Faragiri

Learn Speed SEO · Core Web Vitals, Field data and lab data, Reading Lighthouse and 2 more

The three Core Web Vitals with their thresholds, why a Ludlow merchant passed one test and failed the other in the same week, and what speed is worth.

5Branches
5Nodes below
3Depth

Hollowbrook Wines, an online merchant working out of a unit in Ludlow, scored 91 for performance in Lighthouse in April 2026 and failed Core Web Vitals in the same week. A Lighthouse score of 91 and a field LCP of 3.4 seconds are both true at once, and understanding why is most of what this subject is.

The three measurements are worth memorising because they are short. LCP under 2.5 seconds, INP under 200 milliseconds, CLS under 0.1. Largest Contentful Paint is when the biggest thing on the screen finishes drawing. Interaction to Next Paint is how long the page takes to visibly respond after somebody touches it. Cumulative Layout Shift is how far things move after they have appeared, which is the one measured in no unit at all.

The number that fails is not the average

The thresholds are read at the 75th percentile, not the average, which changes what a fix has to achieve. Hollowbrook's average LCP was 2.1 seconds and it failed anyway, because a quarter of loads were slower than 3.4 and the assessment is taken from that end. Fixing the median does nothing here. The work is on whatever makes the slowest quarter slow, and at Hollowbrook that was a producer photograph served at full resolution on the eleven pages nobody had rebuilt.

Slowest quarter rules

Why a 2.1 second average LCP still failed at the 75th percentile

Core Web Vitals assess the slowest quarter of real loads; unoptimised producer images pushed that threshold to 3.4 seconds.

LCP duration in seconds (75th percentile decides pass or fail)

Passing threshold (under 2.5 seconds)
0s5s
  1. Hollowbrook average LCP2.1 secondsLooks healthy on a mean, but averages do not count
  2. 75th percentile LCP3.4Quarter of loads slower than 3.4 due to full-size producer photos

Note also that INP replaced FID in March 2024, shifting focus from first response to the worst interaction across the entire visit.

One of the three changed recently enough to matter for anything written before it. INP replaced FID in March 2024, and the two measure different things: the old one timed the first response and the new one takes the worst interaction across the whole visit. A page that passed comfortably on the old measure can fail the new one without changing.

Lab data says why, field data says whether

Lighthouse runs the page once, on a simulated connection, on the machine asking. It sees nothing about real visitors and it can tell you which request blocked the render, which is exactly the information a field measurement cannot carry.

The same page, two numbers

What each of the two measurements is actually able to know

Neither column is the honest one. They answer different questions, and only the left one is read when a position is decided.

Field dataReal Chrome visitors, 28 rolling days, read at the 75th percentile
Lab dataOne simulated load, on the machine that asked for it
Who it measuredPeople who loaded the pageNobody
How many loads28 days of themOne
The connectionWhatever each visitor hadA profile the tool chose
What it can tell youWhether the page passesWhich request blocked the render
How fast a fix showsUp to 28 daysImmediately
Used for rankingYesNo

Drawn in compare mode, with no side marked as the answer, because the argument this figure exists to end is the one about which of the two is real.

The field number comes from Chrome users who actually loaded the page, over a 28 day window. That window is the reason patience is compulsory: field data is a 28 day rolling window, so a fix takes 28 days to show fully, and a fortnight after a deployment the number is a blend of the old page and the new one. Hollowbrook shipped the image fix on 2 May 2026 and the field figure did not settle until the end of the month.

The division is clean enough to keep. Lab data says why, field data says whether. Only one of the two is used for ranking, and it is not the one with the score out of a hundred.

The honest size of it

Speed is a tiebreaker between comparable pages, not a way past a better one. A page that answers the question badly and loads instantly loses to a page that answers it well and takes four seconds, and no amount of work on the second number reverses that.

Honest size of speed

Where page speed actually moves numbers, from crawl cost to checkout

Speed acts as a narrow search tiebreaker, while stopping layout shift at payment directly protects completed orders.

Practical impact on business outcomes

Minor effectLarger direct effect
  1. Crawl budget on small sitesLarge sites feel slow-fetch indexing drag; small sites barely notice
  2. Search ranking tiebreakerReal between comparable pages, never beats a better answer
  3. Checkout CLS stabilityFixing shifting buttons at payment earned more orders than search did

Conflating the small search effect with the larger commercial checkout effect is how speed work gets sold at the wrong price.

What it is worth is easier to see on the other side of the ledger. Hollowbrook's checkout lost people at the payment step while the layout was still moving under their thumbs, and the CLS fix was worth more in orders than anything it did in search. The search effect was real and small; the commercial effect was real and larger, and conflating the two is how speed work gets sold at the wrong price.

There is a second effect that is easier to miss. A slow site is crawled less, because a crawler that finds pages expensive to fetch fetches fewer of them, which turns a speed problem into an indexing problem on a large site and into nothing at all on a small one.

Where this stalls: optimising the number in front of you

The stall is that the lab score is the number on the screen, it moves immediately, and it is satisfying. So teams drive it from 78 to 91 with deferred scripts and a smaller font payload, none of which touched the producer photograph, and the field assessment does not move because nothing a real visitor experienced has changed.

Where teams stall

The Lighthouse score Hollowbrook moved from 78 to 91 while failing in the field

Deferred scripts and smaller fonts pushed the lab dial into the green band, leaving the heavy producer photograph and real visitor wait untouched.

Lighthouse performance score (0 to 100)

91Lighthouse performance, one run

0100

The green band, which this page sits inside while failing in the field

A report leading with a score out of a hundred without stating the 75th percentile field result hides whether real users waited less.

The tell is a report that leads with a score out of a hundred and never says what the 75th percentile did. Hollowbrook had three of those, from three suppliers, across two years.

What to open in 2026

PageSpeed Insights shows the field data and the lab run on one screen, in that order, which is the order they should be read in. It is free, and it replaces running Lighthouse alone and assuming the resulting score was the number Google was using.

What to open in 2026

Read PageSpeed Insights in order, or query CrUX when one page lacks data

Both tools are free and separate real visitor behaviour from lab diagnostics, though neither knows if your checkout actually takes money.

Does the page have enough Chrome traffic for field data in CrUX?
  • Traffic is sufficient
    Read field data first, then labPageSpeed Insights shows field results above the lab run; query CrUX directly for wider slices.
    • Judge pass or fail from the 28-day field section
    • Use the lab waterfall below it to locate blocking requests
  • New or low-traffic page
    Debug with lab runs, withhold verdictCrUX excludes pages below its traffic floor, so field status cannot be assessed yet.
    • Catch heavy images and render-blocking scripts early
    • Wait for real visitor volume before declaring a pass

Both tools would report the exact same 91 on a broken checkout page that accepted zero payments.

The Chrome User Experience Report is where the field half comes from, and it is queryable directly when a single page's figure is not enough. It is free, and it replaces guessing whether the machine you tested on is representative of anybody. It only holds pages with enough traffic to report on, so a new page has no field data at all and cannot be assessed until it does.

Neither tool knows what the page is for. Both would have reported the same 91 for a checkout that took nobody's money.

The five parts of this

The three measurements and their numbers are the three measurements, and the thresholds Google uses. The 91 against the 3.4 is why your score and your report disagree, and it is the distinction the other four rest on.

The five parts

Five pieces of speed work, and why seasonal noise clouds attribution

Thresholds, the 91 vs 3.4 split, actionable lab lines, ranking vs crawl impact, and serving fixes all connect before you measure outcomes.

Speed system
  • Three thresholdsLCP, INP, and CLS targets evaluated at the 75th percentile
  • 91 vs 3.4 splitWhy lab score and real field report routinely disagree
  • Actionable lab linesFixing the producer photo instead of vanity script tweaks
  • Ranking and crawlModest tiebreaker plus indexing efficiency on larger sites
  • Serving and attributionJune 2026 pass coincided with warm weather, blurring wine sales proof

Telling the search lift apart from checkout gains and seasonal weather requires holding other variables steady.

Knowing which lines in a lab run are worth acting on is which numbers to act on, and which to ignore. What all of it buys in search, said plainly, is the honest size of the ranking effect, and the crawl consequence is what a slow site costs you in pages indexed.

The producer photograph and the deferred scripts are both changes to what gets served, which is the layer that decides whether any of the rest counts. Telling the search effect apart from the checkout effect afterwards is knowing whether any of it worked.

Hollowbrook fixed the images and passed in June 2026. Orders rose in the same month and so did the weather, and nobody held anything still long enough to know which of the two the wine was responding to.

Inside this, 5

Alongside this

Above this