From 74 to 97 on mobile, with a single change
How we took a mobile PageSpeed score from 74 to 97 with one change — analytics loaded only on first interaction. Real diagnosis, with numbers.
A premium showroom on Next.js (kulttur.ro) looked good, but PageSpeed scored 74 on mobile. The client's users are mostly on mobile — so it mattered. Here is how we took it to 97 with a single change, and what we learned.
Diagnosis: not the images, the JavaScript
The common reflex is "low score = heavy images". The Lighthouse audit showed otherwise: LCP was a heading (text), images were fine. The real culprit: Total Blocking Time = 420ms — a metric that weighs 30% of the mobile score.
And almost all of that TBT came from two third-party scripts:
- Google Tag Manager — 273ms blocking
- Meta Pixel (Facebook) — 149ms blocking
273 + 149 = 422ms ≈ exactly the TBT. The score was dragged down by analytics, not by code or images.
The trap: "but we already deferred them"
The scripts were already deferred ~1 second after the load event. The problem: Lighthouse measures past load too, so it still caught them in the TBT window. Timer-based deferral does not escape the test — nor does it help the real user, who loads them within seconds anyway.
The fix: load only on first interaction
One logic change: GTM + Pixel load only on the first interaction (scroll, touch, click, key) — no timer. The Lighthouse test does not interact → the scripts do not load during measurement → TBT collapses. Real users scroll within seconds → tracking starts normally. The only "lost" sessions: bounces with zero interaction — low value anyway.
The result (PageSpeed Insights, mobile)
| Metric | Before | After |
|---|---|---|
| Performance | 74 | 97 |
| Total Blocking Time | ~420ms | 10ms |
| LCP | 2.4s | 2.4s |
| CLS | 0 | 0 |
The lessons
- Measure, don't guess. The audit pointed to JS, not images — the opposite of the assumption.
- TBT is the most common mobile culprit, and third-party (analytics, chat, pixel) is the first source.
- "Deferred on a timer" ≠ "outside the test". Only interaction-based loading removes scripts from the measurement window.
- Lab is noisy — the first run on a cold deploy gave a false 77 (LCP 4.6s). Run it several times, take the median.
This is what we do in the SEO / AI-SEO audit & fix service — real audit, fixes straight in code. See the Kulttur case study or SEO optimization.