Storefront Speed Journal

Reviewed November 16, 2026

Your Store Is Not Slow on Your Phone: Testing on the Mid Range Device Your Customers Use

Test on the device class the grade comes from. Google field data is dominated by mid range Android phones on ordinary mobile connections, and Lighthouse lab throttling is calibrated to a Moto G class device for exactly that reason. A store that feels instant on a flagship phone over Wi-Fi can post a failing 75th percentile LCP in the Chrome User Experience Report. Match the test to the traffic: throttled lab runs, a real mid range handset if you can get one, and field data as the verdict.

Why your own testing keeps passing a failing store

Every part of a developer test setup is flattering. The flagship phone has a processor several times faster than the median device in your traffic. The office connection is faster and more stable than a bus route 4G connection with competing signals. The browser cache is warm because you visit the store all day, while your customers arrive cold from an ad click. And you test the pages you built, at the times the office is quiet. Stack those four advantages and a store can feel sub second in the hand of the person who made it while its field LCP sits over four seconds for the people paying for it.

Field data corrects for all of it by construction. The Chrome User Experience Report aggregates real visits, real devices, real networks, over 28 days, and the Core Web Vitals grade is read at the 75th percentile, meaning a quarter of your visits are worse than the number you are graded on. That design is deliberate: a store is not fast if it is fast for its luckiest visitors. When the lab and the field disagree, the field is not being unfair. It is describing a different, larger, slower world than the one in your office, and it is the world your revenue comes from.

Lighthouse throttling exists to bridge the two without buying a phone farm. Its simulated throttling applies a slower CPU, around four times slowdown in the default mobile configuration, and a constrained network profile modeled on real mid range performance. The Moto G class reference device is not an insult to your customers. It is a statistical portrait of them, and a page that cannot perform there cannot perform for the median visit, however it behaves on the device in your pocket.

A testing setup that tells the truth

Start with the lab discipline: private window, cache disabled or a fresh profile, mobile emulation with throttling on, and the templates that matter, home, a top collection, a top product, run more than once and read the median, because single runs swing. Record the LCP element and the phase breakdown each time, since the point of the lab is diagnosis, which phase is spending the time, not the score, which is a lagging summary of the phases. A lab run read properly gives you the suspect list. Read as a grade, it gives you a number to argue with, and arguments do not ship fixes.

Add one real device if the budget allows any hardware at all: a current mid range Android, bought once, kept on a prepaid SIM, used on the mobile network away from the office router. The gap between emulation and a real handset on a real network still surprises experienced teams: radio behavior, thermal throttling on a warm phone, and the true cost of heavy JavaScript on a modest chip all show up there first. Test the full journey on it monthly, product to cart to checkout handoff, and note where your own thumb starts waiting. Your thumb, on that device, is a better instrument than your opinion on your flagship.

Then let the field data referee. After a fix ships, the lab improves immediately and the field window refills over weeks, and confusing those two clocks causes most speed arguments inside companies. Report them separately and by name: lab, today, diagnostic, and field, 28 day, the grade. When a stakeholder asks why Search Console has not turned green the morning after a fix, the answer is the window, and it should be given before the fix ships, not discovered in the meeting where the fix gets declared a failure.

What mid range testing changes about priorities

JavaScript cost moves up the list. On a flagship, a heavy app script parses in a blink. On a mid range chip, parse and execute time scale brutally, and the same storefront can spend seconds of main thread time before it responds properly to a tap, which is where Interaction to Next Paint failures are born. Stores that test only on fast hardware systematically undervalue script removal, because on their device the scripts are free. On the customer device they are the tax everything else pays, and the audit order flips: scripts first, then images, then the rest.

Image priorities sharpen too. A hero that loads acceptably on Wi-Fi and a flagship becomes the whole story on a throttled connection, and decisions that looked like taste, slider versus single image, video hero versus poster, stop being taste. They are arithmetic at 1.6 megabits with a four times CPU penalty. The mid range test does not just find problems. It settles internal arguments, because a preference that costs two seconds on the reference device has to argue for its two seconds in front of everyone, with the number on the screen.

Close the loop by writing the test conditions into the release checklist, the same document that holds the drawer test and the spacing test: throttled lab run on the three templates, LCP phases recorded, real device journey monthly, field review at the window boundary. Speed work fails organizationally far more often than it fails technically, because the conditions that reveal slowness are absent from the room where the store is judged. Put the slow phone in the room, literally or by emulation, and the store starts getting judged by the standard its customers already use.

Related reading

Related reading on this site: Shopify App Audit: Which Apps Load on Every Page and Cost You LCP, Image Widths in Liquid: The srcset Fix That Shrinks Every Hero.

Frequently asked questions

Why does my store feel fast but fail Core Web Vitals?

Because your test setup is flattering: a flagship phone, fast office Wi-Fi, a warm cache, and quiet hours. Field data grades real visits at the 75th percentile over 28 days, dominated by mid range phones on ordinary networks. When lab impressions and field data disagree, the field is describing your customers world, not your office.

What device does Lighthouse mobile testing simulate?

A Moto G class mid range Android, with roughly four times CPU slowdown and a constrained network profile in the default mobile configuration. That reference device is a statistical portrait of median traffic, which is why passing on a flagship phone proves little about the grade.

Should I buy a cheap Android phone for speed testing?

If you can, yes. One current mid range Android on a real mobile network reveals radio behavior, thermal throttling, and true JavaScript cost that emulation approximates. Test the full journey on it monthly, product to cart to checkout, and use throttled lab runs for day to day diagnosis.

Why has Search Console not improved right after my speed fix?

The field grade uses a rolling 28 day window of real visits, so it refills gradually as new, faster visits replace old ones. Lab scores improve the day the fix ships and are the right immediate check. Report lab and field separately, and set the field expectation before the fix ships.