Skip to content
Interface language: English
Design / Signal

Website speed and Core Web Vitals: what to measure and which fixes come first

Assess LCP, INP and CLS, separate field data from lab tests, and plan improvements to images, fonts and JavaScript.

Article language: English

A woman watches a glass sphere on a red cord beside water channels in a coastal garden.
In this article

A fast page shows important content, responds to a click and does not move a different button under someone's finger at the last moment. That is the experience worth measuring. One green score cannot tell the whole story, especially if the wait for a contact form is long enough to cool both coffee and the desire to enquire. Check the measurement conditions, find the cause of delay and repeat the test after a change.

Core Web Vitals are Google's measures of loading, responsiveness and stability. They do not replace checking a form, the clarity of an offer or accessibility. The procedure below helps a website owner commission and accept a meaningful optimisation rather than buy a promise of “100 points” without knowing the test conditions.

Understand LCP, INP and CLS

web.dev's documentation describes three core measures: LCP concerns when the largest eligible image or text block appears in the viewport; INP concerns responsiveness to interactions; and CLS concerns unexpected layout shifts. The recommended good thresholds are at most 2.5 seconds, 200 milliseconds and 0.1 respectively. CLS is not a duration and is not expressed in seconds.

Assessment uses the 75th percentile: the value at or below which 75% of measurements fall, considering mobile and desktop separately. It is not the single best attempt. Nor is it an unconditional guarantee of every person's experience. Networks, devices and usage still differ.

These are thresholds in Google's documentation, not numbers invented by a supplier. An image-size or initial-JavaScript budget, however, should follow the particular project. There is no universal permissible photograph weight that automatically proves good performance on every B2B page.

Separate user data from a controlled experiment

Field data describes real visits when a sufficient sample is available. A lab test runs a page in fixed conditions and helps reproduce a problem. These views can differ. Missing field data means neither a pass nor a failure; describe the measurement's limitation.

Record the URL, date, page version, tool and test conditions. Note whether a result concerns one URL or a broader set of data. Do not compare the home page before a change with one article afterwards: they are different views, with different resources and potentially different experiences.

Choose representative examples: home, service, a long article and contact. Test direct entry by URL for each. Website acceptance (in Polish) should cover the whole user task. A screen that loads quickly but cannot send an enquiry remains an unfinished release.

Find the actual LCP element

Do not assume a photograph is the main obstacle merely because it occupies a lot of space. The largest element may be text, while delay may come from the server response, late resource discovery or waiting for a script. Inspect the network trace and rendering time before choosing a fix.

The LCP optimisation guide says, among other things, that an image serving as the LCP element should not use deferred loading, called lazy loading. An important resource needs early discovery. Raising the priority of one key image can be justified; assigning the highest priority to an entire gallery defeats the hint's purpose.

Consider a working scenario: the first slide appears only after JavaScript fetches configuration. Shrinking its file may not remove the earlier wait. The supplier's task should therefore be “identify and reduce the image discovery delay”, not merely “compress the photo”. Compare the same path after the change.

Fit images to where they appear

Keep the originals and prepare variants at the required dimensions for the site. The browser can choose between them using srcset and an accurate description of the layout slot size. Naming a file “mobile” is not enough if the site still downloads the full large variant in every view.

Inspect the downloaded file in network tools. Also assess quality after compression: hair, fine fabric and text in a real document can deteriorate sooner than a calm surface. Format choice belongs in the test. A WebP extension alone does not prove appropriate dimensions or reasonable transfer size.

Brand photography on a website (in Polish) addresses direction, crops and image selection. Performance should not remove an image's meaning. If a panoramic scene needs a separate mobile composition, plan that variant instead of merely reducing resolution until the subject is no longer visible.

Aeneia conceptual photography

Limit layout shifts

web.dev describes reserving image space with correct dimensions or an aspect ratio. Text then need not move only after the image downloads. Check elements inserted later too: embeds, banners and large modules that change the height of content already being read.

Do not assess stability only during the first second. Scroll an article, open the menu and inspect material loaded further down. A change after a deliberate user action and an unexpected jump while reading are different situations. Report the particular movement, moment and cause rather than calling the whole screen “unstable”.

Record font behaviour too. A fallback with different metrics can change line count and button position. Loading strategy and a fallback-font test belong to premium website typography (in Polish). Removing one heavy image will not fix a problem caused by text.

Test interaction response, not just page startup

Fast loading does not automatically mean a fast response to a click. Open the menu, switch a tab and type a search query. Watch the pause between action and visible response. The INP optimisation guide distinguishes delays associated with event handling and rendering.

The supplier should identify what occupies the browser's main thread, which runs code and updates the view among other work. It could be a large JavaScript task, an expensive view update or work unrelated to the current action. An animated icon does not shorten that work. It can signal waiting if the message reflects the real state. If a 3D scene causes substantial load, also compare a complete path without WebGL (in Polish), preserving the information and the user's main task.

For micro-interactions (in Polish), test rapid repeated action and navigation during a transition. The effect should not block the next decision or leave an overlay on screen. Connect a measurement to an observable task in the report: “the menu responds after a click” is more useful than an animation library's score alone.

Set the order and acceptance conditions

First remove the problem blocking main content or an action. Then address unnecessarily downloaded resources and view stability. Optimising a small icon before fixing a multi-second wait for the opening image may be technically correct but have little effect on the user's journey.

Each task needs a diagnosis, planned change, risk and repeat-measurement method. After reducing scripts, check that search and the form still work. After changing a font, check characters and typesetting in every supported language, including Polish diacritics. After moving image loading, test the scrolled view, not just the first photograph.

Do not promise search rankings from a performance score. Google describes page experience as a broader whole. Speed works with content, accessibility and usability. Conduct technical SEO acceptance separately using the pre-publication check (in Polish).

Avoid comparing random runs

Agree whether you are testing a first visit with an empty cache or a return visit. Keep the same browser version and network throttling settings. Run several trials and explain an outlier instead of selecting the prettiest result. If results differ, also record whether other resources were downloading in the background.

Before completing the work, repeat the key scenario on a real device. A controlled test helps find the cause; ordinary use shows whether the fix harmed content or functions. Accept both kinds of evidence. Closure can then follow a described procedure instead of one green indicator. If lab results do not explain the team's difficulty, supplement the diagnosis with observation of website use at work (in Polish), describing the surrounding conditions.

How video is embedded (in Polish) can strongly affect page behaviour. Check what loads before playback and what a customer receives on a weak connection. A smaller video file alone will not solve the cost of the whole player.

What belongs in the report?

The report should support a decision and a repeatable check. Record the starting point, the changed element and the result on the same URLs. If field measurement does not yet reflect the new version, say so. A post-fix lab result is useful evidence but does not replace later observation of real visits.

  • URLs and measurement conditions are specified.
  • Field and lab data remain separate.
  • The LCP element and its delay have been identified.
  • Image and font variants have been checked.
  • Main interactions were tested after the change.
  • Fixes have an owner and evidence of renewed acceptance.

Agree the next measurement with the site owner after adding a large gallery, an analytics tool or another integration. An acceptance result describes a particular version, not all future changes. A small project can use a fixed URL list and short comparison record. What matters is a clear baseline for deterioration and a person who decides on a fix.

Want to set an order for performance work? See the Design scope, then describe the situation through contact.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear