LCP is measured per URL — your problem is per template
Six Lighthouse runs across six page classes on two public sites: LCP varied 2.7× inside one origin while TTFB never passed 50 ms. Fix templates, not pages.
Your LCP score is reported per URL, but almost nothing you can fix lives at URL level. It lives at template level — the homepage, the article, the listing, the locale twin. Run six page classes on two public sites and the spread inside a single origin reaches 2.7× on LCP and 20× on layout shift, with a server response time that never exceeded 50 milliseconds on any of them. The template is the unit of measurement, the unit of blame and the unit of fix.
That has a consequence most performance reports ignore: a site can be simultaneously "passing Core Web Vitals" and shipping a landing template that takes five seconds to paint, because the number at the top of the report describes an average of layouts that share nothing but a domain.
Field data sees URLs, and only the popular ones
Google's Chrome UX Report aggregates real user experiences into two distributions: page-level and origin-level. To appear as a page-level row, a URL has to satisfy the eligibility rules — it must be publicly discoverable, judged by the same indexability criteria search engines apply, and it must be sufficiently popular. The CrUX methodology is explicit that an exact traffic number is not disclosed, only that a page needs "a minimum number of visitors."
Read that twice. Page-level field data is a dataset about your most-visited URLs. The paginated archive you just rebuilt, the city pages, the locale twins, the tag listings, the 300 product variants — these either never appear as rows or are folded into an origin aggregate that mixes an article template with a checkout.
So the two numbers in the report are structurally guaranteed to disagree with the experience of the page a buyer actually lands on. Origin LCP is an average across classes. Page-level LCP exists only for classes with traffic. The classes in between — the ones that are numerous, templated and slow — are exactly the ones neither number describes.
What six runs showed
Method, before the numbers: Lighthouse 13.0.3, mobile form factor with Lighthouse's simulated slow-4G throttling, one lab run per URL, run on 19 November 2026 from a single machine. One URL represents each class. These are readings, not verdicts on either property, and no fix was applied to anything — the point is what variance looks like before anyone touches a line of code.
| Origin | Template class | Performance | LCP | Total blocking time | CLS | TTFB |
|---|---|---|---|---|---|---|
| bbc.com | Homepage | 33 | 10.6 s | 10,408 ms | 0.065 | 19 ms |
| bbc.com | Section index (/news/world) | 35 | 8.4 s | 6,621 ms | 0.014 | 13 ms |
| bbc.com | Article | 23 | 7.8 s | 7,461 ms | 0.221 | 6 ms |
| bbc.com | Locale twin (/mundo) | 51 | 3.9 s | 5,384 ms | 0.011 | 11 ms |
| smashingmagazine.com | Homepage | 81 | 3.0 s | 478 ms | 0 | 47 ms |
| smashingmagazine.com | Category index | 97 | 2.4 s | 48 ms | 0.012 | 39 ms |
Three things in that table are worth more than the scores.
TTFB is a non-issue on every row. Between 6 and 47 milliseconds. Whatever produces a 10.6-second LCP is not the host, the CDN or the database. This is the single most useful line in the table, because the reflex fix for a bad score — move hosting, add a cache layer, upgrade the plan — cannot explain a spread that exists inside one origin on one platform.
The same platform produces 10.6 seconds and 3.9 seconds. The English homepage and the Spanish-language homepage are the same product, same design system, same delivery, same organization. The 2.7× gap is the template and everything the template pulls in — not the infrastructure. Any diagnosis that starts with infrastructure has already missed it.
Layout shift is a template property, not a site property. The article template shifts 20× more than the locale twin on the same origin (0.221 against 0.011) — and 0.221 already sits in "needs improvement" territory, one ad slot away from failing, while the origin average looks unremarkable. On the other origin, the category index blocks the main thread for 48 ms against 478 ms on the homepage: 9.9× between two templates behind one domain. Different class, different JavaScript inventory, different problem.
The four LCP sub-parts map onto classes
LCP decomposes into four phases: time to first byte, the delay before the LCP resource starts loading, the time that resource takes to load, and the delay before the element renders. Each template class tends to fail in a different phase, which is why a single fix applied site-wide so often moves nothing measurable.
| Template class | Usual LCP element | Dominant sub-part | What actually moves it | What does not |
|---|---|---|---|---|
| Homepage / campaign landing | Hero image, or the H1 when the hero is text | Resource load delay | Preload the hero, fetchpriority="high", cut the number of competing above-fold assets | Compressing images that were never the LCP element |
| Article / detail page | Inline hero image | Resource load time, then render delay | Serve the width the layout really renders at; stop shipping a 2000 px original to a 390 px viewport | Lazy-loading the hero — it makes LCP later, not earlier |
| Listing / archive / pagination | A text block in the first card | Element render delay | Self-hosted, subset fonts with font-display: swap; remove render-blocking CSS from the critical path | Image optimisation — there is no image to optimise |
| Locale twin | Same element as the parent locale | Render delay, occasionally TTFB | Make sure the locale does not load a second copy of the framework or a translation payload before first paint | Copying the parent's fix and assuming parity |
| Search results / filtered view | Result text | Time to first byte | Cache the query surface, move the fetch server-side | Client-side rendering tricks |
The pattern is worth stating plainly: on image-led classes, LCP is a loading problem; on text-led classes, it is a rendering problem. Those two have almost no fixes in common, and most performance programmes are built around the first one because it is the one with a vendor for it.
The class where the usual fix does nothing
Listing and archive templates are the clearest case. Their LCP element is a text block, so an image-compression project buys exactly zero milliseconds there, and the delay sits in fonts, CSS and the order of the head. Meanwhile CrUX will rarely give you a page-level row for deep pagination — the URLs are low-traffic, and page two onwards is often not even indexed — so this can be the slowest class on a site while remaining invisible in every dashboard that reads field data.
That combination — expensive, common, and unmeasured — is why archives are usually the last template anyone fixes and the first one users hit on the way out.
How to pick which template to fix first
Three questions, in this order:
- How many URLs does one change cover? A template fix is worth its cost in proportion to the URLs that inherit it. One change to a shared layout that improves 4,000 URLs beats forty page-level tweaks.
- Does this class exist in field data at all? If it has no page-level row, you need lab measurement per class plus a real-user instrument tagged by template before you can claim a win. Do not report an improvement you cannot observe.
- Is it a landing surface? A class people enter on beats a class they only browse into. The article template that receives campaign traffic is worth more than the tag page nobody lands on.
A practical shortcut we use: pull the top entry URLs for the last 90 days, label each one with its template, and count. Three or four classes will account for almost everything — and the one with the worst measured LCP is usually not the one on the fix list.
What to report instead of one number
The single Core Web Vitals figure at the top of a monthly report is the most expensive kind of lie: technically true and operationally useless. It cannot tell you which template to change, and it flatters the site precisely when the landing class is failing.
The fix is unglamorous. Add a template dimension to your real-user monitoring — the web-vitals library plus whatever analytics you already load — and report five rows instead of one. One afternoon of instrumentation, and every decision that follows changes: the rebuild scope, the vendor brief, the definition of done.
We are not claiming a conversion lift here. On most sites the attribution is too thin to carry that claim, and dressing performance work up as revenue is how performance programmes lose their budget the first time a quarter is flat. What is measurable is the class-level gap, and it is large enough to justify the work on its own.
If you are scoping a rebuild, the quote should name the template classes it covers and the class-level target it commits to. A proposal that promises "better performance" without naming the templates is a proposal to improve the homepage and hope.
Two more readings that make this easier to act on: how field data splits the LCP sub-parts in practice in Cloudflare's BEACON dataset, and why CrUX now measures soft navigations separately in the soft-navigation change. For the image-led class specifically, the hero video dilemma covers the decision most marketing sites get wrong.
If you want the six-class read done on your own templates — field where it exists, lab where it does not, mapped to the fix per class — that is the web design and development engagement we run, and it starts with one measurement instead of one promise.
Frequently Asked Questions
What does it mean to measure LCP by template?
It means grouping URLs by the layout that renders them — homepage, article, listing, locale twin, search results — and reading a separate LCP for each group instead of one number for the site. Templates share markup, fonts, CSS and JavaScript, so a fix applied to a template moves every URL in that class at once, while a fix applied to a single URL usually moves nothing.
Why does CrUX not show LCP for my long-tail pages?
CrUX publishes page-level data only for URLs that are publicly discoverable under the same indexability criteria search engines use, and that are sufficiently popular, meaning they received a minimum number of visitors that Google does not disclose. Low-traffic pages, deep pagination and URLs that are not indexed either fall out of the dataset entirely or appear only inside the origin-level aggregate.
Should I optimise LCP on every page?
No. Fix by template, not by page. A template fix covers every URL that shares the layout, and the failure mode is usually identical inside a class: the same LCP element, the same dominant sub-part, the same blocking resources. Chasing individual URLs produces a long list of one-off changes that never touch the class that is actually slow.


