Most “AI for ads” tutorials stop at “generate 5 headlines.” That gets you nothing testable — five rewrites of the same angle won’t move CTR. Worse, it ignores how Meta actually ranks ads in 2026: since the Andromeda ranking system rolled out (Feb–Mar 2026), Meta evaluates thousands of times more creative variants in parallel than its predecessor, which means creative diversity, not audience targeting, is now the primary performance lever. This walks through a workflow that produces 9–12 genuinely differentiated variants, each paired with a visual brief and a numeric kill rule, so you feed the algorithm the diversity it now rewards.
TL;DR
- Generate 10 hooks across 10 labeled angles (pain, contrarian, stat, etc.), pick the top 3, then write each in 3 primary-text lengths and brief 3 visuals — a 3×3 learning grid.
- Run 3–5 distinct creatives per ad set, not 30. That is Meta’s sweet spot for the learning phase: enough diversity for Andromeda to find a winner without fragmenting the signal.
- Set a numeric kill rule up front (e.g. link CTR below 1.0% after $20 spend). Don’t kill before ~18 hours — the ad set is still in Meta’s learning phase (50 optimization events in 7 days).
- Best models for ad copy as of June 2026: Claude Sonnet 4.6 or GPT-5.5 for volume and angle diversity; a Custom GPT or Claude Project loaded with your 10 best past ads for tight brand voice.
What this covers
A repeatable system for performance creatives on Meta, TikTok, and Google. The outputs are testable: distinct angles, three primary-text lengths per hook, a visual brief, and a 48–72h kill rule. The point is not “make AI write your ads” — it’s “use AI to widen the test matrix without adding headcount,” which is exactly what Andromeda now pays you for.
Who this is for
Performance marketers, founders running paid ads themselves, indie ecommerce operators, and growth leads on small teams. If you’re spending $1k+/month on paid and have fewer than 5 active creatives, this workflow alone will roughly triple your test cadence — and pull you out of the under-diversified zone Meta’s 2026 algorithm punishes.
Know your baseline first
Before you judge any variant, know what “good” looks like on each platform. As of June 2026, these are the cross-industry medians to benchmark against — your kill rule should sit near your account’s own average, not these numbers, but use them as a sanity floor.
| Platform / placement | Typical CTR (2026) | Notes |
|---|---|---|
| Meta feed (all-industry median) | ~2.19% (link CTR ~1.0%) | Fashion/home decor ~2.8%; tech ~1.0%; healthcare ~0.7% |
| Meta — Lead Gen objective | ~2.59% | ~61% higher CTR than Traffic objective |
| TikTok In-Feed | ~1.5–3% | Spark Ads run ~24% higher CTR than standard In-Feed |
| TikTok (all formats, avg) | ~0.84% | TopView/Brand Takeover skew the high end |
Source figures: industry benchmark reports (digitalapplied.com, triplewhale.com, adamigo.ai), June 2026.
When to reach for it
Launching a new product or campaign; refreshing fatigued creatives (CTR halved or frequency above 3.5); scaling a winner and needing 5+ variations that share its DNA. Don’t use it when you’re still figuring out positioning — fix the offer before optimizing the headline. If you don’t yet know which angle to test, climb one level with the AI campaign ideation tutorial first.
Before you start
- Have one paragraph of clear positioning: who it’s for, what it does, what changes for them, what they pay. Without this, hooks come out generic.
- Pull 3–5 winning competitor ads (the Meta Ad Library is free and public). Paste them into the prompt as “voice and angles already in market — do not repeat.”
- Decide your kill threshold up front. “Link CTR below 1.0% at $20 spend = kill” is testable. “It feels weak” is not.
- Note that since February 2026, new Sales, Leads, and App Promotion campaigns launch with every Advantage+ Creative enhancement turned on by default. If you want to test your exact copy/visual against a control, toggle the enhancements off per ad, or the system may rewrite your headline and re-crop your image.
Step by step
- Define the brief: product, audience persona, ad placement (Meta feed / Reels / TikTok In-Feed / Google Search), and the one metric you’ll judge by (link CTR, CPC, signup rate).
- Generate 10 hooks. Use this prompt:
Give me 10 first-line hooks for a Meta feed ad selling [product] to [persona]. Use these angles in order: pain, desired outcome, contrarian take, surprising stat, founder story, before/after, social proof, objection-flip, urgency, curiosity gap. One sentence each. No emojis. - Pick the top 3 hooks by gut plus the test “would I stop scrolling for this?”
- For each chosen hook, generate primary text in 3 lengths: 15 words, 50 words, 90 words. Different placements reward different lengths (Reels rewards short; feed tolerates long).
- Write a visual brief for each:
Photo of [subject] doing [action], shot on iPhone, natural light, single product visible in frame.Send to a designer or image generator. - Launch 9 variants (3 hooks × 3 visuals) as separate ads inside one ad set — not one dynamic-creative ad. Same audience and budget. Let it run 48–72 hours. Kill anything below threshold; scale anything 1.5× above your account average.
Meta vs. TikTok vs. Google: how many of each asset
The 3×3 grid maps differently onto each platform’s ingestion. As of June 2026:
| Platform | What you upload | Recommended count |
|---|---|---|
| Meta (separate ads) | Full ads (hook + body + visual) | 3–5 distinct creatives per ad set |
| Meta (dynamic creative, one ad) | Asset pools | 5–8 images, 3–5 headlines, 3–5 primary texts, 2–3 CTAs |
| TikTok | Native video variants | 3–5 hooks tested in first 3 seconds |
| Google RSA | Text assets, Google assembles | up to 15 headlines (30 chars), 4 descriptions (90 chars) |
For Meta, prefer separate ads over dynamic creative when you want to learn which angle won — dynamic creative tells you the winning combination but blurs which single element drove it.
First-run exercise
- Pick one ad set, not your whole account. Even better: pick a low-stakes audience (lookalike, not retargeting).
- Run the 9-variant test exactly as above. Don’t tweak budget mid-flight — under Andromeda, a budget change above ±20%, an audience tweak, or a creative swap can reset the learning phase, and you’ll burn another 50 events relearning.
- After 72 hours, write down which angle won, which lost, and your best guess at why. The guess matters — it’s how you build account-specific intuition.
- For round 2, change only the visuals (keep the winning hook). You’re isolating one variable at a time.
Quality check
- Are the 10 hooks actually different angles, or 10 rewrites of “we’re the best”? If they share more than 3 words across variants, the prompt didn’t enforce difference — rerun with explicit angle labels.
- Does each visual brief produce something a real photographer could shoot? “Eye-catching” is not a brief; “Woman in her 30s, kitchen counter, holding mug, soft morning light” is.
- Is the kill rule numeric?
link CTR < 1.0% at $20 spend— not “it feels off.”
How to reuse this workflow
- Save the winning hook, visual brief, and full prompt as a template. Next launch, swap product/persona only; keep the angle scaffolding.
- Maintain a kill log: dead hooks by angle. After 30 tests, patterns emerge (“contrarian wins for us; aspirational doesn’t”).
- Refresh the competitor pull every 6 weeks — the market voice drifts and so do your angles.
A realistic one-day launch
Brief at 9am → 10 hooks by 9:15 (Claude Sonnet 4.6 or GPT-5.5) → top 3 picked → 9 visual + copy variants briefed by 10am → designer or image tool returns assets by 4pm → ads live by 5pm → kill / scale call on day 3 morning, after the ad set has cleared its first ~18 hours of learning. One person, one day, nine learning units.
Common mistakes
- Launching one variant — you learn nothing about angle, only about that exact ad, and you starve Andromeda of the diversity it now rewards.
- Killing before ~18 hours of spend — the ad set is still in the learning phase (it needs ~50 optimization events in 7 days); you’re killing noise.
- Letting AI write the offer, not just the hook — the offer is the lever; AI optimizes packaging.
- Editing live to “fix” a loser — any meaningful edit can reset learning under Andromeda; kill and relaunch instead.
- One visual paired with three hooks — you can’t tell if the hook or the visual won. Test the 3×3 grid.
- Skipping the angle labels in the prompt — generic prompts give you 10 versions of “amazing product.”
FAQ
- Best models for ad copy in 2026?: For volume and angle diversity, Claude Sonnet 4.6 or GPT-5.5 as of June 2026. For very tight brand voice, a Custom GPT or a Claude Project loaded with 10 of your best past ads so the model writes inside your house voice instead of a generic one.
- How many variants is too many?: For a single Meta ad set, run 3–5 distinct creatives during the learning phase — more than that fragments the signal and each creative gets too few impressions to learn from. Generate 9–12 in your workflow, then ship the strongest 3–5 per ad set.
- What about video ads?: Same workflow, but the visual brief becomes a 3-shot storyboard. The hook in the first 3 seconds is non-negotiable on TikTok — that window decides view-through and CTR.
- Does Meta’s Advantage+ already do this for me?: It optimizes combinations of assets you supply and, since Feb 2026, applies creative enhancements by default. It does not invent fresh angles. You still need distinct hooks to feed it — this workflow produces them.
- How do I generate the full campaign concept, not just ad variants?: Step up a layer with the AI campaign ideation tutorial — start with offer + angle + audience, then cascade into hooks and visuals.
- Can AI judge which hook will win?: No. It can predict pattern matches with past winners; it cannot predict tomorrow’s audience mood. Run the test.
Related
- Ad creative prompts
- Campaign slogan prompts
- AI ad creative ideation: 20 distinct angles
- AI generates ad creative variants
- AI writes platform-native social captions
- AI campaign ideation tutorial
- AI Amazon Listing Tutorial: Title, Bullets, A+ Content
- AI Content Series Planning Tutorial: 12-Episode Roadmap
- AI LinkedIn Thought-Leadership Post Tutorial Without Cringe
- AI TikTok Script Tutorial: Hook + Reveal + CTA in 60 Seconds