Choose the question first
Are you comparing two code changes, looking for a heavy resource or trying to understand the mobile journey customers actually use? The first often suits a repeatable lab test. The last needs data that represents real visits, if that data exists and can be used.
Select relevant page types and define the start point, interactions, device and network profile. Save the date and version so the comparison does not mix different content or environments.
Use lab data for diagnosis and regression checks
A controlled environment makes it easier to compare a change with a baseline. Inspect the loading sequence, resource sizes and main-thread work; record cache, viewport and simulation. Repeat a run when results vary.
A lab tool that does not interact with the page cannot show the entire interactive experience. web.dev explains that some field metrics require real interactions and that lab proxies have limits. State exactly what each tool measured.
Read field data as a distribution
Field measurement reflects actual visits and varies by device, network and location. Check which population and time period the data represents, whether the page type has enough evidence and whether consent or measurement coverage affects the sample.
Lab and field results can differ without either automatically being wrong. A lab profile may not include slower devices, repeat visits, personalized content or interactions that happen in real use.
Create a simple improvement loop
1) Record a hypothesis and baseline. 2) Change one scoped component. 3) Rerun the same lab scenario. 4) Manually check the affected user tasks. 5) Follow field data when comparable evidence is available. Also note changes to images, third-party scripts and content.
Do not set a universal size budget or promise a score without understanding the page’s purpose and audience. Choose limits the team can measure and revisit as usage and features change.
| Method | Useful for | Cannot answer alone |
|---|---|---|
| Lab test | Pre-release diagnosis and comparable runs | How every real visitor experiences the page |
| Field data | Experiences in the measured visitor group | Which line of code caused the result |
| Manual task check | Barriers in a specific journey | Performance across all devices and networks |