evals.sh
ProductModelsSecurityPricingDocsBlog
Sign inGet started
Blog

Guides

What a frontend audit actually measures

Core Web Vitals, accessibility, SEO and security headers: what each metric means, and which fixes move the score most.

Evals.sh Engineering·October 23, 2026·9 min read

A frontend audit compresses a lot of measurement into a few numbers. Those numbers are useful, but only if you know what sits underneath them. This is a field guide to what an audit actually checks, what the thresholds mean, and which fixes tend to change the result.

What gets measured

When you run a frontend audit in Evals.sh, you give it a public URL and choose which checks to run: Performance, Accessibility and Security. The audit loads the page, collects measurements for each selected area, captures screenshots, and adds written insights for each category explaining what it found and what to change. The results roll up into a single evaluation score that combines performance, SEO, security headers and accessibility violations, and the full report can be exported as a PDF.

You can also add instructions, such as “focus on the mobile homepage” or “explain the performance findings for a junior developer”, which shape the written insights without changing the underlying measurements.

CategoryWhat the report shows
PerformancePerformance score, SEO score, and the lab metrics LCP, FCP, CLS and TBT.
AccessibilityTotal violation count, then each rule that failed with its impact level, a description and how many elements it affects.
SecurityWhether HTTPS, Content-Security-Policy, X-Frame-Options and HSTS are present, plus a list of missing headers.
InsightsWritten findings for each category, with concrete suggestions tied to what was measured.

Core Web Vitals

Core Web Vitals are Google’s published set of user-centred performance metrics. Each one has a public “good” threshold, and they are worth knowing by heart.

Largest Contentful Paint (LCP) measures when the largest visible element, usually a hero image or a headline block, finishes rendering. Good is 2.5 seconds or less; anything over 4 seconds is poor. LCP is the closest single number to “when did this page look loaded?”

Cumulative Layout Shift (CLS) measures how much visible content moves around unexpectedly while the page is in use. It is a unitless score: 0.1 or less is good, above 0.25 is poor. The classic cause is an image or ad slot without reserved dimensions that pushes text down just as someone goes to tap it.

Interaction to Next Paint (INP) measures how quickly the page responds visually after a click, tap or key press, across the whole visit. Good is 200 milliseconds or less; above 500 is poor. INP replaced First Input Delay as a Core Web Vital in 2024.

Two supporting metrics round out the picture. First Contentful Paint (FCP) is when anything first appears; it is good at 1.8 seconds or less. Time to First Byte (TTFB)is how long the server takes to start responding. It isn’t a Core Web Vital, but a slow TTFB delays every metric after it, so it is often the first thing to look at when LCP is bad across the board.

THRESHOLDSlower is betterLCPloadinggoodneeds improvementpoor2.5s4.0sCLSvisual stabilitygoodneeds improvementpoor0.10.25INPresponsiveness · fieldgoodneeds improvementpoor200ms500msTBTmain thread · labgoodneeds improvementpoor200ms600msVitals are judged at the 75th percentile of page loads. TBT uses Lighthouse lab bands.
Published thresholds for the main loading, stability and responsiveness metrics. Bands are not drawn to scale; what matters is which side of each boundary you land on.

Lab numbers vs. field numbers

An audit loads your page once, under controlled conditions. That is a lab measurement. Google’s Core Web Vitals assessment uses field data: real visits from real devices, judged at the 75th percentile. The two can disagree, and both are useful. Lab data is repeatable, so it is the right tool for catching regressions and comparing submissions fairly. Field data tells you what your actual users experience.

INP can only be measured from real interactions, so a lab audit reports Total Blocking Time (TBT) instead. TBT adds up how long the main thread was blocked by long tasks during load. In Lighthouse scoring, under 200 milliseconds is good and over 600 is poor. A page with high TBT will almost always have a poor INP in the field, because the same long tasks get in the way of responding to input.

SEO basics

The SEO score is not a ranking prediction. It checks the technical basics that let search engines read and index a page: a descriptive <title>, a meta description, a valid viewport tag, crawlable links with real href attributes, images with alt text, a page that isn’t accidentally blocked from indexing, and legible font sizes on mobile. These are cheap to fix and tend to be all-or-nothing, so a low SEO score usually points to two or three specific omissions.

Accessibility

Automated accessibility checks test a page against rules derived from the Web Content Accessibility Guidelines (WCAG). The report lists each failing rule with an impact level (critical, serious, moderate or minor) and how many elements it affects. The most common findings fall into a few groups:

  • Colour contrast. WCAG AA asks for a contrast ratio of at least 4.5:1 for body text and 3:1 for large text. Light grey on white and muted text on dark backgrounds are the usual offenders.
  • Labels. Every form input needs an accessible name, and icon-only buttons need a text label (an aria-label or visually hidden text) so a screen reader can announce them.
  • Semantics. Use real <button>, <a>, <nav> and heading elements in a logical order. A clickable <div> looks fine to a mouse user and is invisible to a keyboard user.
  • Document basics. A lang attribute on the page, unique IDs, and alt text that describes the image rather than repeating the filename.

What automation can’t tell you

Automated rules catch a meaningful share of accessibility problems, but not all of them. They can confirm an image has alt text, not that the alt text is accurate. A clean report is a good floor; a quick keyboard-only pass through the main flow is still worth doing.

Security headers

Security headers are instructions the server sends with every response, telling the browser how to treat the page. They cost almost nothing to add and close off whole classes of attack.

  • HTTPS. The baseline. Everything else depends on the connection being encrypted.
  • Strict-Transport-Security (HSTS). Tells the browser to use HTTPS for this site from now on, even if someone types or links http://. It prevents downgrade attacks on later visits.
  • Content-Security-Policy (CSP). Declares which sources scripts, styles, images and frames may load from. A good CSP turns many cross-site scripting bugs from exploitable into blocked. It takes the most care to configure, so start in report-only mode and tighten from there.
  • X-Frame-Options / frame-ancestors. Controls whether other sites can embed your page in a frame, which is how clickjacking works. The modern form is the CSP frame-ancestors directive; X-Frame-Options: DENY or SAMEORIGIN covers older browsers.
  • Referrer-Policy. Limits how much of the current URL is sent to other sites when a user follows a link. strict-origin-when-cross-origin is a sensible default and keeps query strings with tokens or IDs from leaking.

Fixes that move the score

Most audit reports list dozens of opportunities. In practice, a handful of changes account for most of the improvement on a typical site.

FixWhat it improvesHow
Size and prioritise the hero imageLCPServe a correctly sized, modern-format image; mark it high priority and don’t lazy-load it.
Reserve layout spaceCLSSet width and height (or aspect-ratio) on images, embeds and ad slots; avoid inserting banners above existing content.
Load fonts deliberatelyLCP, CLSPreload the one or two fonts used above the fold, use font-display: swap, and pick fallbacks with similar metrics.
Split and defer JavaScriptTBT, INPCode-split by route, defer non-critical scripts, and move heavy third-party tags off the critical path.
Fix the server responseTTFB, LCPCache rendered pages or static assets close to users, and avoid slow work before the first byte.
Add the missing headersSecuritySet HSTS, Referrer-Policy and frame-ancestors at the server or edge; roll out CSP in report-only mode first.

Work through them in roughly that order, re-running the audit after each change. Because the audit is a lab measurement, results can move a little between runs; look for consistent shifts rather than reacting to a single number. And if you are grading a cohort, run every submission through the same audit so the comparison is fair. The score is a summary. The value is in the list of specific things to fix.

On this page

What gets measuredCore Web VitalsLab numbers vs. field numbersSEO basicsAccessibilitySecurity headersFixes that move the score

Try it on your own project

Free credits every month. No card required.

Get started

Keep reading

Guides

Writing rubrics an AI grader can apply consistently

October 23, 2026

Guides

Grading a full cohort in one session with Groups

April 2, 2026

Guides

Choosing a project evaluation tool in 2026: a buyer's framework

April 2, 2026

evals.sh

Autonomous evaluation for software teams and educators. Agents that review code, drive real browsers and grade work against your rubric.

Product

  • Project evaluation
  • Code review
  • Frontend audit
  • API testing
  • Assignment grading

Platform

  • Models
  • Security
  • Pricing
  • Documentation

Resources

  • Blog
  • Guides
  • Contact

Legal

  • Privacy policy
  • Terms of service
  • Data handling

© 2026 Evals.sh. All rights reserved.

Customer data is never used to train AI models.