Last updated: October 2026 · By The Ecommerce Benchmark Team
Reverse-engineering an ad is defined as identifying which of its layers is responsible for its performance, rather than describing what the ad contains. To reverse-engineer why an ad is winning, decompose it into five layers, audience, offer, claim, hook and execution, then swap each for a generic version and see which swap collapses the ad. Drawn from 7,000+ websites and 15+ industries, the CROBenchmark dataset grades each site on 248+ audit criteria [CROBenchmark Report 2026, Omniconvert]. This guide shows the Five-Layer Teardown and the Substitution Test, and how to turn a diagnosis into a test.
- Split every ad into 5 layers: audience, offer, claim, hook and execution.
- Run the Substitution Test by swapping each layer for a generic version and noting which swap kills the ad.
- Check the offer first, because offer wins masquerade as creative wins and copying the creative alone gives a false negative.
- Copy the objection an ad answers, which is the most transferable element, and leave the artwork.
- Write the diagnosis as a hypothesis that could fail, then run the test that would fail it.
Working out why an ad is winning means identifying which part of it is carrying the performance, which is a much harder question than noticing that it has run for four months. Most competitor analysis stops at the easy half and then copies the wrong thing: the edit, the colour grade, the voiceover style, none of which is usually what made the ad work. Omniconvert has measured creative performance across 70,000+ experiments and 2,500+ Shopify stores over 13 years in eCommerce, and the single most common cause of a disappointing creative test is a confident diagnosis that named the wrong layer. This guide sets out the Five-Layer Teardown, the Substitution Test that isolates the load-bearing element, and how to turn a diagnosis into something you can actually prove.
Why descriptions are not diagnoses
The question a diagnosis has to answer is counterfactual. If this ad had been shot in a studio instead of on a phone, would it still work? If it had opened with a product shot instead of a complaint, would it still work? If the offer had been ten percent off instead of a bundle, would it still work? Only one or two of those swaps will plausibly break it, and those are the answer.
The Five-Layer Teardown
- Audience. Who is this for? Inferred from vocabulary, assumed category knowledge, the objection opened with, the price implied and the context shown. Confidence is moderate and usually sufficient.
- Offer. What is being proposed? Bundle, trial, subscription, guarantee, straight discount, or nothing at all. This is fully visible and frequently the real answer.
- Claim. What is asserted, and which worry does it settle? The claim is the argument of the ad, stripped of its delivery.
- Hook. What happens in the first two or three seconds, and why does it stop a scroll? Hooks are a small set of recurring structures rather than infinite inventions.
- Execution. Format, pacing, voice, casting, grade, music. Highly visible, highly expensive to copy, and rarely the mechanism.
How to reverse-engineer why an ad is winning, layer by layer
The Substitution Test is defined as replacing one layer at a time with its most ordinary version and judging whether the ad survives. It is deliberately crude because precision is unavailable from outside, and it gives a ranking, which is enough to decide what to build.
In most cases exactly one substitution causes a visible collapse. That is the load-bearing layer. Occasionally two interact, typically a claim and the offer that makes it credible, and that pairing is itself the finding.
The public libraries make this easier than it sounds, because they show you the advertiser's variant set. Google's Ads Transparency Center and Meta's equivalent let you see several creatives from one advertiser side by side, and the element they hold constant across variants is the element they believe is doing the work. That is the advertiser's own diagnosis, revealed by their behaviour rather than stated.
Offer wins dressed as creative wins
What transfers, and what does not
| Layer | Transfers between brands? | The usual failure when copied directly |
|---|---|---|
| Objection answered | Strongly, within a category | Rarely fails; the main risk is answering an objection you do not actually have |
| Hook structure | Well, if rewritten in your own language | Copied word for word it reads as borrowed and dates quickly |
| Offer shape | Well, if your margins allow it | Copied without the margin behind it, it wins volume and loses money |
| Claim | Only if you can substantiate it | An unsupportable claim is a compliance problem, not a creative one |
| Audience framing | Partly; depends on overlap with your buyer | Right argument aimed at a buyer who was never yours |
| Execution and production style | Poorly | Most of the budget, least of the effect, and it looks derivative |
Turning a diagnosis into a test
A clean test holds your brand, format and production constant and varies only the diagnosed layer. If the diagnosis is the sizing objection, both arms are your creative, in your voice, and one addresses sizing while the other addresses something else. A result from that test is directly actionable and remains true after the competitor pulls their ad.
What a performance marketer should do this week
- Pick a single long-running competitor creative. One, not a folder.
- Write one sentence per layer: audience, offer, claim, hook, execution.
- Substitute each layer with a generic version and note which swap collapses the ad.
- Check explicitly whether the offer is carrying it, and whether your margin could match it.
- Look at the advertiser's variant set: whatever they hold constant is their own diagnosis.
- Write the hypothesis in one sentence, then design the two-arm test that would disprove it.
Creative diagnosis only pays if the traffic lands somewhere that converts. The free Ecommerce Benchmark leaderboard score rates a store across six dimensions, Creative and Ads, Reviews and UGC, AI Visibility, Agentic Commerce, Competitor Synthesis and CRO, benchmarked against real competitors in your category and country, with the paid audit report returning the full prioritised detail. For the page-side half of the problem, the CRO audit checklist covers the destination. And where a team wants the creative hypotheses ranked and generated continuously rather than one teardown at a time, Omniconvert Commander (formerly Nexus) is the AI for eCommerce growth engine that unifies commerce data, prioritises experiments by True Profit and produces the creative you approve before it goes live.
Related reading: DTC competitor research stack and spot a winning ad.
Frequently asked questions
How do you work out why a competitor ad is working?
Break it into five layers and test each one for whether the ad would survive without it. The layers are the audience it is aimed at, the offer it carries, the claim it makes, the hook in the first seconds, and the execution. Most teams look only at execution, which is the layer that transfers least. Run the Substitution Test on each layer in turn, imagining the ad with that element swapped for a generic version, and the layer whose removal kills it is the one carrying the performance.
What is the difference between a creative win and an offer win?
A creative win survives a change of offer; an offer win does not. If you imagine the same ad with a standard discount instead of its unusual one and it still looks compelling, the creative is doing the work. If it collapses into an ordinary ad, you are looking at a pricing decision wearing a creative costume. This distinction matters enormously, because copying the creative without the offer reliably produces a disappointing test and a wrong conclusion about the angle.
Can you tell who an ad is targeting from outside?
Not precisely, but you can infer a great deal from the creative itself. The vocabulary, the objection it opens with, the context shown, the price point implied and the level of category knowledge assumed all narrow the audience considerably. An ad that opens by explaining what a category is addresses a newcomer; one that opens with a comparison against a named alternative addresses someone already shopping. That inference is usually enough to decide whether the angle is relevant to your buyer.
What is the most transferable part of a winning ad?
The objection it answers, followed by the hook structure. The specific visual, voice and edit are the least transferable, because they are tied to a brand and a production budget. If a competitor creative is winning by addressing a worry about sizing, durability or delivery, that worry exists in your category too and your version can answer it in your own voice. Copy the argument, not the artwork, and the test tells you something you can use.
How do you confirm your diagnosis is right?
Write the diagnosis as a testable hypothesis before you build anything, then run one test that would fail if you were wrong. If you believe the win is carried by the sizing objection, build two variants: your normal creative answering the sizing objection, and your normal creative answering something else. If the first outperforms, the diagnosis holds independently of the competitor execution. A diagnosis never tested is just a confident opinion about another company advertisement.
With the load-bearing layer named, the next step is to turn a competitor's winning ad into your own without reproducing its expression.
Most teams that analyse a winning competitor ad write an inventory: fast cuts, phone footage, a text overlay, a discount in the last frame. The brief then reproduces every surface feature and none of the mechanism, and the result underperforms. The most common miss is an offer that carried the ad, such as a long guarantee or a bundle, copied without the margin behind it.
The two-arm test in your own brand language is the kind of step a structured programme runs routinely. AliveCor ran a homepage redesign through Omniconvert as a single A/B test across all devices and recorded a +10.59% conversion rate, +10.79% revenue per user, and a 95.25% chance to win. [Omniconvert, AliveCor case study]
Name the load-bearing layer, then try to disprove it
Noticing that an ad has run for months is observation; naming the layer that keeps it running is analysis, and only the second one changes what you build. Decompose every creative into audience, offer, claim, hook and execution, then substitute each layer with a generic version and watch for the swap that collapses it. Check the offer explicitly, because offer wins impersonate creative wins more often than anything else and copying the artwork without the pricing produces a false negative you will act on for a year. Then write the diagnosis as a hypothesis that could fail, and run the two-arm test in your own brand language that would fail it.
