Faragiri

Learn Field data and lab data Speed

Five pages at an antiques dealer, ranked by lab score and then by what visitors actually waited, and the two orders are almost exactly reversed.

5Sections
4Min read
4Depth

A lab run is one load on one machine on a simulated connection. Field data is real loads by real visitors over 28 days. Alnmouth Antiques has five pages with enough traffic to be measured both ways, and ranking them by each produces almost opposite orders: the page scoring 94 in the lab made visitors wait 4.1 seconds.

The five read, in field seconds against lab score: furniture 4.1 and 94, silver 3.3 and 88, ceramics 2.6 and 79, the shop's about page 2.0 and 71, and clocks 1.8 and 61. The window is the 28 days to 30 May 2026.

Neither number is wrong. Lab data says why, field data says whether, and only the field number is used for the assessment, which is the single fact that decides where an afternoon should go.

Why the orders reverse

A lab run happens on whatever machine asks for it, on a throttled connection chosen by the tool, with nothing cached and nobody logged in. The furniture page that scores 94 is light, and it loads a map and three photographs after the first paint, which a lab run finishes before it measures and a visitor on a train does not.

Field data

What visitors actually waited on five pages

The order here is almost the reverse of the order by lab score, and the lab score sits in the note beside each bar.

Largest Contentful Paint in seconds at the 75th percentile, Alnmouth Antiques, 28 days to 30 May 2026.

  1. FurnitureScores 94 in the lab. A map and three photographs load after the first paint.4.1
  2. SilverScores 88. Same template, fewer photographs.3.3
  3. CeramicsScores 79, and misses the threshold by a tenth of a second.2.6
  4. About the shopScores 71. Almost nothing on it, and no map.2.0
  5. ClocksScores 61, the worst in the lab and the fastest for visitors.1.8

Lab runs stop before late-loading maps and photos finish on a visitor's train connection, while the clocks page fetches its single heavy image right away.

The clock page scores 61 and its visitors wait 1.8 seconds, because it is heavy and everything heavy about it is a single image the browser starts fetching immediately. A score is a weighted average of laboratory measurements; a field figure is what happened to people.

The percentile, and why the average is no help

The field figure is read at the 75th percentile, not the average. Alnmouth's furniture page averages 2.4 seconds and is assessed on 4.1, because a quarter of loads are slower than that and the assessment is taken from that end of the distribution.

Distribution

Why improving the average misses the assessment

Alnmouth's furniture page looks reasonable on average, yet a quarter of visitors wait past 4.1 seconds.

Largest Contentful Paint in seconds on the furniture page

Average load75th percentile (assessed)

  1. Furniture page2.4 seconds4.1Older phones running a slow script hold back the slowest quarter of visits.

Making the median faster leaves the 75th percentile untouched unless the fix speeds up older phones.

This changes what a fix has to do. Making the median faster moves the average and not the assessment; the work is on whatever makes the slowest quarter slow, which at Alnmouth was visitors on older phones running a script that the newer ones finished instantly.

Where this stalls: the 28 day window

A fix takes 28 days to appear fully, because the window rolls. Alnmouth shipped an image change on 6 June and the field figure on 20 June was a blend of a fortnight of the old page and a fortnight of the new one, which read as a disappointing half improvement and was in fact the whole improvement, half reported.

Rolling window

Why a mid-window check reads as half an improvement

Field figures blend the previous 28 days of loads, so checking at two weeks mixes old and new pages.

  1. 6 JuneImage fix shipsNew loads begin improving, while almost four weeks of old loads remain in the window.
  2. 20 JuneHalf-reported blendA fortnight of old page data and a fortnight of new data mask the full gain.
  3. 28 daysFull window clearsShip one change and wait a full month before shipping another so effects stay separable.

Making a second edit in week two leaves a better number with no way to tell which change worked.

The stall is a second change made in week two because the first appeared not to work. Two changes inside one window cannot be told apart afterwards, and the shop then has a better number and no idea which change produced it. The rule that costs nothing is to ship one thing and wait a month.

The page with no field data at all

A page with too few visitors has no field data at all, and a new page therefore cannot be assessed until it has some. That is not a fault to fix; it is the reason lab data exists. On a page nobody has visited yet, a lab run is the only evidence available and it is genuinely useful, so long as nobody reports the score as though it were the assessment.

Low-traffic pages

Choosing evidence when a page has no field data

New or quiet pages have no 28-day visitor record, making realistic lab testing the only option.

Does the page have enough traffic for a field figure?
  • Yes
    Use the 28-day field report to judge pass or failReal visitor loads settle the assessment.
    • Read lab runs only to diagnose bottlenecks
    • Do not report a lab score as a pass
  • No
    Run a lab test on realistic visitor hardwareLab data is the only evidence until traffic builds.
    • Throttle to a mid-range phone and poor connection
    • Do not rely on an office desktop on fibre

Testing on office fibre hides delays experienced by visitors on phones in rural Northumberland.

Test on a mid range phone and a poor connection, not on the office machine. Alnmouth's whole team works on desktops on fibre, which is the environment in which every one of these pages is fast, and the shop's own visitors are largely on phones in rural Northumberland.

What to open

PageSpeed Insights puts both on one screen with the field data first, free, and it replaced running a lab tool alone and assuming its score was the number being used. The order on that screen is the argument of this page rendered as an interface.

Tool layout

How PageSpeed Insights orders field and lab data

Putting real visitor data above the simulated run keeps the assessment separate from the diagnosis.

pagespeed.web.dev
  1. 01Real visitor loads (Core Web Vitals)
  2. 02Sourced from Chrome User Experience Report (CrUX)
  3. 03Single load score and waterfall diagnostics
  4. 04Query the dataset directly for wider patterns
  1. 01Top panel: 28-day field assessmentPlaced first because only field numbers decide whether the page passes.
  2. 02Traffic requirementQuiet or newly launched pages lack enough traffic and show no field data here.
  3. 03Bottom panel: simulated lab runExplains why a page is slow, not whether real visitors experience it.
  4. 04Direct CrUX queryUseful when a single URL's report does not give enough context.

The layout of PageSpeed Insights mirrors the rule: check whether visitors wait before asking why.

The Chrome User Experience Report is where the field half comes from and it can be queried directly when one page's figure is not enough, free, replacing guessing whether the machine you tested on is representative of anybody. It only holds pages with enough traffic, which is where the missing field data on new pages comes from.

Which lines of a lab run are worth acting on is which numbers to act on, and which to ignore, the three field measurements themselves are the three measurements, and the thresholds Google uses, and what a pass is worth once it arrives is the honest size of the ranking effect. All of it is how fast it loads, and what that is worth in search.

Alongside this

Above this