Mobile CWV regression still active — 1,374 poor URLs, PSI lab scores 11–43/100 across all 24 tracked pages
Across the 30 Apr – 18 Jun window, $452,680 was spent on Google + Meta paid landing pages, generating $840,924 in GA4 revenue on paid-tier pages (blended ROAS 1.86x). 68% of all paid spend ($309,799) sits on just 11 HIGH-priority URLs. PSI monitoring has expanded from 5 to 24 URLs this period — every one of them still scores "Poor" on lab metrics, but field LCP and field INP tell a more nuanced story (see CWV Status tab). The /promo/new-customer tracking gap has worsened, not improved: GA4 now shows just 2 purchases against $109K combined spend.
Early signal that fix work is taking hold — Good mobile URLs up 24% (37 → 46)
While the Poor URL count keeps climbing overall, the Good URL count has started moving in the right direction for the first time since the May 1 regression. This lines up with engineering's recent page-level optimisation work: removing hidden elements that were slowing page load, compressing/optimising images, and converting the embedded product video to a lightbox pattern (so it no longer loads and decodes inline on first paint). See the GSC Mobile Timeline tab for the full trend and the Fix Timeline tab for how this slots into the phased plan.
Root cause found — Meta Ads is being misattributed as "Organic Social" in GA4
The $260K spend gap flagged in the last pass turned out to be a known Meta-side tagging/UTM issue, not missing data: GA4 is recording Meta's paid traffic under the "Organic Social" channel instead of "Paid Social". Re-running the channel-spend allocation with this corrected, $290,399 of Meta spend across 26 URLs — including /promo/new-customer ($91.7K), /promo/aus/flash-sale ($33.8K), and /promo/lp/aus-mothersday ($25.6K) — now maps cleanly to Organic Social sessions and is flagged accordingly in the Priority URL List's channel breakdown. Only $63 across 1 URL remains genuinely unattributed. Until Meta's tagging is fixed, "Organic Social" session/revenue numbers on these URLs should be read as paid performance, not free organic reach.
Tier definitions used throughout this dashboard
HIGH = combined Google + Meta spend ≥ $10,000 in this window (11 URLs, directly exposed to the mobile CWV penalty). MEDIUM = combined spend $250–$10,000 (53 URLs). LOW = organic-only, zero paid spend — 5,294 such URLs exist in the GA4 export; the Priority URL List tab shows the top 20 by revenue as a representative sample (164,247 organic sessions / $1.13M revenue collectively across all 5,294 — long tail, out of scope for paid CWV prioritisation). 25 additional URLs with under $250 spend ($1,706 total) are excluded as noise.
1,374 poor mobile URLs as of 18 Jun — but Good URLs are rising for the first time
Poor mobile URLs have more than doubled since mid-March (625 → 1,374). Needs-improvement URLs sit at 430. The headline number to watch, though, is Good URLs: 37 → 46 (+24%) in the last few weeks — the first sustained improvement since the May 1 regression. The chart below maps all three GSC categories against campaign and engineering events.
| Date | Event | Markets | Data Signal | Implication |
|---|---|---|---|---|
| Mar 1 | US (Southern) Meta Advantage+ live | US | New Meta campaign structure | Baseline CPM established for US before regression |
| Apr 23 | Desktop CWV fixed ✓ | All | CPA stable. Desktop QS recovering. | Fix applied desktop only — mobile branch missed. Desktop poor URLs begin declining. |
| Apr 30 | NCA live on Demand Gen & Meta Advantage+ | AU | $109K combined spend in this window | Launched into the regression window — GA4 tracking gap has worsened (71 → 2 purchases) over this period |
| May 1 | Code regression deployed 🚨 | All | CPA calm. Mobile LCP/TBT/CLS degrade silently. | 28 days of clean CrUX data buffers the impact initially. |
| May 29 | CrUX 28-day window fully degraded | All | Rolling average now mostly post-regression data | Quality Score pressure builds across all landing pages. |
| Early Jun | Engineering begins page-level optimisation | All | Good URLs begin climbing (37 → 46) | Hidden elements removed, images optimised, embedded video → lightbox. Date approximate — first measurable effect visible in GSC from this point. |
| Jun 18 | Current checkpoint | All | 1,374 Poor · 430 Needs Improvement · 46 Good | 24 URLs now under active PSI monitoring (up from 5) — see CWV Status tab. |
How to read this table
Google / Meta columns are the authoritative top-level ad-platform spend split (pulled directly from the Google Ads and Meta Ads exports) — CPA / AOV / ROAS / CPS are blended across both using GA4 purchases and revenue. Click any row to expand the GA4 channel-grouping breakdown for that URL, with each platform's top-level spend allocated down to its GA4 channels pro-rata by session share. Known issue: Meta Ads traffic is currently being misattributed as "Organic Social" in GA4 instead of "Paid Social" — Meta spend is allocated against Organic Social sessions where this applies, and flagged with a purple "⚠ Meta, misattributed" badge so it's not mistaken for free organic reach. The breakdown's Total row is highlighted in light grey as the aggregate.
| # | Tier▾ | URL Path▾ | Total Spend▾ | Google▾ | Meta▾ | Sessions▾ | Eng. Rate▾ | Bounce▾ | Purchases▾ | Revenue▾ | CVR%▾ | CPA▾ | AOV▾ | ROAS▾ | CPS▾ | CWV Mobile▾ |
|---|
All 24 tracked URLs score "Poor" on lab metrics (11–43/100) — but field data is more mixed
PSI monitoring has expanded from 5 URLs to 24 since the last capture. Every single one still lands in PSI's "Poor" lab bucket (score <50) — lab LCP averages a catastrophic 15.5s (range 1.1s–31.1s) and lab Total Blocking Time averages over 2,000ms (one page hits 15,530ms), both consistent with severe main-thread blocking under PSI's simulated mobile conditions. Field data tells a different story for some pages: field INP is "Good" (<200ms) on 20 of 24 URLs, meaning real users aren't experiencing anywhere near the lab-reported blocking. Field LCP is more concerning — only 2 of 23 measured URLs are field-Good, 5 are Needs Improvement, and 16 are field-Poor (≥4s) — this is a wider spread of field degradation than the last capture, where only isolated pages showed poor field LCP.
Every tracked page is in the Poor lab bucket. Google's Ad Quality Score reads this signal; CPC bids stay inflated while this persists.
2.5–4s field LCP. Closest to crossing into Good — prioritise these for quick wins alongside the optimisation work already in motion.
Real-user interaction responsiveness is mostly fine despite catastrophic lab TBT — reinforces that the lab/field gap, not raw user pain, is the main diagnostic puzzle.
| URL▾ | Lab Score▾ | Lab LCP▾ | Lab TBT▾ | Lab CLS▾ | Field LCP▾ | Field INP▾ | Field CLS▾ | Ad Spend▾ | ROAS▾ |
|---|
A group of user interactions with the site within a given time window (resets after 30 minutes of inactivity, or at midnight). One visitor can generate multiple sessions.
A session lasting 10+ seconds, OR with a conversion event, OR with 2+ pageviews/screenviews. This is GA4's bar for "the visitor actually did something."
Share of sessions that met the engaged-session bar above. Higher is better — >60% is generally considered strong for a landing page.
Share of sessions with no engagement at all (under 10s, no conversion, single pageview). >60% is high/concerning, <40% is strong.
Average length of a session in seconds, across all sessions landing on that URL.
Count of completed GA4 purchase events attributed to sessions that landed on this URL.
Total transaction revenue from those purchase events, in AUD.
Conversion rate. >3% is strong for a paid landing page, <1% needs work.
GA4's own default attribution model — Paid Search, Cross-network, Organic Search, Direct, Paid Social, Display, etc. This is GA4's view of where the session came from, separate from the literal ad-platform billing data shown in the Google/Meta spend columns.
Cost per acquisition, blended across Google + Meta spend against GA4 purchases. Lower is better. Shown as "—" when there are zero purchases.
Average order value — how much each purchase is worth, on average, for sessions landing on this URL.
Return on ad spend. 1.0x = breakeven on spend vs revenue (not accounting for margin). This dashboard's overall paid-tier ROAS is 1.86x.
Cost per session — how expensive it is, on average, to bring one visitor to this URL via paid channels.
Time until the largest visible element (usually a hero image or headline) finishes rendering. Measures perceived load speed. Good <2.5s.
Lab-only metric: total time the main thread was blocked by long tasks, preventing the page from responding to input during load. Good <200ms.
Field-only metric: time from a real user's interaction (tap/click/keypress) to the next visual update on screen. Good <200ms. This is the field-data cousin of TBT.
Measures unexpected visual movement as the page loads — buttons or text jumping around. Good <0.1.
A 0–100 composite lab score combining LCP, TBT, CLS, Speed Index and First Contentful Paint under simulated throttled-mobile conditions. 0–49 = Poor, 50–89 = Needs Improvement, 90–100 = Good.
Lab = one simulated PageSpeed Insights/Lighthouse run under fixed network and device throttling — repeatable but synthetic. Field = real, anonymised Chrome user data aggregated over a rolling 28-day window (Chrome UX Report / CrUX) — what actual visitors actually experienced. The two can diverge sharply, as they do across most URLs in this dashboard.
Google's gTech team audited 3 key paid landing pages using real CrUX field data (75th percentile, June 2026). While the audit found LCP, CLS and INP are technically within "Good" thresholds on the AU mystery picnic page (LCP 0.8s, INP 79ms, CLS 0), the US mystery picnic and NCA pages show INP at 152–154ms — borderline Needs Improvement. The audit identified two primary root causes causing the PSI lab score gap (scores 12–47 despite good CrUX field values): long-running JavaScript tasks blocking the main thread, and unoptimised above-fold image delivery. The 19 Jun PSI re-capture across 24 URLs confirms this is systemic, not isolated: lab TBT now averages over 2,000ms site-wide (peak 15,530ms on one URL), while field INP remains Good on 20 of 24 pages — the lab/field gap the audit flagged is still the dominant explanatory pattern.
Engineering has already begun page-level optimisation: removing hidden DOM elements that were silently slowing load, compressing/optimising images, and converting the embedded product video to a lightbox pattern (deferring video load/decode until the user opts in). This is exactly the kind of "Phase 1" work outlined below, and it's already showing up in GSC — Good mobile URLs rose from 37 to 46 (+24%) over the same period. Keep going: the lab data below confirms there's substantially more main-thread blocking and image-delivery work left across the full 24-URL set.
✅ Immediate: Audit & Diagnosis — partially underway
PSI monitoring is now live on 24 URLs (up from 5). Initial page-level cleanup (hidden elements, image optimisation, video → lightbox) is already shipping and showing up in GSC's Good-URL count. Remaining work: confirm root cause per-template and extend the same treatment to the rest of the HIGH/MEDIUM tier.
- ✓ DonePSI automation expanded from 5 to 24 priority URLs
- ✓ In progressHidden/unused DOM elements removed on initial pages
- ✓ In progressImage optimisation pass underway
- ✓ In progressEmbedded video converted to lightbox (defers video weight off first paint)
- Still to do: run Chrome DevTools Performance trace on the 3 worst lab-TBT pages (scavenger-hunt-usa at 15,530ms, gift-aus at 3,130ms, things-to-do at 3,480ms) to confirm whether the long tasks are template-wide or page-specific
- Still to do: resolve the /promo/new-customer GA4 tracking gap (2 purchases on $109K spend) — this is a tracking/UX issue, not a CWV one, and is actively masking whether CWV fixes are helping that page at all
CrUX field data shows AU mystery picnic LCP is already in reasonable shape relative to lab (lab 9.9s vs field 2.4s) — the PSI lab score is driven by JavaScript execution blocking interactivity in the lab environment, not slow image delivery per se. Tools: Chrome DevTools, WebPageTest, PageSpeed Insights.
🔧 Phase 1: LCP Image Optimisation — Preload Above-Fold Images
Highest ROI fixes shippable without major refactors. Focus entirely on mobile above-fold content delivery. Google gTech specifically recommends preloading the LCP hero image as a priority fix for the NCA page.
- Preload LCP hero image: Add
<link rel="preload" as="image" href="/hero.webp" fetchpriority="high">in<head> - Responsive image preloading: Use
imagesrcsetandimagesizesattributes to preload only the mobile viewport version - Convert hero images: WebP/AVIF format, set explicit
widthandheightto prevent layout shift - Critical path CSS: Inline above-fold styles, defer non-critical stylesheets
- Fix layout shift sources: Several HIGH-tier pages now show lab CLS well above 0.8 (e.g. /promo/aus/mystery-picnic at 0.90) — identify the injected element (ad slot, dynamic banner, or late-loading content) causing the shift
<link
rel="preload"
as="image"
href="/hero-mobile.webp"
media="(max-width:600px)"
/>
</head>
Preloading the LCP image instructs the browser to fetch it sooner than it normally would. Use sparingly — when everything is prioritised, nothing is. Preload only the mobile viewport version of the hero using media queries. Do not preload multiple types of the same resource.
⚙️ Phase 2: Break Up Long Tasks + Optimise JS Execution
Google gTech identified a 445ms long task blocking the main thread on the mystery picnic pages — visible in Chrome DevTools as an XHR Load → Function Call → onSuccess chain. The 19 Jun PSI re-capture across all 24 tracked URLs shows this is far from an isolated case: lab TBT now ranges from 1,040ms to 15,530ms across the set, averaging over 2,000ms — every single tracked URL is well past the 600ms "Poor" threshold. This is the primary driver of the lab scores (11–43) sitting well below field reality (field INP is Good on 20 of 24 URLs).
- Profile main thread: Chrome DevTools Performance panel → record page load → look for red-flagged tasks >50ms (Google gTech found one at 445ms; the worst tracked URL now shows 15,530ms of blocking time in a single lab run)
- Break long tasks into chunks: Refactor tasks to run in <50ms segments — use
setTimeout()to defer sequential functions off the critical path - Move non-UI work to Web Workers: Data sorting, analytics processing, and pixel scripts that don't need DOM access should run on a background thread
- Minimise client-side data processing: Handle heavy operations server-side before sending to browser. For SSR, generate content at build time and avoid hydrating unused components
- Third-party script audit: Delay chat widgets, tag managers, and Meta/Google pixels until after page interactive — these are prime long-task candidates
- Defer React hydration: Do not hydrate components the user cannot yet see — beware of hydrating unused components blocking the main thread
const worker = new Worker("sort-worker.js");
worker.postMessage(dataToProcess);
// Main thread stays free
worker.addEventListener("message", (evt) => {
// Worker done — update UI
renderResults(evt.data);
});
function runInChunks(tasks) {
const task = tasks.shift();
if (!task) return;
task();
// Yield to browser between tasks
setTimeout(() => runInChunks(tasks), 0);
}
A 445ms long task was observed in the DevTools trace running as an XHR Load → Function Call → onSuccess chain. Break this into chunks under 50ms each. Move computational tasks that don't need DOM access (analytics, sorting, data transforms) to Web Workers using the comlink library. Handle data post-processing on the server before sending to the browser wherever possible.
🎯 Phase 3: Target 90+ PSI Score — Field Data Validation + Pipeline Gates
Close the lab/field gap fully. Ensure all high-spend templates meet Good threshold consistently in real CrUX data, then lock it in with automated pipeline checks to prevent any future regression.
- CrUX monitoring: Validate 75th-percentile field data improvement 28 days post-fix in Search Console. CrUX takes 28 days to fully re-average — check weekly.
- Above-fold layout restructure: Address 10–16% scroll depth (Clarity data). Surface booking CTA, key content, and social proof above the fold — most users never scroll past the partially-rendered hero
- A/B test mobile vs desktop layouts to validate CVR uplift from CWV improvements with a controlled experiment before full rollout
- Lighthouse CI in deployment pipeline: Use
lighthouse-cito gate new builds — fail deployments that don't meet the performance budget. This would have caught the May 1 regression before production. - web-vitals npm package: Instrument real user metrics and pipe to your analytics endpoint for ongoing field data visibility beyond Search Console
Scroll Depth Problem — Clarity (carried forward — not refreshed this cycle)
75% of mobile users cite page load speed as the #1 factor in their experience. Users form a first impression in just 50ms. A 0.1s speed improvement increases CVR by 8.4% for retail and 10.1% for travel sites (Google/Deloitte, 2019).