The AI Ad Copy Workflow We Run for Clients

The short version: We run ad copy the way we run visual creative: high volume, tight taxonomy, human filter. Five hook angles (pain, proof, contrarian, price, social norm), variants generated per angle per awareness level, a kill pass for claims, policy, and voice, then results read at the angle level. One strategist runs a full sprint in about a day.

Tuesday morning, one client, one product: 40 ad copy variants sitting in a doc. Eight hooks per angle across five angles, each line tagged with the awareness level it targets. To someone new, it looks like a wall of text. The tags make it legible: every line is one argument, aimed at one state of mind.

About 15 of those lines will be dead by lunch. Not because the model wrote badly, but because they claim things we cannot back, risk platform policy flags, or sound like a different brand. The filter is the point of the workflow, and this post walks through all of it.

Copy is the quiet half of our AI performance creative work. The visual workflow gets the attention because photoreal avatars are impressive. The copy system underneath moves numbers on the same scale and costs little to run.

What is the AI ad copy workflow?

The workflow has five steps. Define a taxonomy of hook angles (pain, proof, contrarian, price, social norm). Generate variants per angle per awareness level with an LLM. Filter by hand for substantiation, policy, and voice. Pair survivors with creative, one angle per cell. Then read results at the angle level, not the variant level.

The reason to invest here is the same reason to invest in visuals. According to Nielsen’s Catalina study, creative drives about 47% of sales lift in advertising while targeting drives roughly 9%. The words carry the hook, offer, and objection handling. Most teams still treat them as an afterthought typed into Ads Manager at launch.

How do you build a hook and angle taxonomy?

Start with five angle families: pain, proof, contrarian, price, and social norm. Every hook gets one family tag plus an awareness level (problem-aware, solution-aware, or product-aware). The taxonomy turns a pile of lines into a structured test. Without it you learn which sentence won. With it you learn which argument won.

Five families cover most of paid social. We have tried larger taxonomies and they start to blur: two strategists tag the same hook differently, and the angle-level reads fall apart.

Angle family	What it sounds like	When it wins
Pain	"You have re-downloaded this app four times"	Cold, problem-aware traffic that has never heard of you
Proof	"4,000 reviews and the same sentence keeps showing up"	Solution-aware buyers comparing options, retargeting
Contrarian	"Skip the 10-step routine. You need two."	Saturated categories where every ad makes the same promise
Price	"That is $0.40 a day, less than the tip on your coffee"	Product-aware audiences stalling at checkout, promo windows
Social norm	"Half your group chat already switched"	Broad consumer products with word-of-mouth loops

The taxonomy is also a brief-writing tool. When a client asks for “more ads,” the useful question back is which angle they believe is underexplored, and the table gives that conversation a shared vocabulary.

How do you generate variants per angle and awareness level?

Generate in small, constrained batches: one prompt per angle per awareness level, eight hooks per prompt. Constraints do the work. Word caps, a spoken register note, a ban on invented statistics, and a voice reference from past winners keep the output usable. A wide-open “write me 40 ad hooks” prompt returns 40 versions of the same idea.

Here is the hook prompt, shortened:

For [product], write 8 ad hooks in the PAIN angle for a
PROBLEM-AWARE audience. The reader knows the frustration
([specific frustration]) but not the product category.
Rules: under 12 words each, no brand name in the hook,
no exclamation points, spoken register, each hook names
a different concrete moment of the pain.
Output as a numbered list, one line each.

Winners from the hook pass get expanded into primary text with a second prompt:

Take hook #3 and write 3 versions of primary text, 40 to
80 words each. Voice reference: [3 pasted lines of past
winning copy]. Stay inside this claims inventory:
[approved claims list]. Do not invent statistics, awards,
guarantees, or customer quotes.

The claims inventory is a client-specific doc: every claim the founder and legal have signed off on, with the source next to it. The model never gets to argue from outside it.

What gets killed in the human filter step?

Three things: claims we cannot substantiate, lines that create platform policy risk, and voice drift. Across our sprints, roughly a third of a generated batch dies here. That ratio is our own observation, not an industry benchmark. The pass takes one strategist under an hour, and it is the only step we never automate.

Unsubstantiated claims. The model writes confident numbers with no source behind them. “Clinically proven to double retention” is a great hook and a legal problem. If a claim is not in the inventory, the line dies or gets rewritten around a claim that is.

Platform policy risk. Meta restricts ads that assert or imply personal attributes, and “you” plus a named insecurity is exactly the construction the pain angle wants to write. So we rewrite those lines to describe the situation rather than the person. “Your acne” becomes “the third concealer this month.”

Brand voice drift. By variant 30, the model slides toward the category-average voice. We paste past winners into every prompt and still catch it. The test is blunt: if a line could run under a competitor’s logo unchanged, it dies.

How do you pair copy with creative without exploding the test matrix?

Do not fully cross them. Six copy survivors against six creatives is 36 cells, and most budgets cannot feed 36 cells to significance. Pair within angle instead: pain copy runs on the pain-concept visual, proof copy on the proof visual. One angle becomes one test cell with 4 to 5 variants inside it.

The pairing rule holds because a mismatched cell is unreadable anyway. Contrarian copy over a testimonial visual is two arguments interrupting each other. When that ad loses, you cannot say which half failed, so the result teaches you nothing.

For one consumer subscription client, this took a proposed 48-combination launch down to 5 angle cells with about 22 ads total. Same learning agenda, far less budget needed to reach a read.

Why read results at the angle level, not the variant level?

Because variant-level reads are mostly noise. At typical test budgets, a single variant rarely collects enough conversions to separate skill from luck. Aggregating spend to the angle level reaches a stable read in days instead of weeks. And it answers the question that feeds the next batch: which argument is working.

Meta’s own research on creative quality points creative to more than half of ad performance on its platforms. The angle read makes that lever usable: next sprint, the winning argument gets 16 fresh hooks and the losing one gets retired.

Angle-level reads also change how you handle fatigue. When a winning angle’s numbers sag, you rotate fresh variants inside that angle instead of scrapping the argument. Our guide to detecting ad creative fatigue covers the warning signals; the taxonomy is what makes the refresh cheap.

Where this fits in a full creative system

Copy and visuals are one system for us, not two departments. The batch, the filter, and the angle read run inside the same sprints that produce our avatar and video work, and the two halves share a taxonomy so the learning compounds across both.

If you want this running on your account, Book a Free Strategy Call and bring your current ads. We will tag them into angles on the call and show you which arguments you have never tested.