Vilvah · Creative case study · Meta Ads · Jan to Jul 2026

Increasing ROAS was not about
how the creatives felt.
It was about how they performed.

ROAS moved from 1.92x in January to 2.94x in July. Not one of the decisions behind that started with an opinion about a video. All 416 video creatives were measured at three fixed points, every verdict traces back to the number the video failed, and every recommendation below points at a specific second of a specific edit. This is what seven months of judging performance instead of feel looks like.

Video creatives only. $87,839 of static, GIF and carousel spend is excluded and explained below.
$176,965Video spend
5,520Purchases
2.27xBlended ROAS
63 of 290Creatives clearing all three gates
The method

The three numbers we judge every video on

Performance is not one score, it is three. A video loses people in stages, and each stage is a different job. Measuring all three separately means a weak creative tells you which second of the edit to change, instead of leaving you with a feeling that the video is not working.

Gate 1 · of everyone served

Thumb Stop Ratio

Target > 18%

Share of impressions that became a 3-second view. This is the hook: first frame, opening line, pattern break.

227 of 290 creatives clear it
Gate 2 · of everyone who stopped

3s → 50% view

Target > 20%

Share of 3-second viewers who reached halfway. This tests the middle: pacing, and whether the hook's promise gets paid off fast enough.

217 of 290 creatives clear it
Gate 3 · of everyone at halfway

50% → 100% view

Target > 40%

Share of half-watchers who finished. High numbers mean the offer and the end CTA are actually being seen.

118 of 290 creatives clear it

Vilvah's shape: the hook is not the problem. 227 of 290 creatives clear the 18% target and the account averages 27.9%. The constraint sits at the far end, where only 118 of 290 clear the 40% completion target. Read this deck as an ending problem, not an opening one.

Results

Month by month: spend, ROAS and retention

Video creatives with at least $100 spend. ROAS climbed steadily from 1.92x in January to 2.94x in July while cost per purchase fell from $33 to $26. The look of the work did not change to produce that. What changed is which creatives were kept, cut and re-edited, and every one of those calls came out of the three scores on these cards.

36 video creatives live

Jan 2026

$25,576spend778purchases1.92xROAS
Hook · 3s26.3%target 18%
3s → 50%37.0%target 20%
50% → 100%39.2%target 40%
33 video creatives live

Feb 2026

$17,983spend518purchases1.96xROAS
Hook · 3s29.5%target 18%
3s → 50%31.8%target 20%
50% → 100%38.0%target 40%
54 video creatives live

Mar 2026

$31,529spend970purchases2.32xROAS
Hook · 3s26.9%target 18%
3s → 50%30.8%target 20%
50% → 100%32.1%target 40%
44 video creatives live

Apr 2026

$19,229spend592purchases2.40xROAS
Hook · 3s26.1%target 18%
3s → 50%34.4%target 20%
50% → 100%42.1%target 40%
60 video creatives live

May 2026

$22,228spend594purchases2.04xROAS
Hook · 3s23.6%target 18%
3s → 50%35.3%target 20%
50% → 100%38.1%target 40%
50 video creatives live

Jun 2026

$21,503spend739purchases2.47xROAS
Hook · 3s32.1%target 18%
3s → 50%27.1%target 20%
50% → 100%36.5%target 40%
58 video creatives live

Jul 2026

$28,451spend1100purchases2.94xROAS
Hook · 3s31.1%target 18%
3s → 50%32.7%target 20%
50% → 100%44.0%target 40%
Interactive · pick a month

Where each video loses its audience

Each line is one creative falling away from 100% of impressions. The dashed black line is the benchmark path: 18% hook, then 20% of those to halfway, then 40% of those to the end. Any line sitting under it is losing people faster than it should, and the steepest segment is where the fix belongs.

The vertical scale is logarithmic, so a fall from 100% to 10% takes the same space as 10% to 1%. That keeps the later stages readable instead of squashing them into the floor.

Seven-month trend

Only one of the three numbers has failed consistently

Spend-weighted average across every video creative in the month. Switch metric to see which part of the funnel improved and which one did not.

The important caveat

Retention scores do not predict ROAS at Vilvah

I split the 174 creatives with at least $300 spend into four equal groups by each metric, then measured ROAS in each group. If a benchmark drove revenue, ROAS would climb left to right.

Use these numbers to diagnose creative, not to allocate budget

Correlation between each gate and ROAS across 174 creatives: 0.002 for the hook, 0.026 for the middle, 0.062 for the ending. All three are effectively zero. Creatives in the bottom quartile for completion return 2.37x. Creatives in the top quartile return 2.47x.

This is not a reason to abandon the benchmarks. It means Vilvah has already passed the screen they were designed for. 227 of 290 creatives clear the hook target, so attention is no longer the bottleneck and the metric has stopped discriminating between winners and losers. Use these gates to diagnose why a weak creative is weak, and use ROAS and cost per purchase to decide where the money goes. Do not pause a 3.44x creative because its TSR reads 27%.

0.062Correlation between completion rate and ROAS
2.37x vs 2.47xROAS of the worst and best completion quartiles. Effectively identical
Production source

In-house creative against influencer creative

SourceSpendPurch.ROASHook3s→50%50→100%
Influencer$89,10228672.30x30.0%28.3%37.7%
In-house$75,05723562.33x26.0%38.9%39.8%

A clean trade-off, and both land in the same place. Influencer creative hooks harder at 30.0% against 26.0%. In-house holds far better through the middle at 38.9% against 28.3%. ROAS lands at 2.33x and 2.30x, a difference too small to act on. The obvious test is an influencer hook cut onto an in-house body.

Scope

Why statics and GIFs are not in this study

FormatSpendAdsHook 3s→50%50→100%Why
Video$176,96541627.9%32.7%38.4%The analysis universe for this report.
Static$83,7233120.2%2.9%3.4%No video metrics. Reports 0.2% TSR because there is nothing to play.
GIF$2,919828.1%93.8%41.0%Loops inflate the counter: 93.8% at the 3s→50% gate is impossible. Exclude.
Carousel / image$1,19663.7%23.1%32.4%Mixed formats, too little spend to read.

Statics carry $83,723 and 2452 purchases at 2.22x, almost identical to video. They are working. They simply cannot be judged on video retention, so they sit outside this report rather than dragging the averages down.

By product

Two product lines fail at exactly the same point

ProductSpendPurch.ROASHook3s→50%50→100%
Milkdrop$51,14516222.21x31.1%28.8%39.6%
Mixed / other$31,58910772.34x26.6%36.1%38.2%
Mud Mask$15,4664222.21x23.7%44.7%33.1%
Milk Powder$15,2454582.23x28.5%36.1%35.7%
Milkdrop + Brightening$13,1203962.11x24.2%31.4%44.2%
SPF 50$10,0782671.99x22.1%32.2%29.5%
AM/PM Combo$6,0361682.56x21.9%30.0%38.6%
Face Brightening Kit$4,7932043.13x31.1%22.4%46.1%

Mud Mask and SPF 50 share one failure. Both hold unusually well to halfway, 44.7% and 32.2%, then finish at 33.1% and 29.5%, the two weakest endings on the board. Face Brightening Kit does the opposite and returns 3.13x, the best of any product line.

Worked examples

Four videos, four different numbers

The same three measurements applied to four real assets. Each one felt fine in review. The numbers say otherwise, and they disagree in different places. Scroll sideways. Each card names the number that separates it from the others and points at the part of the edit responsible. Tap a video to play it.

49.6s
Highest hook in the account

Muthu Akka · 4 Week ChallengeIn-house

51.8%Hook
27.7%3s→50%
47.1%50→100%
2.45xROAS

A real customer, face on camera, talking straight into the lens from frame one. 51.8% of everyone served stopped to watch, nearly three times the 18% target and the highest hook of any video here. It then holds 27.7% of those viewers to the 25-second mark on a 49.6-second runtime, which is the longest ask of any creative in this set and it still clears the target. This is the format to clone.

$2,564 spend · 90 purchases · INH_Insta_Video_MuthuAkka_4weekChallenge_Milkdrop_27June26
22.5s
Best middle and best ending

Priya · All The WorkInfluencer

19.1%Hook
43.8%3s→50%
50.0%50→100%
3.75xROAS

A modest 19.1% hook, barely over target, and then it does not lose anyone. 43.8% reach halfway and 50.0% of those finish, and it returns 3.75x, one of the highest in the account. At 22.5 seconds the halfway mark lands at 11 seconds, which is a far easier ask than Muthu Akka's 25. This is the video to study when fixing anything that dies in the middle.

$1,130 spend · 52 purchases · INF_Priya_Video_AllTheWork_FacebrighteningKit_17July26
16.9s
Strong start, collapses at the end

Hema · Milk Powder Face WashInfluencer

38.5%Hook
35.6%3s→50%
24.3%50→100%
1.83xROAS

38.5% hook and 35.6% to halfway, both well clear of target, and then only 24.3% finish against a 40% benchmark. On a 16.9-second video the back half is barely 8 seconds, so losing three quarters of the audience there is a content problem, not a length problem. ROAS sits at 1.83x against a 2.27x video average. The end of this video is where the money leaks.

$334 spend · 8 purchases · INF_Hema_Video_Milkpowderfacewash_POM_01May26
17.6s
Lowest hook, only creative under 1x

Hand Shot · Face Brightening KitIn-house

11.6%Hook
31.8%3s→50%
33.1%50→100%
0.94xROAS

Product in hand, no face, no voice, no claim in the opening frame. 11.6% hook, well under the 18% target, and the only creative in this set returning under 1x. The middle and the ending are respectable at 31.8% and 33.1%, so the people who do stop are reasonably engaged. The problem is that almost nobody stops. Compare frame one of this against frame one of Muthu Akka.

$183 spend · 3 purchases · INH_Video_Handshot_FacebrighteningKit_21Mar26

Runtime is why the middle number is not comparable across videos

Performance still has to be read in context. These four assets run 49.6, 22.5, 17.6 and 16.9 seconds. The 50% gate therefore lands at 25 seconds for Muthu Akka and at 11 seconds for Priya. Holding 27.7% of viewers to 25 seconds is a much harder job than holding 43.8% to 11 seconds, so the two scores are not directly comparable even though both clear the 20% target.

Compare the middle number between videos of similar length, and compare the hook and the ending across everything. When a long video posts a weak middle, check the runtime before briefing a re-cut: the fix may be a trim rather than a reshoot.

The output

Which videos to change, and what to change

Ordered by spend. Every verdict below comes from the number the video failed, not from how the video looks or how it felt in review. This is an edit brief, not a budget decision.

VideoSpendPurch.ROAS Hook3s→50%50→100%Fails atVerdict and the actual edit
L24H · MudmaskINH$4,0151343.00x15.3%76.5%35.5%hookBest hold in the account, worst opening. 76.5% of 3-second viewers reach halfway, the highest of any creative here, but only 15.3% stop the scroll. The body of this video works. Cut three new openings onto the same edit and run them as separate ads.
Brunali · Recommending · MilkDropINF$3,219971.86x22.6%20.2%38.1%middleHighest influencer spend, weakest engine. Clears the hook, then only 20.2% reach halfway and ROAS sits at 1.86x against a 2.27x video average. Seconds 3 to 8 are where this loses. Re-cut or retire before the next flight.
Dhrithi · Challenge · SPF50INF$2,984751.90x23.4%45.5%22.1%endingStrong middle, collapses at the end. 45.5% reach halfway and then only 22.1% finish, the worst completion of any big spender. Whatever sits in the back half is not earning its runtime. Trim the length and move the offer forward.
Review · Removesuntan · MudmaskINH$2,932792.02x24.3%41.4%24.9%endingSame shape as the SPF 50 video. Both are Mud Mask and SPF creatives that hold well to halfway and then lose three quarters of the audience. Check whether these two share an end card or outro template.
Diya · ViralSerum · MilkdropINF$2,9141031.96x24.9%43.2%37.6%endingClose to clearing. Marginal fix. 37.6% completion against the 40% target. A shorter outro or an earlier CTA likely gets it over the line without a reshoot.
MuthuAkka · MilkdropINH$2,564902.45x51.8%27.7%47.1%clears allTemplate asset. Clone this format. 51.8% hook is the second highest in the account and it clears all three gates. The 4-week challenge structure is repeatable across products. Brief the next batch off this.
Kavitha · MilkdropINH$2,8781122.54x48.4%24.1%41.4%clears allSecond template asset. 48.4% hook, clears all three, ran three months straight and still returns 2.54x. Proof the Akka-led format has legs beyond a single flight.
Ram · 5DaysChallenge · FB KitINH$1,734893.36x44.2%20.1%71.2%clears allHighest completion in the account at 71.2%. Also 3.36x ROAS. Whatever the last third of this video does, it should be reverse-engineered into the Mud Mask and SPF creatives that fail exactly there.
Rajini · BotoxinBottle · BASINF$1,252593.48x45.9%44.2%61.0%clears allThe best all-round creative in the account. Clears every gate with room, 3.48x ROAS on $1,252. Underfunded relative to how it performs. Scale it.
Ayshyani · Instagrammademebuy · FB KitINF$3,084801.88x25.2%15.6%25.5%middleRetire. Two versions, same failure. The original and its copy together took $3,084 at 1.88x. Both drop to 15.6% at halfway and 25.5% completion. Nothing here is worth re-cutting.
Karthika · FB KitINF$1,402532.60x37.7%11.4%38.2%middleGreat hook, broken second scene. 37.7% stop, then 11.4% reach halfway, one of the sharpest drops in the account. The first 3 seconds are worth keeping. Rebuild everything after.