Above the fold, or the first screen, is the part of a web page visible in the browser window when the page opens, before any scrolling. The fold is the bottom edge of that window, so its position belongs to the device and the window size rather than to the page: one URL has one fold on a phone and another on a desktop monitor. It matters for speed because LCP looks only at this region. In web.dev’s definition, the metric “reports the render time of the largest image, text block, or video visible in the viewport.”
One page, two different largest elements
Of the home pages we audited (snapshot 2026-11-09: the original sites, not our proxy, past our validity gate, latest measurement per host), PageSpeed Insights named a different kind of LCP element on mobile and on desktop for 1,163 of 6,877, or 16.9%. Kind here means text or image.
| LCP on mobile | LCP on desktop | Pages |
|---|---|---|
| Text | Image | 853 |
| Image | Text | 310 |
The usual case is a block of text as the largest thing on a phone and an image on a wide screen. To tell this apart from run-to-run variation, we compared repeat measurements of the same page on the same device, a median of 18 days apart: the kind changed on 3.0–3.5% of 3,356 pairs. The gap between devices is about five times that. Nor is it a platform effect. Every platform with at least 250 pages in the sample falls between 15% and 20%, and the two halves of the sample split by measurement date give 16.7% and 17.1%.
Where the fold falls in a PageSpeed run
Lighthouse, the engine behind PageSpeed Insights, does not test on a real phone or monitor. It emulates two windows, 412 × 823 CSS pixels for mobile and 1350 × 940 for desktop, as set in its emulation constants. Those are the two folds behind every lab score, and your visitors’ phones and windows draw the line somewhere else. Google’s thresholds treat the two as separate populations as well: web.dev measures “the 75th percentile of page loads, segmented across mobile and desktop devices.”
What breaks when a page is built for one fold
- Lazy loading by count. A tool that sees only the markup cannot know where the fold will be,
so it counts images. WordPress leaves the first three content images without a
loadingattribute and lazy-loads the ones after them (the threshold was raised from 1 to 3 in version 6.3). The count is the same on every device, and the first screen is not. web.dev is direct about the cost of a miss: “Never lazy-load your LCP image.” - One hero preloaded for every device. A
preloadwithfetchpriority="high"picked from a desktop layout is sent to phones too. Where the phone’s largest element is text, as on 853 of the pages above, that image is not the phone’s LCP, and the priority it gets is taken from the element that is. - Critical CSS cut for one window. The inlined slice is sized to a first screen. Cut at a desktop width, it can miss what a phone shows first, and those parts paint unstyled until the full stylesheet arrives.
Preloading for one device, in code
When the hero exists on one layout only, the preload can carry a media condition, and a browser whose window does not match skips the file. MDN calls this “responsive preloading”:
<link rel="preload" as="image"
href="/hero-wide.webp"
- fetchpriority="high">
+ fetchpriority="high"
+ media="(min-width: 800px)">
How our proxy draws the line
Tilda pages go through a separate pipeline. On the others, where we have measured the page in our own headless Chrome, the proxy takes the LCP element from that measurement, made at three window sizes: the two PageSpeed uses and 1024 × 800. It sends the same HTML to every device, so only one image can get priority. If the windows agree that the LCP is an image but name different ones, the phone’s wins. If one says text and another says image, the measurement is set aside and the proxy reads the markup instead. If they all say text, no image preload is added. Its lazy loading leaves the first three images outside the header and navigation untouched, the same count WordPress uses. Whether the hero can be found in the markup at all is a separate question, covered in our hero image study.