Home › SEO Guide › Speed and what Google actually measures
Speed and what Google actually measures
“The site is slow” is an observation. Google measures three concrete things, and they do not overlap with the feeling of slowness. A site can seem fast and fail two out of three.
The three metrics
| What it measures | In plain language | Good threshold |
|---|---|---|
| LCP main content speed | How long it takes for the largest thing on the screen to appear: the main photo or the headline | under 2.5 s |
| CLS visual stability | How much the page jumps under the reader's finger while it loads. You have experienced it: you tap a button and it moves | under 0.1 |
| INP response speed | How long the page takes to react after you tap something | under 200 ms |
Google uses them as a ranking signal, but a modest one: they do not beat relevance. A slow page that answers the question exactly will still sit above a fast one that does not. Speed matters most between otherwise equal pages, and enormously for people who actually buy.
The most instructive case: the page that jumped
saltim.ro, a mattress store
Measured on 5 July 2026, visual stability was 0.224, that is, more than twice as bad as the acceptable threshold. In practice, on a phone, the content visibly rearranged itself while the person was reading.
The cause was not server speed, but a combination of two settings that were each correct on their own. Images were lazy-loaded, so as not to weigh down the page, but their dimensions were not declared. Until a product photo loaded, the browser put a placeholder with different proportions in its place. When the real, square image arrived, everything below it moved.
The fix was a style rule that reserves the correct space for each type of image from the start. Result: from 0.224 to 0.01. Twenty times below the threshold.
It is worth remembering why the problem appeared: nobody made a mistake. Someone turned on lazy loading, which is good practice. The theme did not declare image dimensions, which is an old oversight. Each decision was reasonable, the combination was not.
The improvement we refused to make
Also on saltim.ro, the biggest possible improvement in response time was obvious: keeping pages ready-built, instead of rebuilding them for every visitor. It would have brought the response time down from almost a second to almost nothing and would have visibly raised the score.
We did not do it.
The reason: the platform puts a unique security key in every page, and the “add to cart” button checks it. If you keep the page ready-built and serve it to several people, they all get the same key, expired.
The effect would have been a better speed score and a shopping cart that refuses orders. That is, the exact opposite of the reason the site exists.
It can be done safely, but it requires extra construction, which delivers the key separately from the page, and testing on a copy of the site, not on the live one. We wrote this in the report as a postponed item, with the reason; we did not stay silent and we did not tick it off either.
There is also a convenient variant, which you will come across: keeping pages ready-built only for bots, not for people. The score goes up, because the measuring tool is a bot. Real customers feel nothing. It is not an optimization, it is for show.
Two small interventions with real effect
From goai.ro, both made in August 2026, both a few lines long.
Priority for the main image
We measured what exactly “the largest thing on the screen” is on articles: the cover photo. The browser, however, treated it like any other image, so it discovered it late. An attribute that tells it “this one is important, fetch it first” solved the problem. Checked afterwards: all article pages now mark it as a priority, and the rest of the images remain lazy-loaded, as they should.
Tracking scripts, deferred until the first interaction
The analytics code loaded together with the page and consumed over 250 kilobytes before the reader saw anything. Now it loads at the first sign that there is a human in front of the screen: a click, a touch, a keystroke.
It has a cost, which we tell the client every time: the visitor who opens the page, reads without touching anything and leaves is not counted. You gain speed, you lose a little of the accuracy of the statistics. It is a choice, not a free improvement.
Where slowness usually comes from
- Unprepared images. A 4 MB photo taken with a phone, uploaded directly. The most frequent and the easiest to fix.
- Too many plugins. Each plugin adds code to every page, including the ones where it does nothing.
- Undersized hosting. If the server takes over a second to respond to a simple page, no optimization within the site makes up for it.
- Accumulated tracking code. Analytics, heatmaps, chat, remarketing. Each was added by someone, once, for something.
A special case, from the same area: on saltim.ro we found debug mode active in production. It slowed the site down and, worse, displayed internal details on any error. Turning it off brought the response time on the product page down from 1.5 to 1.2 seconds. This is the kind of thing that does not show up in any automated speed report.
How to measure it yourself, correctly
- Use PageSpeed Insights, Google's free tool. Enter the URL and read the result for mobile, not desktop. Most visitors are on a phone, and that is where Google looks.
- Look at the top, at the field data, if there is any. Those are measurements from real visitors. The figures below are a lab test, useful for diagnosis, but not reality.
- Do not go by a single test. The score swings by five to ten points from one run to the next, depending on server load. Test three times.
- Also test an inner page, not just the homepage. The homepage is usually the most polished one.
Key takeaways
- The three metrics are: how long it takes for the content to appear, how much the page jumps, how long the response is delayed.
- Speed is a modest ranking signal and a major sales factor.
- Test on mobile, three times, and on an inner page.
- The most frequent real cause is unresized images.
- If someone proposes an improvement that sounds too simple, ask what can break. Sometimes the answer is “the shopping cart”.
- A good score obtained only for the measuring tool does not help any real customer.