Derived appeals: keeping an improvement trail on your creatives

Say the "material · detail" appeal worked: one close-up of a nacre watch dial, two lines of copy, click-through at nearly double your account average. Now what?

Most teams pick one of two wrong answers. The first is to run the winner until it dies: three weeks in, click-through drifts down, budget gets cut, the appeal is gone. The second is to jump to a brand new appeal — price this time, warranty next — so you never learn what about the material angle landed.

The right answer is the one nobody schedules: push on what worked and find its edge.

What a derived appeal is

Children of the root "material · detail" look like this.

• Stronger material proof — don't assert the dial is special; show three angles and the finishing process.
• Material against the alternative — put it beside a printed dial and let the difference argue.
• Material in use — not a product shot but a wrist, in changing light, on a real day.

None abandons the reason — the buyer is still persuaded this watch is made differently; only the evidence moved. A new appeal is different. "Three-year warranty" says nothing about material, so it is a sibling of the root, not a child. Lose that distinction and the tree becomes a flat list with indentation.

Appeal, copy and creative are separate layers, covered in what an ad appeal is. Derivation happens inside the appeal layer; rewriting an appeal in a snappier sentence is a copy test.

The one-variable rule

A derived appeal may change exactly one of four things: the evidence, the framing, the audience, or the medium — whether copy or image carries the argument.

Change two and you learn nothing. Strengthen the proof and shift the audience, watch click-through rise, and you cannot attribute it. Next quarter you want that win back and don't know what to repeat.

Four one-variable children of one root:

The last row is the one teams skip, though the same appeal performs very differently depending on whether copy or image leads. Testing wording variants inside one creative is a separate job, covered in A/B testing ad copy.

When to derive, and when to move on

Four triggers cover nearly every case.

That last row is the single most common waste in this workflow. An appeal underperforms, someone says "maybe we executed it badly," and three derivatives get built on a foundation that never held. If people don't buy for that reason, re-arguing it changes nothing. The counterweight: don't judge a root on one creative — run it behind two or three images first.

The tree stays shallow

Eventually someone wants to derive from a derivative. Don't — two levels is almost always enough.

• Level 1 (root) answers which reason we sell on. That's the skeleton of the library.
• Level 2 (derived) answers how best to argue that reason.
• Level 3 answers no question anyone has. It tests noise.

The hard constraint is budget. A root with four children is five branches, and buying each enough impressions to read costs real money. A level deeper is ten-plus branches carrying numbers too small to separate from chance.

The soft constraint is readability. A two-level tree reads at a glance: the idea's best result so far, and how much it moves when the argument changes. At three levels that breaks, since grandchildren differ from siblings on one variable and from cousins on two. Wanting to go deeper usually means the derivative has grown into a new appeal — promote it to a root instead.

Reading a tree

Read a root and its children as one unit. Two rules.

1. The root's number is the ceiling of this idea so far.
2. The spread between children says whether execution or the idea is the constraint.

Baseline here — the running average of existing creatives — is 0.80% CTR at ₩42,000 CPA, and every row was read at 30,000+ impressions.

The root cleared baseline by 1.8×, so the reason is real. The evidence child pushed it another 19%, so the ceiling isn't 1.44% — it's at least 1.71%. Then the spread: 1.09% to 1.71%, a factor of 1.6. Children that fan out that widely mean the constraint is execution, and headroom is left in the idea.

Now the opposite shape.

Everything sits between 0.81% and 0.88% against a 0.80% baseline. Nothing done to the argument moves the number, so the constraint is the idea — the answer is a different root, not a fifth derivative. Pulling numbers by period is covered in comparing creative performance.

The trail is the point

The real value of a derived appeal isn't this quarter's click-through. It shows up six months later.

A new hire says: "What if we showed the nacre dial next to a printed one?" Without a tree that sounds like a good idea, and two weeks and a test budget go to reaching last year's answer. With a tree it takes five seconds: we ran that in March, it came in below the root, and the comparison made us read as the cheap option.

An appeal tree records what worked, but more importantly what was tried and rejected — the expensive half. Successes stay visible because they're still running; failures are invisible unless someone wrote them down, which is why they get reinvented by whoever joined last.

copy hound keeps derived appeals as children of their root, each carrying an improvement note saying what was changed. Hanging off each child are the per-channel creatives — Kakao Moment, Naver GFA and Meta copy checked against character limits — the image gallery, performance rolled up by period, and a dated stop / keep / scale verdict. For structuring the roots themselves, see building an appeal library.

In short

• A derived appeal keeps the reason and changes one of evidence, framing, audience, medium.
• Two failed derivatives means move on; a root that never beat baseline means don't start.
• Two levels deep, never three. Wide spread between children means execution is the constraint; narrow spread means the idea is.

FAQ

How do I tell a derived appeal from a new one?

Ask whether the reason to buy changed. Still arguing "made differently"? Derived. Moved to "backed by a warranty"? New root. When it's ambiguous, pull the key noun from each headline — same noun, same appeal.

How many derivatives should one appeal have?

Three or four in total is plenty, and run no more than two at once — launching four together splits impressions until none reads cleanly. One variable at a time is faster in practice.

Can I derive from an appeal that underperformed?

Not recommended. Give the root a fair shake first — two or three images behind the same appeal — but if it still can't clear baseline, the reason failed, not the argument.

Can I manage this in a spreadsheet?

You can start there. Parent-child relationships break first: as derivatives accumulate it stops being obvious which row belongs to which root and what was changed. copy hound keeps the tree, improvement notes, per-channel copy, numbers and verdict log on one line. If all you need is a character count, the free counter is public.

Read next