Faragiri

Learn Reading Lighthouse Speed

A brush maker worked through a Lighthouse report from the top and spent a fortnight on savings that were never going to arrive.

4Sections
4Min read
4Depth

The estimated savings are a model, not a measurement. Bickerstaffe Brushes makes household and industrial brushes in Lancashire, ran a lab report on its main category page in April 2026, and got eleven opportunities listed, and three of them touched the LCP element. The other eight added up to a larger predicted saving and were worth almost nothing.

The savings column is a simulation: it estimates what would happen if a request were removed, on the connection the tool chose, holding everything else fixed. Real pages do not hold everything else fixed, and a removed request is often replaced by waiting for something further down the queue.

Read the LCP element first

The LCP element is named in the report and it is the first thing to read. At Bickerstaffe it was a photograph of a yard brush, 1.1 megabytes, loaded by a script after the stylesheet had finished. Everything that delays that one file is worth doing and everything else is worth reading and mostly not doing.

Opportunities

Three fixes touching the LCP photo vs the other eight

Simulated savings add up bigger on paper for the eight unrelated items, yet only the three on the LCP chain moved field speed.

Three LCP opportunitiesAll three touched the 1.1-megabyte brush photograph.
Eight other opportunitiesDid not touch the LCP element.
Predicted lab savings0.4, 0.3 and 0.2 seconds2.9 seconds
Actual field improvement1.6 secondsNothing anybody could see
Why it behaved that wayBroke a blocking chain (stylesheet to script to photo)Replaced waiting with another item in the queue

Read the named LCP element first and sort work by what delays that file rather than by simulated savings.

Sort by what the page is actually waiting for, not by the size of the saving. The three opportunities that touched the brush photograph predicted savings of 0.4, 0.3 and 0.2 seconds and together moved the field figure by 1.6 seconds, because they removed a chain rather than a file. The eight that did not touch it predicted 2.9 seconds between them and moved the field figure by nothing anybody could see.

The lines that are usually noise

An unused CSS warning on a page that renders in 1.4 seconds is noise. It is nearly always present, it is nearly always large, and removing unused rules from a stylesheet is a week of work on a shared theme for a saving the visitor cannot perceive.

Lab reports

Six lines from one report, and whether they were worth acting on

The three worth doing touched the file the page waited for, while the largest predicted savings were noise.

One category page, April 2026

  1. YesPreload the largest imageThe yard brush photograph, 1.1 megabytes, and the thing every visitor waits for.Yes
  2. YesRemove the third party font stylesheet from the headRender blocking, one line to change, and it shortened the chain.Yes
  3. YesServe the image in a modern formatSame file, a fifth of the bytes, and no visible difference on a phone.Yes
  4. NoRemove unused CSSThe largest predicted saving on the list and a week of work on a shared theme.No
  5. NoReduce unused JavaScriptSecond largest. None of it runs before the first paint.No
  6. Not checkedWhether the three fixes hold on a busy FridayThe server has never been measured under load, and nobody has looked.Not checked

Look below opportunities to the diagnostics request chain diagram: that is where the stylesheet-to-script delay appeared.

The render blocking list is worth reading and the rest usually is not. Anything in the head that has to finish before the first paint delays every visitor, and that list is short and specific. Bickerstaffe had a font stylesheet from a third party in it, which is one line to change and was one of the three that mattered.

The diagnostics section is more useful than the opportunities section, and it is below it, unheaded and easy to miss. The request chain diagram there is what showed the brush photograph waiting on a script that was waiting on a stylesheet, which is the actual fault and appears in none of the eleven opportunities as such.

Where this stalls: working from the top

The report is ordered by predicted saving, which reads as a priority list and is not one. Working from the top is the natural thing to do, it is what Bickerstaffe did for a fortnight, and the field assessment did not move because nothing that a visitor waits for had changed.

False progress

Lab score rose after working from the top while field speed stayed flat

Sorting by predicted saving pushed Bickerstaffe to spend a fortnight on items visitors never wait for.

Lighthouse lab score (weighted average), Bickerstaffe category page

Start of fortnightAfter two weeks

  1. Category page lab score6481The score climbed on paper, yet the field assessment did not move at all.

Because the lab score is a weighted average, fixing low-impact items raises the score while leaving the LCP bottleneck intact.

A score is a weighted average, so it moves when unimportant things move. The score went from 64 to 81 in that fortnight, which felt like progress and was a report agreeing with itself. What the field data said is the subject of why your score and your report disagree.

Running it honestly

Run it three times, because one run varies by several points. A lab run on a shared machine competes with everything else on that machine, and Bickerstaffe's page scored 64, 71 and 68 in three consecutive runs on the same afternoon with nothing changed. A single run is a sample of one and it is treated as a fact more often than any other number in this cluster.

Test discipline

Three rules for running a lab audit without fooling yourself

Single local runs bounce around with machine load, so treat one score as a noisy sample rather than a fact.

Reliable lab checks
  • Run three timesOne run varies by several points on a shared machine; average or compare repeated runs.
  • Use PageSpeed InsightsRuns on Google's hardware and prints real field data right above the lab numbers.
  • Use Chrome DevTools for draftsBuilt into the browser since replacing the old pre-2019 extension; best for unpublished pages.

Three runs on the same afternoon scored 64, 71 and 68 with nothing changed

Hosted runs remove local office machine variation, though small run-to-run jitter still happens.

PageSpeed Insights runs the same audit on Google's own machines and prints the field data above it, free, and it replaced running a lab tool alone and assuming its score was the number being used. Using it rather than a local run removes the variation from the office machine and does not remove the variation.

Lighthouse itself is built into Chrome's developer tools, free, and it replaced the separate extension that had to be installed before 2019. It is the right tool for a page that is not published yet, and its report is identical to the lab half of the hosted one.

Which of the three field measurements a fix is aimed at is the three measurements, and the thresholds Google uses, and what the whole exercise is worth in search is the honest size of the ranking effect. Both are inside how fast it loads, and what that is worth in search, and the yard brush photograph is the only line in the report that ever mattered.

Alongside this

Above this