Two PageSpeed runs produce different scores, and the team mistakes one good result for a fix. To diagnose the issue, you need consistent conditions, multiple measurements, and an understanding of the difference between lab tests and data from real visitors.
Record the settings and page version
Lab and field data reflect different conditions and may show different results. Official documentation.
Alongside the result, record the URL, release date, viewport width, device mode, and network and CPU throttling settings. Add the measurement location, browser, and browser version if available. Don’t treat the default mobile preset as a complete model of all phones. It creates specific lab conditions. If you run the test through a service, save a link to the specific result and its settings. For a manual browser test, note any open extensions and background activity that could affect performance. The next run should use the same settings so that differences reflect the site change.
Separate first and repeat visits
| Data | How it was collected | How to compare it | What to record alongside it |
|---|---|---|---|
| Cold cache | First load | Separate series | Keep conditions consistent |
| Warm cache | Repeat visit | Separate series | Don’t mix it with the first load |
| Mobile emulation | Lab mode | Record CPU and network settings | Not representative of all phones |
| Field data | Real visits | Specify the period | Compare separately |
A first load without a cache and a repeat visit with saved resources answer different questions. Run them as separate series. Specify whether the HTTP cache, service worker, and intended application state were cleared. Don’t delete user data on production for a lab test; use your own test profile. Third-party services may retain state independently, so note their actual requests. If you opened the page without a cache before the change and with a cache afterward, a large difference doesn’t necessarily indicate an optimization. Compare matching modes, keeping the original conditions and actual loaded resources consistent.
Save every run in the series
Run the same unchanged release several times and save every result. For a summary, you can use the median—the middle result when ordered—and show the range. For example, for 2,1; 2,2; 2,3; 2,4; 3,0 seconds, the median is 2,3, and the range is 2,1–3,0 seconds; this is a suggested method, not a universal guarantee of accuracy. Don’t discard a slow run just because it makes the results look worse. Check the waterfall, errors, and changes in third-party responses. If you know why a run was an outlier, describe that separately. Keep the time, metrics, and relevant conditions for each result. This series shows the variation in lab results and helps you assess whether the expected change exceeds the test’s normal fluctuations.
For a local Chrome series, select one Network throttling profile and one CPU throttling profile in Performance. For first loads, keep Disable cache enabled; run a separate series for repeat visits. Record the URL, viewport width, date, and settings alongside each result. The “Slow 4G” profile only sets lab conditions; when comparing with third-party PageSpeed results, account for their own settings.
Change one cause at a time
Suppose you’re reducing the size of the main image for one product. Before the change, load the product page several times without a cache and with the selected network throttling. Then replace only the image and repeat the series. Leave the code, text, and third-party widgets unchanged. Compare when the image request starts, when the image is received, and when it appears. If you also enable caching, run a separate comparison using the previous cache settings; otherwise, you can’t distinguish the effect of the image from the reuse of files. In your results, include the actual measurements for each run and the changes you tested.
Compare a specific metric
The overall score can change because of different components and hide a specific regression. Compare individual metrics and their causes. For LCP, check the candidate element and loading stage; for responsiveness, check the specific click; for shifts, check which elements moved. You can’t claim an INP improvement based on one fast page load without any interaction. If the LCP element changed between runs, find out why: the viewport, content, or font may have had an effect. State exactly what changed—for example, the main image on this product page loaded earlier. Don’t generalize a result from one page to the entire site.
Review field data separately
Field data reflects real devices, networks, and behavior over its own time period. It doesn’t have to match your lab preset. Keep the source, period, and aggregation level: a specific URL or a group of pages. If there isn’t enough data, that doesn’t mean the delay was zero. After a release, the historical period may include both the old and new versions, so don’t expect them to match immediately. Lab testing is useful for diagnosis and checking a change under controlled conditions; field observations help you understand how widespread an issue is among real visitors. These results explain each other but don’t replace each other.
Compare the series before and after the change
Review the median and range across several identical runs, then examine the request or task that changed. One good score doesn’t describe the entire audience. For real-visitor data, keep the time period and level of aggregation: a URL or the whole site.
Read also: Site speed and conversion: how much each extra second of load time costs and What to check on your site right after publishing changes. Site checks are available on the reChecker technical audit page.