Andromeda-aware creative ladder — audience-first, concept-first, no micro-variations

Andromeda no longer optimizes across your entire ad account against one ranking pool — it optimizes inside audience-segment-level auctions. Two creatives competing for "men 30-45" with only minor differences (hook word swapped, button color changed) get de-prioritized because Andromeda treats them as one concept inside that segment's auction. The framework: start from the audience profile → write a creative-concept brief for each audience → if a talking-head spokesperson is used, build a persona profile (age, voice, look, accent) per concept → generate concept-tied assets via your AI image+video pipeline. Micro-variations are banned. Each new asset has to be a different concept end-to-end, not a cosmetic edit of an existing one.

This framework is the operational answer to why the perceptual-hash dedup pipeline ( ) works. Pair them: dedup catches micro-variations before upload; the ladder structures the brief so your creative team doesn't produce micro-variations in the first place. Meta's own engineering posts on Andromeda (the retrieval engine) and GEM (the generative ad model) are the canonical reference — this case is the operator-translated playbook for working inside Meta's new design. Multiple operators in source confirm the audience-segment auction is the dominant ranking surface in 2025+; one notes the entire effort of a performance marketer now sits in "ability to generate genuinely different concepts at scale" rather than in targeting.

1. Define audience segments at Andromeda's granularity. Not "men 25-44" but the segment grain Andromeda actually auctions inside — work backwards from your account's segment-level performance reports, not from your demographic intuitions. Each segment becomes a creative-concept brief.
2. Write one creative-concept brief per segment. Story, persona, hook, location, time of day, emotional arc. Treat each as a separate ideation — not as a variation of one master concept. If two briefs feel like they could share a hook, they're the same concept and you've collapsed the segment grain.
3. If using talking-head video, persona-profile per concept. Build a persona spec for each spokesperson: age, voice, look, accent, why they care about the product. The persona is part of the concept — switching personas on the same hook is a different concept; switching hooks on the same persona is a micro-variation.
4. Generate concept-tied assets via the AI pipeline. nano-banana for keyframes, Veo / Kling for video, voice generation per persona. Each concept's assets stay within its concept family — do not cross-pollinate, that's how micro-variations sneak back in.
5. Ban cosmetic edits. No "let's try the same creative with a different CTA button". If you want to A/B that, you're below Andromeda's resolution and you're spending budget on slots Meta has already collapsed.
6. Set every Meta AI checkbox to on. Targeting 18+ all-genders with language + country only. The creative carries the segment signal; Andromeda routes it. Old "I'll force my audience" targeting actively fights the auction.
7. Validate against the published Meta engineering posts. When you're stuck on whether a new concept is "different enough", re-read the two engineering.fb.com posts (linked below) and ask: would Andromeda's retrieval engine pick this asset for a different segment auction than the parent? If no, it's a micro-variation.

The canonical Meta engineering references this framework is built against:

Andromeda — retrieval engine that powers next-gen personalized ads:
https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/

GEM — Generative Ads Model, the central brain accelerating ads recommendation AI:
https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/

The audience-first framework:

1. Audience profile (segment-level grain Andromeda auctions inside)
2. Creative concept per audience (story, hook, location)
3. Persona profile per concept (when a talking-head spokesperson is used)
4. Concept-tied assets (no cross-pollinating hooks across concepts)

Banned: hook-swap, button-color, CTA-text — Andromeda collapses these into
 one concept inside the segment auction.

• Old "I'll force my audience" targeting fights the auction. Tight demographics + interest stacking made sense when Meta couldn't auction on creative; now they actively starve the algorithm. 18+ all-genders + language + country is the working default in source.
• Hook swap = same concept. A creative with the original hook replaced is not a new concept — Andromeda hashes it together with the parent. If you would have called it "variation B" under the old playbook, don't ship it.
• AI-generated video isn't yet universally winning. One operator on consumer apps + web subscriptions: AI-generated video dominates the top of their ranking. A different operator on e-commerce $1M+ brands: zero AI-generated video in their top 20. Pick your stack to your category, not to the chat consensus.
• FTC fake-review rule applies to AI testimonials. Generating an AI "person who never used the product" testimonial is a regulatory violation in the US (FTC's 2024 final rule). Static images, b-rolls, variations, animated visuals are all fine; imitating a real customer who never had the experience is not. One operator's framing: "where you can use AI, use it — where you legally can't is testimonial fabrication."
• AI-disclosure labels can drop CTR ~33%. Meta has signaled it may auto-label AI-generated content. Plan a budget reallocation if the rollout hits, and pivot AI-generated portions toward "actor portrayal" framings where applicable.