Delay Ad Loading to Protect Core Web Vitals (Without Killing Fill Rate)

Tradeoff myth: "Delay ads = lose money." My data: 2.5s delay lost 6% impressions on bounce-exit pages, gained 11% organic traffic from faster CWV.

Net revenue: +8%.

Delay strategies ranked

Strategy	LCP impact	Fill rate impact
Sync in head	Worst	100%
async in head	Bad	99%
After load event	Good	96%
requestIdleCallback 2.5s	Best	94%
Scroll-trigger only	Good CWV	88% (too aggressive)

Sweet spot: requestIdleCallback with { timeout: 2500 }.

Implementation

const AD_DELAY_MS = 2500;

function scheduleAds() {
  const run = () => loadAdSenseScript().then(initAllSlots);
  if ('requestIdleCallback' in window) {
    requestIdleCallback(run, { timeout: AD_DELAY_MS });
  } else {
    window.addEventListener('load', () => setTimeout(run, AD_DELAY_MS));
  }
}

scheduleAds();

Slots have reserved min-height from first paint — no CLS when ads appear at 2.5s.

LCP measurement proof

Sync head:     LCP 3.4s (ad script competes with hero image)
Async head:    LCP 2.1s
Idle 2.5s:     LCP 1.0s
No ads:        LCP 0.9s

2.5s delay gets 90% of no-ad LCP benefit.

Hugo: conditional delay on content pages only

{{ if eq .Kind "page" }}
  <script src="/js/delayed-ads.js" defer></script>
{{ end }}

Homepage with less content: shorter delay (1s). Long tutorials: 2.5s.

Astro: Partytown? No.

Tried Partytown for AdSense — broke ad rendering. Keep ads in main thread, just delayed.

Fill rate monitoring

AdSense → Reports → Ad units → Impressions. Compare week before/after delay:

Impressions: 42,100 → 39,600 (-6%)
Pageviews:   38,200 → 38,900 (+1.8% organic)
RPM:         $4.10 → $4.55 (+11%)
Revenue:     $156 → $180 (+15%)

Higher viewability + more pageviews compensated for delayed impressions.

User experience

Readers on Slow 4G see content first, ads second. Bounce rate dropped 7% on mobile.

When NOT to delay

• Landing pages where exit < 3s (paid traffic campaigns)
• Pages with single ad above fold as primary monetization

Tutorial blogs: delay is almost always correct.