Meta Ads ROAS Decline—It’s Not Creative Fatigue, It’s Messaging Stagnation

The branding and marketing stories from Soulpapamarketing.

Why Brands Get Creative Refresh Timing Wrong — The Causality Is Reversed

The industry standard says: “As impression frequency rises, creatives get fatigued. Refresh them periodically.” Not wrong. But the causality is flipped.

Rising frequency is the result, not the cause. The actual sequence is:

1. New creative uploaded → Andromeda allocates exploration budget to that creative
2. Highest-purchase-intent audience segments get exhausted first
3. Remaining audience CVR drops
4. Andromeda rates that creative’s auction competitiveness lower → CPM rises
5. Frequency rises as a byproduct of this cycle

In other words, the real purpose of creative refresh is not “fresh perspective” but reaching audience segments that haven’t been exhausted yet. Timing refreshes by frequency or days is always one beat too late.

VCG Auction Structure: CVR Should Determine ROAS by Design

Meta’s ad auction isn’t a simple highest-bidder system. Based on VCG (Vickrey-Clarke-Groves) theory, the formula determining bid value is:

Total Value = Bid × Estimated Action Rate + Ad Quality

Even with equal bids, ads with higher Estimated Action Rate (CVR) achieve higher Total Value. In VCG, higher Total Value leads to lower winning bid. Ads with higher CVR secure more impressions on the same budget, and Andromeda automatically searches for high-purchase-intent audiences that match that creative. This is VCG’s design intent. CVR is the theoretical basis for determining ROAS.

But one real-world tradeoff emerges here. Audiences assigned to high-CVR creatives are premium segments that competitors also want. Competition is fierce, so CPM rises. The VCG discount effect and premium audience cost cancel each other out. Triple Whale’s 2025 data analyzing 35,000 brands shows this structure: on years with +8.3% CVR, CPM also rose +20%.

Why You Don’t Feel ROAS Improvement Even When You Raise CVR

If VCG structure is sound, CVR improvements should show up as ROAS gains. Yet many advertisers don’t directly observe this effect. Two reasons why:

First, VCG effects are structurally impossible to directly observe. We ran Pearson correlation analysis on CVR vs. CPM using real account data (10 KRW accounts, weekly for 14 weeks) to see if CVR improvement appears as CPM decline. The result: no statistically significant relationship between CVR and CPM (r ≈ 0, p > 0.05). Budget changes, audience exhaustion rate, competitor bid intensity, seasonality, and premium audience auto-selection effects all hit CPM simultaneously. VCG operates at the individual auction (impression) level, but when aggregated to weekly metrics across hundreds of thousands of impressions, that signal drowns in other variables. Your CPM doesn’t drop when you raise CVR—this isn’t system failure. The structure was designed so you can’t observe it directly.

Second, brands with stagnant messaging can’t regenerate CVR signal itself. When Andromeda treats new creatives as identical to existing ones, CVR exploration never begins. The VCG premise of “new CVR signal → new audience search → auction competitiveness recovery” never happens. This is why ROAS doesn’t improve even when you refresh creatives.

This is why obsessing over one or two surface metrics is dangerous. No matter how precisely you track CVR%, CPM, and frequency, if those metrics are corrupted by noise, your judgment will be wrong.

Creative Similarity 60% — Andromeda’s Baseline for Messaging Evolution

Andromeda, officially launched in Meta Engineering Blog December 2024 and rolled out to all accounts in October 2025, completely replaced the ad retrieval engine. The retrieval stage model complexity increased 10,000x compared to the previous system and runs on NVIDIA Grace Hopper Superchip infrastructure.

Andromeda reads creatives—pixels, video frames, text, audio—and converts them to dense feature vectors (embeddings). When the embedding similarity between different creatives is 60% or higher, they’re treated as the same ad. Brands with stagnant messaging can’t pass this threshold no matter what creative they make—variations of the same messaging won’t clear it.

60% is not a wall to break through. Proper messaging evolution naturally clears this baseline. Conversely, using completely different messaging—clickbait, sensationalism, careless expansion into messaging unrelated to brand direction—is recognized as new by Andromeda, but without CVR history, you have to reprove your audience from scratch. You abandon the CVR assets your existing messaging built.

Real Data: In Accounts Where VCG Design Works, CVR-ROAS Is r=0.97

In accounts where messaging evolves, the VCG structure actually works. Analyzing 14 weeks of weekly CVR and ROAS for a single ad account (minimizing variables), Pearson correlation came in at r=0.97. CVR alone explains 94% of ROAS variation.

One counterargument needs addressing: ROAS = CVR × AOV / CPC, so higher CVR naturally means higher ROAS. Formulaically yes. But in reality, CPM, CTR, and AOV move independently of CVR. If CPM rises faster than CVR when you improve CVR, ROAS falls. Low-AOV brands have structurally low ROAS even with high CVR. Seasonality, budget changes, and competitor bid intensity simultaneously shake ROAS. For r=0.97 to emerge amid all this noise isn’t just formula reflection. It’s proof that CVR and ROAS have a genuine relationship at the VCG bid foundation. The signal survives the noise.

The difference in this account isn’t CVR level itself. It’s that messaging evolved periodically, letting Andromeda find new audience segments every time. CVR signal regenerated continuously. VCG effects accumulated.

The key insight: absolute CVR level matters less than the CVR × AOV combination determining ROAS. Low CVR + high AOV keeps ROAS high. From ROAS = CVR × AOV / CPC, creatives only affect CVR. If CVR is normal (4~6%) but ROAS is low, the problem isn’t creatives—it’s low AOV.

Creative Lifespan ROAS Progression — 90-Day Real Account Measurement

We measured CVR and ROAS changes by week after creative upload:

• W1 (Days 0-7): CVR 8.89% → ROAS 7.33x
• W2 (Days 8-14): CVR 8.89% → ROAS 4.14x
• W3 (Days 15-21): CVR 6.52% → ROAS 3.38x — Andromeda algorithm transition period, temporary CPM drop
• W4 (Days 22-28): CVR 6.56% → ROAS 4.67x — New audience search complete, partial recovery
• W5-6 (Days 29-42): CVR 5.11% → ROAS 2.95x ← ROAS 3x threshold break begins
• W7+ (Day 43+): CVR 3.5~4.4% → ROAS 2.3~2.7x

W3 shows a period where CVR drops and CPM falls together. Andromeda exhausts the initial-response audience and transitions to exploring new segments. Mistaking this for fatigue and turning off the creative means missing W4’s recovery opportunity. Watch CVR progression, not frequency or days.

The Essential Work Aligned with Andromeda’s Design Intent

If we can’t directly feel VCG effects through CPM, what can advertisers do?

If Andromeda is already a sophisticated system, then do work aligned with that system’s design intent. Andromeda was built to read creatives and connect them with high-purchase-intent audiences. This design works best when creatives are (meaningfully) different—when messaging actually evolves.

CVR is indeed the key directional metric. But clinging to CVR alone is a trap. CVR depends on traffic volume—CTR and CPC.

If ROAS = CTR × CVR × AOV × 1,000 / CPM, then high CVR alone doesn’t work if CTR is low or CPM rises faster. ROAS stays put. But when messaging evolves, the three numerator terms (CTR, CVR, AOV) move favorably in concert, and the denominator (CPM) gets relatively suppressed as Creative Similarity drops.

Messaging evolves → Δembedding ≥ messaging evolution threshold
→ Andromeda searches new audiences
→ CTR↑ × CVR↑ × AOV / CPM(relative ↓)
→ ROAS↑

This is the simplest, clearest direction when you extend thinking through VCG and Andromeda foundations. Watch CVR as a metric, but see the root moving that metric—evolve your messaging.

That evolution must come from branding perspective. Clickbait, sensationalism, variations designed to trick the algorithm won’t work against Andromeda’s embedding analysis. While they might look like they lower Creative Similarity, they don’t build CVR signal. Careless expansion into messaging unrelated to brand direction has the same problem. CVR history resets; you abandon assets your existing messaging built.

Messaging evolution means maintaining direction while understanding your consumer’s problems more concretely and deeply. As that understanding deepens, expression naturally changes. As expression changes, embedding changes, and Andromeda finds audiences you haven’t reached. This is what VCG design demands of advertisers.

A brand whose messaging stagnates is not genuine. How could a brand actively bringing consumers to life have stagnant messaging?

Meta Ad Creative Refresh Guide by CVR — Threshold by Brand Type

CPM or day-based refresh carries too much noise. Consider refresh when the 14-day rolling average CVR falls below threshold.

If you have a creative group where Creative Similarity Score is 60%+, clean up similar creatives before deploying new ones. More similar creatives narrow Andromeda’s exploration scope and accelerate exhaustion.

How to Calculate CVR-ROAS Correlation Yourself

You can directly verify if VCG structure is operating in your account. You’ll need at least 12+ weeks of weekly data.

Data to gather: Meta Ads Manager → Custom Date Range → Group by Week → Export CVR (purchases/link clicks or purchases/impressions), ROAS

Excel / Google Sheets

=CORREL(CVR_range, ROAS_range)
Example: =CORREL(B2:B14, C2:C14)

Python

from scipy.stats import pearsonr
r, p_value = pearsonr(df['cvr'], df['roas'])
print(f"r = {r:.3f}, p = {p_value:.4f}")
# p < 0.05 is statistically significant

Interpreting results: r ≥ 0.7 means CVR drives ROAS. Low r suggests one of two things: you're in a high-noise period (major budget swings, seasonality), or messaging stagnation means CVR signal isn't regenerating. If the latter, check Creative Similarity first.

What It Really Means to Do Advertising Well

One last point: the reason for laying out VCG auction structure, Andromeda embeddings, Pearson correlations, and ROAS decomposition isn't so you'll run these analyses yourself. It's to technically explain and convince why you must focus on the essential.

Actually running this analysis is pointless except for people like us. Analyze, experience, dig deeper, and one conclusion emerges: focus more on the essential. Advanced technique already proves technique is meaningless. Andromeda got 10,000x more complex to read more precisely how deeply a brand truly understands its consumer.

A brand whose messaging stagnates is not genuine. How could a brand actively bringing consumers to life have stagnant messaging?

References

1. Meta Engineering Blog (2024.12.02). Andromeda: The Next-Gen Personalized Ads Retrieval Engine. engineering.fb.com — Official announcement of 10,000x increase in retrieval stage model complexity, 8% improvement in ad quality.
2. Triple Whale (2025). 2025 D2C Advertising Benchmarks. Analysis of 35,000 ad accounts — CVR +8.3% YoY, CPM +20.0% YoY rising simultaneously.
3. Confect / Aurelius Media (2026). Meta Andromeda Impact Analysis. 3,000+ eCommerce advertisers, $834M spend analyzed — ROAS -7%, CVR -17% observed during Andromeda rollout.
4. Motion (2026). 2026 Creative Benchmarks Report. $1.3B ad spend data — confirmed shorter creative lifespan, refresh cycle compressed from monthly to weekly.

Frequently Asked Questions

Insights from Soulpapa Marketing — Korea's digital marketing agency.
Original Korean article: https://soulpapa.co.kr/2026/05/23/meta-ads-roas-drop-creative-similarity-andromeda/