Reference Devices
In short: a reference device is a real phone, tablet or computer you test on, chosen to stand for the devices your users actually have. Pick devices across the range your users own, not just your team's own high-end machines, and record exactly what you tested.
Choosing devices
Clause 7.4 requires client-side energy to be quantified on, or for, devices representative of your users. A practical way to do that is to choose from a grid of device categories and performance classes, guided by your analytics:
| Category | High-end | Mid-range | Low-end |
|---|---|---|---|
| Desktop | Modern workstation (6+ cores, dedicated GPU) | Standard desktop (4 cores, integrated graphics) | Budget desktop (2 cores, integrated graphics) |
| Laptop | Performance laptop (MacBook Pro 16-inch class) | Standard laptop (MacBook Air class) | Budget laptop (Chromebook class) |
| Smartphone | Current-generation flagship | Mid-tier smartphone | Budget or older-generation |
| Tablet | iPad Pro class | iPad Air class | Entry-level tablet |
The brand names describe a class of device, not a recommendation of a product. Any device with similar performance fits the cell. The classes and core counts come from the original draft (§16). (Device class examples last reviewed: 2026-10.)
Some practical points:
- Weight towards your users. If most of your traffic is mid-range smartphones, test more of those.
- Do not test only high-end devices. A test set made up of your team's fastest machines is unlikely to represent your users.
- Cover each major category you serve. At Entry level, a common starting point is one device per major category; at Standard, at least two (see Implementation Tiers and Data Quality).
What to record
Clause 8 item 8 requires the devices used for client-side quantification to be disclosed by type, model, operating system and browser version. For example:
- type: smartphone
model: [manufacturer and model]
os: [operating system and version]
browser: [browser and version]
performance_class: mid-range
This is what makes a client-side measurement repeatable and lets readers judge how representative it is.
Please submit any comments you have here.