Why a site loads slowly on a phone: how we took our home page's Lighthouse score from 40 to 80

A hero video, two WebGL star skies and an animation far down the page were slowing nusrat.tech's home page. We moved their start-up later and in our test (local build, Lighthouse phone profile) the score rose from 40 to 80; this is a lab run, not real-visitor data.

Key points

  • On 8 November 2026, Lighthouse 12.6.1 (phone profile, local build, one run per page) scored the nusrat.tech home page 40 before our changes and 80 after.
  • Total Blocking Time fell from 6,410 to 130 ms and Largest Contentful Paint from 9.0 to 4.8 s, while page weight barely moved: 1,530 to 1,493 KiB.
  • The page was slowed by a first-screen video that, by our reading of the trace, waited for two WebGL star skies to start, and by an animation renderer far down the page that compiled on first load.
  • This is a lab run, not real-visitor data: the field Core Web Vitals in PageSpeed Insights come from 28 days of the Chrome UX Report.

On the phone profile of Lighthouse, the home page of nusrat.tech scored 40 out of 100 in our test. After our changes the same test gives 80, and it looks the same. Below: what we measured, what was slowing it, and exactly what we changed. One caveat up front: this is a lab run on a local build of the site, not the live site, and we explain below how that differs from what Google ranks on.

What we measured

On 8 November 2026 we ran Lighthouse 12.6.1 with the mobile profile (simulated throttling) against a local production build of the site, not the live site. One run per page, before and after the changes. Numbers from the live site, or from Google's PageSpeed Insights, may differ.

Page Score before Score after
/ru (home) 40 80
/ru/services/crm 87 89
/ru/geo 88 91

On the home page the other numbers moved like this:

  • Total Blocking Time: 6,410 ms to 130 ms.
  • Largest Contentful Paint: 9.0 s to 4.8 s.
  • Page weight (Lighthouse's total byte weight): 1,530 KiB to 1,493 KiB.

The weight barely changed. The page got faster not because we removed content but because we stopped starting heavy things too early.

What was slowing it

We read the trace rather than guessing. Three causes:

  1. The video in the first screen. A 515 KB star video replaced the still image when the page woke up, and became the largest paint itself. It waited about 8 seconds, mostly behind the start-up of two WebGL star skies (that cause is our reading of the trace).
  2. The black-hole "fall" animation far down the page. Its renderer compiled about two screens early, during the first load. In the phone simulation that is roughly 3 seconds of blocking.
  3. A font that only appears below the first screen. The daylight font (three weights, three character sets: Latin, Latin Extended and Cyrillic) was preloaded although it is needed only on scrolling.

What we changed

The look is the same; the order of start-up changed:

  • The still image is painted first and stays underneath. The video fades in over it after the page has loaded.
  • Both skies start after the page's load event plus a short idle moment.
  • The fall's renderer compiles when the visitor starts scrolling.
  • The daylight font is no longer preloaded.

Why start-up order beats weight

The page weight moved from 1,530 to 1,493 KiB and the score doubled. On a phone the bottleneck is often the main thread, not the network: while it is busy starting WebGL or compiling a renderer, the page cannot answer taps or paint what matters. So we did not cut content. We moved the heavy parts to the moment when they no longer get in the way of the first screen.

A word on LCP. When the first-screen image is swapped for a video, the browser treats whatever painted later as the largest element and waits for it. We gave the image back the role of first frame: it paints at once and the video appears over it. The cost is one thing only: the video now starts a little later than before.

The CRM and GEO pages were already above 85, so their gains are small: 87 to 89 and 88 to 91. The home page, which had everything listed above, got the big win.

What the metrics mean

Total Blocking Time adds up, for every long task (over 50 ms) between First Contentful Paint and Time to Interactive, the part beyond 50 ms. Google advises staying under 200 ms on average mobile hardware, and says TBT is a lab metric. Ours went from 6,410 ms to 130 ms.

Largest Contentful Paint is when the largest visible element is painted; for a video, the poster or first frame time counts. Google calls 2.5 s or less good and anything above 4.0 s poor. Ours is 4.8 s: better than 9.0 s, but still on the poor side of that line. The work is not finished.

In Lighthouse's scoring TBT carries 30% of the score and LCP 25%. That is why a handful of heavy start-up tasks could hold the whole page at 40. In the before trace they were the fall's renderer (about 3.1 s), two tasks the trace does not attribute (about 2.0 s and 1.3 s) and the script of the second star sky, the one that runs down the page (about 1.1 s; we identified the file by its contents), all in Lighthouse's simulated phone time.

What this run does not prove

  • It is one run per page, on our machine. Another run will give other numbers.
  • PageSpeed Insights runs Lighthouse in a Google data centre, where conditions can vary, so its numbers will differ from ours.
  • Google says it highly recommends good Core Web Vitals (LCP, INP and CLS), which reflect real people's experience. PageSpeed Insights takes them from the Chrome UX Report over 28 days. A lab score of 80 is not what Google sees from your visitors.

What you can check on your own site

  1. Run your page through our free page check: it shows speed and the phone view.
  2. Open PageSpeed Insights and compare its two blocks: real-user data and lab data.
  3. If the lab part is slow, look for what starts at load but is needed later: video, animations, fonts, widgets.
  4. After each fix, measure again on the same page with the same profile, and write down the date.

One measurement guarantees nothing: the numbers shift a little from run to run. That is why we recorded this one with its date and will repeat it after the next changes.

Questions and answers

Why does a site load slowly on a phone?
On a phone the bottleneck is often the main thread, not the network: while it is busy starting WebGL or compiling animations, the page ignores taps and does not paint what matters. Our home page was like that: its weight barely changed, yet the score doubled once the heavy parts started later.
How is a lab score different from what Google sees?
Lighthouse and the lab part of PageSpeed Insights simulate a phone, and the numbers vary from run to run. Google recommends achieving good Core Web Vitals (LCP, INP and CLS), which reflect real people's experience; PageSpeed Insights takes them from 28 days of the Chrome UX Report.
What LCP and TBT values count as good?
Per web.dev, a good LCP is 2.5 s or less and a poor one is over 4.0 s; Google advises keeping TBT under 200 ms on an average mobile device. Our 4.8 s LCP after the changes is still poor by that threshold, so the work is not finished.
Where do I start checking my own site?
Open the page in the free check on nusrat.tech and in PageSpeed Insights, and compare the real-user data with the lab data. If the lab part is slow, look for what starts on load but is needed later: video, animation, fonts, widgets. Measure again after each change and note the date.

Sources

Prepared by the Nusrat Tech team with AI assistance, from the sources below.

  1. Total Blocking Time (TBT) — web.dev (Google), 2025-10-15
  2. Largest Contentful Paint (LCP) — web.dev (Google), 2025-09-04
  3. Understanding Core Web Vitals and Google search results — Google Search Central
  4. About PageSpeed Insights — Google for Developers
  5. Lighthouse performance scoring — Chrome for Developers

More notes

  1. How do you add a business to Yandex Maps in Uzbekistan?

    You add a business through Yandex Business or directly on the map: the profile goes to review, then appears in Maps. Yandex's help does not name Uzbekistan on yandex.ru or on yandex.uz, so check the outcome yourself: after review, search for the company in Tashkent Maps.

  2. llms.txt: what it is, whether your site needs one and how to make it

    llms.txt is a proposal (llmstxt.org): a Markdown file at the site root with a summary and links for AI agents. Google says such files are not needed to appear in its AI search features. According to llmstxt.org, the file can complement robots.txt; in our view it does not replace good pages.

  3. Publishing an iOS app on the Uzbekistan App Store: lessons from two releases

    An App Store app is more than code: the account type sets the seller name, and App Review can send a build back. We went through it twice; here is what to decide before ordering.

All notes

Want this working for your business?

Tell us what you sell and to whom. We answer in Telegram the same day.

Reach usDiscuss your project on Telegram

  • We reply the same working day, Tashkent time.
  • You write to Nodirkhon; he brings in whichever of us your problem needs.