The short answer
Test speed on a real mid-range phone and with PageSpeed Insights, not on your laptop. The targets are Google's three Core Web Vitals: Largest Contentful Paint within 2.5 seconds, Interaction to Next Paint of 200 milliseconds or less, and Cumulative Layout Shift of 0.1 or less, measured at the 75th percentile of page loads, with mobile and desktop counted apart.
On AI-built sites, the usual causes are the same three: the main content waits for JavaScript, images are far larger than they need to be, and too many outside scripts load at once. All three can be fixed. The first one is the hardest to fix in place.
Why it feels fast to you
You built the site on a fast laptop, on good wifi, with the files already cached from the last hundred previews. Your buyer opens a link from LinkedIn on a three year old phone, on mobile data, for the first time. Those are different tests.
- Processor. A mid-range phone runs JavaScript much more slowly than a laptop. A page that is drawn by script waits for that.
- Network. Mobile data has more delay per request. Many small files hurt.
- Cache. A first visit downloads everything. Your hundredth visit does not.
- Screen. A phone shows a narrow layout that you may never have looked at closely.
Google also uses the mobile version of your site for indexing and ranking. Its {a(MFI, "mobile-first indexing guide")} asks for the same main content, headings, structured data, titles and descriptions on mobile as on desktop, and says not to hide main content behind clicks or swipes, because Google will not trigger them.
The three numbers and what moves them
| Metric | Good | What it measures | Common cause on AI-built sites |
|---|---|---|---|
| LCP | 2.5 s or less | When the biggest thing on screen, often the hero image or heading, appears | Hero waits for JavaScript; huge hero image |
| INP | 200 ms or less | How fast the page responds to taps and clicks | Heavy scripts busy on the main thread |
| CLS | 0.1 or less | How much the layout jumps while loading | Images without set sizes; fonts and banners that push content |
The LCP point matters most for rendering. web.dev's LCP guide says the LCP resource should be discoverable from the HTML source. An image added by JavaScript is found late. It also says never to lazy-load the LCP image, and to use fetchpriority="high" on the image most likely to be the LCP element. On a page drawn by script, the browser cannot even start on the hero until the script has run.
Images: the quickest win
Generated sites often use one large image file for every screen size. A phone then downloads an image sized for a wide desktop monitor. Fix it like this:
- Several sizes. web.dev's responsive images guide explains
srcsetandsizes, so the browser picks the right file. Three to five sizes is common. - Set width and height. Give each image its size in the HTML so the layout does not jump when it loads.
- Lazy-load below the fold only. Use
loading="lazy"for images further down. Never on the hero. - Modern formats. WebP or AVIF are usually smaller than PNG and JPEG for the same look. Check the result by eye.
- Text as text. A headline inside an image is slow, invisible to crawlers and hard to edit.
Our post on image formats and sizes has more detail for Webflow sites.
Scripts: count them
Open your page source and count the script tags, then look at where each one loads from. AI-built sites often pick up extra pieces over time: an animation library for one effect, a chat widget, two analytics tools, a font loader.
- List every outside script and who asked for it. Remove the ones nobody can name.
- Load analytics and tag managers after consent, which also helps with privacy rules.
- Replace a large library used for one small effect with CSS where you can.
- Self-host fonts and load only the weights you use.
Stop the layout jumping
Layout shift is the one people notice and forgive least: they go to tap a button and the page moves. On AI-built sites it usually comes from things that arrive late and push content down.
- Images and videos without a set width and height, or an aspect ratio in CSS.
- Banners for cookies, offers or chat that slide in above the content. Put them over the page, not in its flow.
- Fonts that swap in at a different size. Pick a fallback font with similar widths.
- Content drawn by script after the first paint, such as a pricing table or logo strip that fills in a moment later. Reserve its space, or put it in the first HTML.
Testing: lab, field and a real phone
Use three kinds of test. Each one shows something the others miss.
| Test | What it is | Use it for |
|---|---|---|
| PageSpeed Insights, field data | Real-user data from the Chrome UX Report over the previous 28 days | How real visitors experience the page |
| PageSpeed Insights, lab data | A Lighthouse test in a simulated environment | Finding causes and checking a fix right away |
| A real mid-range phone | You, holding the phone, on mobile data | Layout, tap targets, forms and how it feels |
Google's PageSpeed Insights docs say that when a page has too few real-user samples, it shows data for the whole origin, and when the origin has too few, it cannot show field data at all. That is normal for a new site. Use the lab test and the phone until real data builds up.
On the phone, do these by hand: open the home page from a link in a chat app, scroll every page, open and close the menu, send every form and check the lead arrives, and turn the phone sideways once.
Decision rule If your LCP element is text or an image that only appears after JavaScript runs, image and script fixes will help but will not solve it. That is the same rendering problem as chapter 2, and chapter 5 decides it.


