Ad creative angles - a working taxonomy for ecommerce

An angle is the claim an ad makes about why the product matters. It is not the hook, which is how the first two seconds earn attention, and it is not the format, which is the container the claim arrives in. Most creative testing runs ten versions of one claim, reads a flat result and concludes the product doesn't sell.

What is an ad angle, and how is it different from a hook?

Three things get called "the creative" and they are separately choosable.

The angle is the claim. It's the sentence you would have to defend if someone argued with the ad: this stops your hair breaking, this costs less than what you're doing now, people like you already switched. An angle can be true or false, and that's the test of whether you've actually named one.

The hook is how the first two seconds earn attention. A question, a mess on a counter, a number on screen, a face mid-sentence. There's no argument in a hook. It buys the time to make one.

The format is the container. UGC talking head, static image, product demo, founder to camera, review screenshot. Format mostly decides what the asset costs to produce and how quickly you can make ten more of them.

Separating them matters because they fail differently and they're fixed differently. A weak hook means nobody heard the claim. A weak angle means they heard it and it wasn't about them. A weak format means the claim was legible but the ad looked like an ad. Three symptoms, three fixes and only one of them is solved by hiring a better editor.

What are the angle families?

There's no canonical list, so here's one. It's drawn from the patterns that recur across DTC ad libraries rather than from a framework, and it's deliberately finite: twelve families, each answering a question a buyer is already asking themselves.

They aren't mutually exclusive. Most good ads lead with one family and support with a second. Naming the lead is what makes them countable.

The last column carries more weight than it looks. Mechanism angles fail on products where nobody believes there's a mechanism — a tote bag doesn't have one — and identity angles fail on products bought under duress. A blocked drain isn't an identity purchase.

Two of these families have unusually good public evidence behind them. In Baymard's checkout survey, 40% of abandonments come down to extra costs — shipping, tax and fees — being too high, and 18% to the site demanding an account before checkout. That is a documented list of the objections your buyers actually hold, which makes objection-handling and price-reframe angles the two you can write with the least guessing. Two caveats worth carrying: the survey excludes people who were only browsing, so these shares describe shoppers with real intent, and it is separate work from Baymard's 70.22% abandonment rate, which is synthesised across 50 studies.

Why is this a matrix rather than a list?

Because the three axes multiply. Twelve angle families, five or six hook types and four formats you can realistically produce is somewhere north of 240 distinct ads. Nobody runs 240 ads, and nobody should try.

A 5% hit rate means roughly twenty shots to find one winner. Twenty shots at the whole space and twenty shots at one cell are completely different experiments, and they cost the same. If the cell you picked happens to be empty, twenty samples will confirm it's empty with real precision and tell you nothing about the other 239.

The cost of picking wrong keeps going up. Meta's own results for 2025, published in January 2026, show ad impressions up 12% across the year and the average price per ad up 9%. That's a blended worldwide figure across every objective and country rather than a US ecommerce CPM, so read it as direction and not as your number. The direction is that inventory grew and got more expensive at the same time.

How do you sample the matrix without running hundreds of ads?

Vary one axis at a time, in order of how much variance it carries. Angle first, hook second, format last.

Angle is the coarsest axis. The gap between "this stops the thing you hate" and "people like you already own this" is not a different ad, it is a different customer. Hook is finer: it changes who hears the claim, not what the claim is. Format is finest of all: it moves production cost far more than it moves response, which is precisely why it is the axis most briefs vary.

Around twenty ads across three rounds is one expected winner at the measured hit rate. The difference is that this winner comes with a description. You know which claim carried it, which is what separates an ad you got from an ad you can make again.

One honest gap. Nobody has published a variance decomposition of creative performance by angle, hook and format, the study that would say what share of the spread in outcomes each axis explains. The angle-first ordering here is inference from how the three axes behave, not a measured result. If that study ever gets published and it says format carries the variance, run the rounds backwards.

How do you extract an angle from a competitor ad you can only watch?

You can't see their spend, and you don't need to. For ordinary US commercial ads the Meta Ad Library shows no spend or reach at all. The two disclosures are separate and they move in opposite directions: spend and impression ranges appear only on political and social-issue ads, while total reach appears on ads delivered into the EU and the UK and generally not on political ads. The UK is listed by Meta as its own disclosure region rather than part of the EU, and the Digital Services Act does not apply to it. What you get everywhere is the creative itself, the date it started running and how many variants exist. That's enough.

1. 01

   Transcribe the first sentence verbatim

   Not what it means. The actual words, in order. Paraphrasing at this stage is how you accidentally record your own theory of the ad instead of the ad.

2. 02

   Strike out everything that is production

   The face, the music, the captions, the aspect ratio, the cut on the beat. All of that's hook and format. Whatever survives the deletion is the claim.

3. 03

   Ask what the claim would be wrong about

   An angle is falsifiable-shaped. "This works because of the enzyme" fails if there is no enzyme. If nothing about the surviving sentence could be false, you have found decoration and the real angle is somewhere else in the ad, usually in the on-screen text.

4. 04

   Name the lead family and the support family

   Most ads carry two. A before-and-after that opens on a dermatologist is leading with authority and supporting with before-and-after, and those are two different bets you can test separately.

5. 05

   Record how long it has been running, not how it feels

   Start date is the only performance-adjacent signal in the public record. An ad still live after ninety days is being paid for by someone who can see the numbers. Comments and likes aren't that signal.

6. 06

   Tally the whole library, then look at what is missing

   Count families across every ad a competitor is running, not ad by ad. Forty ads across three families tells you which claims survive in your category. The nine families nobody in the category is using are the cheapest information in the exercise, and the only part of it your competitors haven't already paid for.

Reading a competitor ad library covers what else the public record contains and the specific ways it misleads. For this exercise the tally is the output. Everything else is context.

What happens if you test ten executions of one angle?

You get a precise answer to a question you didn't mean to ask.

It happens because briefs are counted in assets. "We need ten videos this month" is a production target, and an editor handed a production target varies the things that are cheap to vary, which are hook and format, the two axes that carry the least variance. The claim stays fixed because nobody wrote it down as a variable in the first place.

The diagnostic takes five minutes. Print your ten scripts and underline the one sentence in each that makes the case. Now try swapping two of those sentences between two ads. If both ads still work, you ran one test.

What it costs is a wrong conclusion at the worst possible altitude. The account reads flat, and the sentence people say next is that the product doesn't work on Meta. The honest conclusion available from that data is narrower: this claim, in this format, at this budget, did not work. Those aren't the same statement and only one of them should change what you do next.

How do you tell whether the angle failed or the ad did?

Three checks, in order, before you retire a family.

Delivery first. An ad that never got enough impressions to be read did not test anything, and the temptation to read it anyway is strongest exactly when the budget is tightest.

Then hook rate, used as a validity check rather than a score. Low hook rate means nobody heard the claim: you tested your first frame, not your angle. A healthy hook rate with nothing behind it means they did hear it and it wasn't about them, and that is a genuine angle result. Hook rate measures the first three seconds and nothing after them, which is why a strong one so often sits beside flat sales — and why validity is the only job it should be given here. Hook rate and hold rate covers the two competing denominators that make the numbers non-comparable between tools.

Then truth. Some angles are false for your product and no execution rescues them. If the mechanism angle keeps losing, the useful question is whether your product actually has a mechanism a buyer would credit, or whether you have been describing a manufacturing process to someone who wants their floor clean.

Where does this taxonomy stop being useful?

It is descriptive, not predictive. Nothing in the twelve rows tells you which family wins for your product. It tells you which ones you have not tried, which is a smaller claim and the only one the evidence supports.

Execution still dominates inside a cell. A strong execution of a mediocre claim beats a weak execution of a good one, so a single round-one loss is not proof a family is empty. If a family should work for your product on first principles, give it one more shot with a different writer before retiring it.

And the scoreboard you score angles on is noisy in a direction that matters. Across 640 Meta incrementality experiments published by Haus in July 2025, brands already doing at least a quarter of their business outside DTC saw about 32% of Meta's impact land on those other channels rather than the DTC store — a finding about multi-channel brands, not about every advertiser. In the same work, DTC-only brands measuring on click-only attribution saw Meta under-report its own incremental contribution by roughly 15%. Neither number describes Meta's default seven-day-click-plus-one-day-view setting, which was not the configuration tested. Angles that work upper-funnel — identity, authority, novelty — are the ones that lose most under in-platform last-click scoring, which means the families most likely to be wrongly retired are the ones that build a brand. MER, ROAS and what to steer on goes into which number to judge that with.

None of that makes the taxonomy less worth having. It makes it a search order rather than a ranking. Twelve rows, a tally of what your category already runs, three rounds that vary one axis at a time. That is a way of spending twenty ads on twenty questions instead of one.