Marketing teams prove ROI from AI-generated creative with head-to-head tests against existing creative, matched spend, and the right metric at each layer.
Marketing teams prove ROI from AI-generated creative by testing it head-to-head against their existing creative on the same product and audience, then judging both on purchase ROAS or conversions, not output volume or clicks alone. Budget moves only behind a win. Whilter.AI’s Wonderchef engagement was run this way, on Meta, over 90 days.
Proof is a like-for-like result: the AI-generated asset and the creative you were already running, live against the same product and the same audience, scored on a business metric. Anything softer is a demo.
The Wonderchef engagement is the cleanest example we have. Every ElevateOS asset launched against the incumbent control for the same product and audience, and budget only followed proven lift. The figures come from Wonderchef’s own Meta Ads campaign export, with ROAS defined as purchase conversion value divided by amount spent. Twister went from 0.05× to 3.09× purchase ROAS. Chai Magic went from 0.54× to 2.47×. Link CTR across the matched experiments rose 166%. Per-SKU detail is on the Wonderchef case page, and the solutions page files this kind of work under “Scale the creative, then prove it worked.”
Hold everything except the creative constant: product, audience, platform, window, and a fair share of spend for the control. If the AI-generated set ran a month later with more budget behind it, you’ve measured the calendar and the media plan, not the creative.
Bombay Shaving Co., one of the Retail & D2C accounts, is the compact version. Its sample head-to-head put the Whilter “AI Assets” ad set against the incumbent control over the same window, December 2025 to January 2026, in Meta Ads Manager. Purchase ROAS was 6.49× against 3.54×, and link CTR was 2.45% against 1.36%. Two arms, one window, one platform’s reporting.
Two more steps tighten the read. A concurrent control beats a before-and-after comparison, because seasonality and promotions hit both arms. And a holdout, split by audience or geography where the platform allows it, answers a different question. A head-to-head asks “is this better than what we had?” A holdout asks “did it cause sales we wouldn’t have had anyway?” The first decides where budget goes. The second is the one finance asks. Ask any vendor, us included, which of the two they ran.
Score four layers separately, and don’t let one stand in for another.
| Layer | Question it answers | Typical metrics |
|---|---|---|
| Creative | Did people respond to it? | Link CTR, read rate |
| Funnel | Did they move forward? | Conversion rate, funnel completion, registrations |
| Business | Did it pay back in media terms? | Purchase ROAS, revenue against spend |
| Operations | Could you produce and ship it? | Variants shipped per week, share passing brand review, time to launch |
The reason to keep them apart is on the Wonderchef page. On Chef Magic, link CTR rose from 1.08% to 2.84% while purchase ROAS slipped from 4.33× to 3.75×. Clicks went up and return went down. Judge on CTR and that’s a win; judge on ROAS and you ask why. The case page lists it per SKU, so the question stays visible.
Other accounts report at more than one layer too. PolicyBazaar shows +40% CTR and +10% conversions side by side, across 100M+ personalised creatives in seven languages. ABHI pointed every video at one action, download the app and register the same day, which keeps the funnel metric unambiguous: +44% app-download conversions and 10× same-day registrations. Ooredoo Kuwait scored a personalised video campaign at four steps against a non-personalised control: delivery rate 84.4% to 94.0%, read rate 35.2% to 49.5%, click rate 1.2% to 2.5%, and CTR 3.4% to 5.1%.
Find out what the number is a number of: one ad set, one SKU, a blend, or the whole account. Then ask whether the control ran at the same time on the same audience, whether the figure is spend-weighted and on how much spend, and what the sample size and window were.
Bombay Shaving Co. shows why. The sample ad set moved purchase ROAS from 3.54× to 6.49×, which the case page tables as +83%. The engagement-wide headline on the same page is ~85% ROAS uplift, and the methodology note labels that figure directional. Same brand, two levels of evidence, both disclosed.
Wonderchef’s “8× ROAS turnaround” is a blend too. The methodology describes it as the blended purchase-ROAS recovery across the loss-making SKUs, from ≈0.3× to ≈2.8× across Twister and Chai Magic, stated conservatively. The blended cohort figures are spend- and impression-weighted and sit on ₹8.46L of spend, ₹29.7L of revenue and 1,256 purchases. Those denominators are what let you judge the ratio.
Treat any ROI figure that arrives without a sample size and a campaign window as unproven, ours included.
The cost side should hold everything it took to make, check and ship the creative, not just the media spend. ROAS is purchase conversion value divided by amount spent, so on its own it’s a media-return figure. Production hours, tool fees and review time sit outside it. Count them for the AI-generated set and for the creative it’s competing with, or the comparison is lopsided.
Review is the cost people forget. Bombay Shaving Co. started with every variant needing compliance review before it could go live, and it tracked creative approval, the share of generated variants passing brand review, at 68%. Read that as a yield figure: you pay to generate everything and can only run what passes. The engagement produced 50+ on-brand variants a week.
ElevateOS enforces brand rules inside the engine rather than after the fact, generates 40–60 variants per campaign, and takes a finished system live across Meta, Google and TikTok in under 30 minutes. That makes approval rate and time to launch the two operational numbers worth putting on the same scorecard as ROAS. Benchmarking Brand Compliance covers how to measure whether output is on-brand.
If you want to run this test on your own account, talk to us.
Published 2026-10-06 · Whilter.AI
