Core Web Vitals in 2026: Performance Budgets That Actually Hold
Lab scores are easy to game. Field data is not. Here is how to build performance budgets that survive contact with a real roadmap.
Every performance project starts the same way: a synthetic audit, a burst of optimisation, a green score, and a slow regression over the following two quarters as features ship. The score was never the problem. The absence of an enforcement mechanism was.
Performance budgets fix that by turning speed from a periodic project into a constraint the delivery process respects automatically.
Measure field data, not lab scores
Lab tools are diagnostic instruments; they tell you why something is slow. They do not tell you whether users are experiencing it, because they use a fixed device, a fixed network, a cold cache, and no real interaction.
Field data — real user monitoring at the 75th percentile, segmented by device class, country, and template — is the source of truth. The segmentation matters more than the headline number: an aggregate LCP of 2.4 seconds routinely hides a mobile category page at 4.1 seconds that is the actual revenue problem.
The three metrics and what moves them
| Metric | Good (p75) | Dominant causes |
|---|---|---|
| LCP | Under 2.5s | Slow server response, unoptimised hero images, render-blocking CSS, late resource discovery |
| INP | Under 200ms | Long JavaScript tasks, heavy hydration, third-party listeners, expensive re-renders |
| CLS | Under 0.1 | Unsized media, injected banners and ads, late web fonts, dynamically inserted content |
INP is the metric most teams still underinvest in, because it only surfaces when users interact and therefore never appears in a synthetic run. It is also the metric most sensitive to framework choices and third-party tags.
LCP fixes in order of leverage
- Cut server response time — cache HTML at the edge; nothing else matters if TTFB is 800ms.
- Preload the LCP image and serve it in a modern format at correct dimensions.
- Remove render-blocking CSS and inline only what the first viewport needs.
- Never lazy-load the hero image; make it discoverable in the initial HTML.
INP fixes in order of leverage
- Break long tasks: yield to the main thread and defer non-critical work.
- Reduce hydration scope — hydrate interactive islands, not entire pages.
- Audit third-party scripts for main-thread listeners; most can be deferred or removed.
- Debounce expensive handlers and virtualise long lists.
Setting budgets that teams accept
Budgets fail when they are aspirational. Derive them from current field data and a target: if mobile LCP p75 is 3.4 seconds, set the budget at 3.2 with a quarterly step down, not at 2.0. Then express budgets in terms engineers control at build time — bytes and request counts — alongside the outcome metrics.
A workable starting set for a content or commerce site: 170 KB compressed JavaScript per route, 250 KB total image weight above the fold, no more than 40 requests before the largest paint, and no third-party script on the critical path.
Enforcement is the whole point
Wire the budget into CI so a pull request that exceeds it fails with a specific message naming the offending asset. Add a synthetic run per template on every deploy, and alert on field-data regressions weekly rather than waiting for a quarterly audit. Give teams an explicit override path with an owner and an expiry date — budgets without escape hatches get disabled entirely the first time they block a launch.
Third-party governance
Third-party scripts are the leading cause of regression because they are added by teams outside engineering and change behaviour without a deploy. Establish a tag review with three requirements: a named business owner, a measured performance cost, and a removal date to be revalidated. Load everything non-essential after interaction or via a worker, and audit the tag inventory quarterly — in most audits, 20 to 40 percent of tags are firing for campaigns that ended.
Connect performance to money
Performance work loses funding fights because it is argued in milliseconds. Convert it: segment conversion rate by LCP bucket in your own analytics, and present the delta. Internal data almost always shows a measurable conversion difference between fast and slow cohorts, and one honest internal chart outperforms any external benchmark in a prioritisation meeting.
Frequently asked questions
What is a performance budget?
An agreed, automatically enforced limit on metrics such as JavaScript bytes, image weight, request count, or field LCP and INP, so that changes exceeding the limit fail the build or require explicit approval.
What thresholds should we target?
Google's good thresholds at p75 are LCP under 2.5s, INP under 200ms, and CLS under 0.1. Competitive teams target LCP under 2.0s and INP under 130ms to retain headroom for future features.
Why does lab data disagree with field data?
Lab runs use fixed devices, networks, and cold caches with no real interaction. Field data reflects actual device and network distributions, consent banners, and third-party behaviour — always treat it as authoritative.
How long does improvement take to show in Search Console?
Field data reports use a 28-day rolling window, so expect four to six weeks before an improvement is fully reflected.