The speed that actually costs you

Speed is discussed as a ranking factor, which undersells it. A slow site loses people before they have seen anything, and they do not report it, they just leave.

The published thresholds

Google publishes three measurements with defined good thresholds, assessed at the 75th percentile of real page loads and measured separately for phones and computers.

• Largest Contentful Paint, how long until the main content appears: good is 2.5 seconds or less, poor is over 4 seconds.
• Interaction to Next Paint, how quickly the page responds when touched: good is 200 milliseconds or less, poor is over 500.
• Cumulative Layout Shift, how much the page moves while loading: good is 0.1 or less, poor is 0.25 or more.

Two things about the method matter as much as the numbers. It is the 75th percentile, so your slowest quarter of visitors sets your score rather than the average. And the field data comes from real users over a rolling 28-day window, which is why a fix does not show up for weeks and why lab tools and field data disagree.

What is actually slow

1. 01Images. Nearly always the largest thing and nearly always larger than needed. Serve them at the size they display, in a modern format, and let the browser choose per screen.
2. 02Third-party scripts. Chat widgets, analytics, pixels, review platforms, consent tools. Each runs on the main thread and they are the usual cause of poor interaction scores.
3. 03Themes and templates with sliders, carousels and video headers, which ship slow by design.
4. 04Fonts, particularly several weights loaded before any text can appear.
5. 05Anything added since the last time you measured, which is the honest answer most of the time.

Fixing layout shift

• Give every image and video an explicit width and height, or a CSS aspect ratio, so the space is reserved before the file arrives.
• Reserve space for anything that appears late: banners, consent notices, promotional bars, advertising slots.
• Never insert content above something the reader is already looking at.
• Load fonts so text does not reflow dramatically when the real font arrives.

Keeping it fast

Sites do not become slow in one decision. They become slow one plugin, one font and one tag at a time, each defensible alone. A written limit on page weight and third-party scripts turns that drift into a decision, and makes adding something a trade rather than an addition.

Images, concretely

Since this is nearly always the largest part of the problem, it is worth being specific rather than saying optimise your images.

• Serve them no larger than they display. A photograph shown 600 pixels wide does not need to be 4000, and a camera or phone original always is.
• Use a modern format. The same picture in a current format is typically a fraction of the size of an old one at the same visible quality.
• Offer several sizes and let the browser pick, so a phone does not download the desktop version.
• Load the ones below the fold lazily, and never the main one at the top, which is usually the measured element.
• Compress properly. Most photographs survive substantial compression invisibly, and the difference is often several times the file size.

Why the tools disagree

A frequent source of confusion. A lab test runs the page once, on a simulated device, under conditions somebody chose. Field data is what happened to real visitors on their own devices and connections over the preceding weeks.

So they measure different things and both are useful. Field data tells you whether there is a problem; lab tools tell you what is causing it and whether a fix worked, immediately, rather than in three weeks. Use the lab tool to iterate and the field data to decide whether you needed to.

The thing worth refusing

Most speed problems arrive as a request rather than as a decision: a chat widget somebody read about, a review badge, a tracking tag for a campaign that ended, a font for one heading. Each costs a fraction of a second and none of them is the problem on its own. A written rule helps, and the rule can be as crude as nothing new loads on the site until something comes off, which forces the comparison that otherwise never happens.