deploying five fixes: async font loading, a WebP headshot with a PNG fallback, an
inertfix on the nav panel, a footer-contrast increase, and dead-CSS removal - Investigating two further levers after the first re-measurement and reporting their honest tradeoffs instead of just doing them
- Running the fresh, non-cached re-measurement and reporting the real numbers, wins and all
Dr. Paul handled
- Setting the hypothesis
- Giving explicit go-ahead to push the full fix batch live
- Reviewing the real measured results and pushing back accurately (“It’s not perfect”) rather than accepting a partial win as full success
- Declining both further optimization levers once he’d heard the tradeoffs, closing the technical work as final
Worked: The nav-panel inert fix closed a real WCAG violation outright (mobile Accessibility 97→100), and the combined font/image/CSS work lifted mobile Performance from 78 to 89. Best Practices and SEO were already, and remain, a literal 100/100 on both devices.
Constraint: Mobile's remaining gap traces to GitHub Pages serving from a single origin with no CDN and a fixed 10-minute cache lifetime that can't be overridden from the repo — a hosting-platform ceiling, not anything left unfixed in the page's own code.
Finding
Not perfect, but close to perfect.
Accessibility, Best Practices, and SEO all reached a literal 100 on both mobile and desktop.
Performance landed at 89 on mobile and 99 on desktop — real, measured, and short of perfect on both counts.
Implication
This experiment split into two different kinds of problems. Every issue that was actually inside the code — a blocking font request, an oversized image, a real accessibility violation, a contrast failure — got found and fixed, and that alone moved the mobile score from 78 to 89 while three of four scoring categories reached a literal 100 on both devices. What's left isn't a code problem anymore: it's the hosting platform underneath, which has no CDN in front of it and a cache lifetime fixed at ten minutes with no setting in the repository that changes it. That's the distinction worth carrying forward — asking AI to fix what's broken in the code, and asking it to fix what the infrastructure underneath limits, are two different requests with two very different costs. The first is close to free. The second means changing where the site lives.
AI closed every gap that was actually in the code. The gap left over belongs to the hosting bill, not the build.
Next Question · 8
Should Dr. Paul become an avatar?
Talking avatar, text chatbot, or shelved entirely?
Back to the Lab