From 79 to 90: The Taylors Cut Speed Teardown
A Taylors Cut teardown: Lighthouse mobile 79 to 90 (LCP 3.96s to 3.31s, TBT 362ms to 182ms) and desktop 99 to 100, measured before and after on the same day. Including the three fixes that did nothing.
Answer in 30 seconds
Lighthouse mobile went 79 (LCP 3.96s, TBT 362ms) to 90 (LCP 3.31s, TBT 182ms) and desktop went 99 to 100 (LCP 0.90s to 0.71s) in one speed pass, measured on the same day with Lighthouse 13.5, three runs, median. The wins were a hero image ladder at quality 65 with a corrected sizes value, content-visibility auto on seven below-fold homepage sections and the footer, and removing a 40 KB snackbar library from the first-load graph. Inline CSS, removing font preloads and disabling the JavaScript preload changed nothing measurable.
Taylors Cut is a self-initiated concept build: a streetwear storefront with weekly drops, cart, promo codes and product structured data. No client affiliation, no invented metrics. The numbers are Lighthouse 13.5, both presets, three runs, median, measured the same day before and after the change, the after run on deployed commit 84faa1d. The same rows are on the build breakdown.
Starting score: 79 · LCP 3.96s · TBT 362ms (mobile). Ending score: 90 · LCP 3.31s · TBT 182ms. Desktop, same day: 99 · LCP 0.90s to 100 · LCP 0.71s.
Eleven points on mobile and most of a second off LCP, with no rewrite. Here is the diagnosis, the fixes in order, and three reasonable-sounding fixes that moved nothing.
The Diagnosis: Three Numbers That Did Not Agree
A score hides which layer failed, so I read three numbers first:
- TBT 362ms. Total Blocking Time is main-thread work. 362ms means JavaScript held the thread long after first paint, which points at the first-load JS graph, not the network. Three contributors: seven below-fold homepage sections plus the footer were laid out and painted on load whether or not anyone scrolled;
notistack, a 40 KB snackbar library imported for one toast, shipped on every route; and MUI was still in the error boundary and on the home page. - LCP 3.96s. The largest element took almost four seconds on the mobile preset (150 ms RTT, 1.6 Mbps, 4x CPU). That element is a photograph, so the browser had to discover, download and decode it.
- The hero file.
sizesdid not describe the box the hero actually fills (calc(100vw - 48px)once padding is applied), so the browser picked a candidate for a wider box and a phone downloaded a desktop-sized photograph.
Bytes and main-thread work were both wrong at once, and both had to be fixed.
The Fixes, In The Order I Applied Them
1. Rebuilt the hero image ladder and corrected sizes. Device widths now start at 360, 420, 500, 640 and 750, sizes is calc(100vw - 48px) because that is the box the hero fills once padding is applied, and quality moved to 65.
2. Added an explicit image preload and a Cloudinary preconnect. The hero URL is in the head, so the fetch starts while CSS is still parsing instead of waiting for layout.
3. Deferred the below-fold work with content-visibility: auto. Seven homepage sections and the footer skip layout and paint until they approach the viewport. The content stays in the HTML, still indexable and still there for a visitor who scrolls.
4. Replaced notistack with a local snackbar. One small component, no dependency, about 40 KB less in the first-load graph.
5. Took MUI off the critical path. No MUI in the error boundary or on the home page, and the mobile drawer mounts on first tap.
6. Raised the browserslist floor. Modern targets mean less transpiled output and fewer legacy polyfills.
Each is a contained change, reviewable in a normal pull request.
Before And After, Both Presets
| Preset | Before | After |
|---|---|---|
| Mobile (150 ms RTT / 1.6 Mbps / 4x CPU) | 79 · LCP 3.96s · TBT 362ms | 90 · LCP 3.31s · TBT 182ms |
| Desktop | 99 · LCP 0.90s | 100 · LCP 0.71s |
TBT roughly halved after the snackbar swap and the deferral. LCP dropped 0.65s after the ladder and the preload. Desktop was near the ceiling and gained the last point.
The older row stays published, unchanged: 73 on the 150 ms lab preset and 83 on a measured 16 ms connection, dated 2026-10-15. It answers a different question, how much of a score is the connection assumption, and it is this build's pre-change evidence. Deleting it to publish a better number is selective publication, the same defect as inventing one.
What Did Not Help
- Inlining critical CSS. It removed a render-blocking request, not the bytes LCP was waiting on. No movement beyond noise.
- Removing the font preloads. The fonts already loaded late enough not to gate the hero, so removing the hints only risked a flash of fallback text.
- Disabling the JavaScript preload. TBT is main-thread execution, not module fetch order, so preloading less JavaScript did not make the thread less busy. The bundle had to get smaller, not arrive slower.
The pattern: a fix helps when it changes what the metric measures. Beyond that it is motion, not progress, which is why each change gets measured before and after, not once at the end.
The Lesson If You Are Buying This Work
The cheapest wins were image sizing and render deferral. Neither is a rewrite, neither is a framework migration, and neither needed a design change. They needed someone to read the network panel, the coverage panel and the main-thread flame chart, then make six small, reversible edits. A 79 is not fixed by a new stack; it is fixed by finding the two or three things the page pays for and does not use.
Both earlier teardowns make the same point differently: a 768 KB preload nobody rendered cost seven seconds of paint, and a hero hidden behind an entrance animation made a fast page look slow. This one adds a third shape: nothing was broken, everything was bigger and earlier than it needed to be.
The live build is at taylorscut.vercel.app, the dated rows are on the breakdown, and the method is under the sprint's scope. If you would rather have the audit, the free grader takes about 30 seconds and will tell you your site is fine when it is fine.
Work with me
Need the same speed work on your site?
Seven live builds with dated measurements and fixed prices published up front. Read the breakdowns, or send me the rough version of your problem.
Not sure which service fits? Find your fix in three questions →
Related reads
I Put This Portfolio on a 200 KB Budget
The root motion provider taxed every route. The measured before and after, the per-route budget gate, and why PageSpeed says 100 and a stricter profile says 93.
How a 768 KB Preload Made a Storefront Take 7.2s to Paint
A Small Batch Brine teardown: Lighthouse mobile 74 with a 7.2s LCP, one root cause, and the fixes that took it to 97 (LCP 2.0s). Dated and reproducible.
From 5.9s to 3.8s: Why a Hidden Hero Makes Everything Slow
A Through Tenon teardown: Lighthouse mobile 75 with a 5.9s LCP caused by an entrance animation, and the fixes that recovered two seconds.