Built for Andromeda.
Won on creative.
The Alpaca Ads methodology for structuring an ecommerce Meta Ads account in 2026. A consolidated, full-funnel system designed so Meta's retrieval engine gets the cleanest signal and the widest creative pool to draw from.
The old playbook now works against you.
Segmented targeting fought for control the algorithm didn't yet have. Under Andromeda it does, and the old tactics now starve the very system built to out-target you.
Many narrow ad sets
Splitting cold / warm / hot into separate ad sets fragments your conversion signal into pools too small to learn from.
Interest & lookalike stacking
Hand-built audiences now constrain retrieval. They fence Andromeda out of buyers it would have found on its own.
ABO everywhere, heavy exclusions
Manual budget control and exclusion layers throttle delivery and slow the model's path out of learning.
Consolidate the account, broaden the targeting, and pour your effort into creative volume and diversity. Every decision in this playbook traces back to that one line.
Meta's ads retrieval engine, announced December 2024, fully rolled out worldwide by October 2025. It is the first stage of delivery: before an ad reaches ranking or the auction, Andromeda sifts tens of millions of eligible ads down to a small, highly personalised candidate pool for each user. Running on NVIDIA Grace Hopper hardware, it uses deep neural networks, computer vision and semantic analysis, to read the actual content of every creative, cluster ads by what they contain, and match those clusters to a user's real-time behaviour. Two consequences drive everything below: creative is now the targeting (every distinct creative gets its own entity ID and becomes an independent shot on goal), and fragmentation starves the engine (consolidation concentrates signal and exits learning faster).
Five principles behind every decision.
Consolidate ruthlessly
Fewer campaigns, fewer ad sets. Every ad set should be on track for 50+ conversions a week. If it can't be, it shouldn't exist as a separate ad set.
Broad is the default
No interests, no lookalikes, minimal exclusions in prospecting. Let Andromeda find the buyer. It has more signal than any audience you can build by hand.
Creative volume is the growth lever
8–15+ distinct creatives live in prospecting at all times. Diversity of concept, not variations of one ad. Concepts are how Andromeda matches you into new pockets of demand.
Protect the signal
One clean dataset, Conversions API running server-side, Purchase as the optimisation event. The model is only ever as good as the data you feed it.
Stop touching it
Edits reset learning. Make changes in deliberate weekly windows, not daily. Patience is a performance setting.
Get the signal right first.
Andromeda's matching is only as good as the conversion signal it learns from. None of the structure below performs if the plumbing is wrong.
One dataset + CAPI server-side
One pixel/dataset per brand, Conversions API running server-side alongside the browser pixel, deduplication confirmed. Server-side events survive iOS signal loss.
Domain verified + AEM
Domain verified and Aggregated Event Measurement configured: 8 events ranked, with Purchase as the #1 priority event.
Catalog connected
Advantage+ catalog connected and the product feed healthy. Needed for catalog ads in retargeting and increasingly used inside Advantage+ Sales itself.
Attribution: 7-day click / 1-day view
Keep the default attribution window, and report on it consistently so every read is comparing like for like.
Creative enhancements reviewed
Advantage+ creative enhancements reviewed at account level. Keep the safe ones on, override per-ad where they distort the creative.
Naming convention locked
Lock the naming convention before launch so creative-level reporting is legible from day one. Renaming live entities is one more edit you don't want.
Four campaigns. That's the whole account.
A consolidated full-funnel build. Each campaign has a single job, and broad targeting runs through all of them.
Prospecting, Advantage+ Sales
Where the majority of spend and nearly all new-customer acquisition happens. Only proven and promising creative lives here. This is not where cold ideas get tested.
Creative Testing, Sandbox
The one place manual structure still wins. Its job: validate net-new creative before it influences the engine campaign. Every batch also teaches Andromeda which messages and formats convert for the brand.
Retargeting, MOFU
With the existing-customer cap already doing some retargeting inside Campaign 01, this campaign is deliberately lean. It guarantees coverage of high-intent warm audiences with the right format.
Retention, Existing Customers
Past customers convert at the highest rate of any audience, but only if you target them deliberately instead of letting the prospecting campaign burn budget on them.
Naming conventions.
Consistent names are what make creative-level reporting legible, and creative is where every decision now lives.
One structure. Every budget.
The four-campaign architecture is the target state. At lower spend you collapse toward it; at higher spend you extend it, but you never fragment it.
Lean
Prospecting Advantage+ Sales + Testing. There isn't enough signal to feed dedicated retargeting or retention.
- Run the existing-customer cap higher (~35–40%) so the engine absorbs retargeting + retention
- Drop Campaigns 03 and 04
Mid
The full architecture exactly as specified above: prospecting, testing, retargeting and retention all running.
- Existing-customer cap at 25–30%
- Dedicated testing sandbox feeding winners up
- Lean catalog retargeting + small retention line
Scaling
Extend, never multiply. Options in order of preference:
- Increase creative throughput first. More batches, more live ads, faster refresh
- Add a 2nd Advantage+ Sales campaign only for a genuinely distinct lever (geo / offer)
- Add a cost-capped scaling campaign for proven top creatives
The mistake at every tier: duplicating ad sets to "scale." Raise budgets on what works in 20–30% steps and add creative. Duplication just splits the signal you worked to concentrate.
Under Andromeda, creative is the media plan.
The structure only performs if it's fed. Andromeda reads creative content and clusters by it. Near-duplicates collapse into the same cluster and compete for the same users.
Different concepts, not different colours. Keep genuine variety across concept: UGC, founder/story, social proof & reviews, product demo, problem→solution, offer-led, lifestyle, comparison, and across format: static, video, carousel, catalog. Build 8–12 distinct concepts at any time, each with 2–3 variations.
The graduation pipeline
Launch a concept batch in Testing. ABO ad set, ~7+ conv/day budget.
Let it run a full 3–4 days minimum and reach learning-phase stability before judging.
Promote when it beats the account benchmark with a trustworthy read, ~15–30+ conversions.
Rebuild the ad fresh inside Campaign 01 so it inherits the engine's signal.
Cut losers cleanly. Hold "maybes" for one iteration, new hook, and re-batch.
Where the money goes.
A mid-tier reference split. Adjust to LTV and retention strength. Ranges overlap by design; these are guides, not rules.
Bidding, start loose
Start every campaign on lowest-cost (highest volume). Introduce a cost cap only once you know your true allowable CPA and are deliberately scaling. Caps set too early just throttle delivery and stall learning.
Learning, clear the bar
An ad set needs ~50 conversions in 7 days to exit learning. Budget each campaign to clear that bar. If it can't, consolidate further. Expect learning to reset after any meaningful edit.
Scaling, go vertical
Raise budget in 20–30% increments, then hold 3–4 days. Vertical scaling (more budget on the winner) beats horizontal scaling (more ad sets) under Andromeda almost every time.
Read it right. Touch it less.
Read at campaign & creative level, not ad-set level
Audience selection is the algorithm's job now; the ad-set view tells you little. Creative performance is the real decision surface.
Read blended
In-platform attribution under-reports. Pair it with MER / blended ROAS and new-customer CAC as the true north-star metrics.
Respect the no-touch window
After launching or editing, leave it 3–4 days before judging or adjusting. Never judge mid-learning-phase or on too few conversions.
Defensible in 2024. Costing you now.
Every one of these once made sense. Under Andromeda they all actively work against the engine.
Splitting prospecting into multiple interest / lookalike ad sets. Fragments signal, constrains retrieval.
Stacking detailed-targeting interests in cold campaigns.
Heavy exclusion layering in prospecting beyond the existing-customer cap and basic brand safety.
Duplicating ad sets to "scale."
Tiny budgets spread across many ad sets. None reach 50 conversions/week.
Frequent edits and budget swings over 30%. Each one resets learning.
Testing minor variations (colour, copy tweak) instead of distinct concepts.
Judging performance mid-learning-phase or on too few conversions.
Running creative testing inside the main Advantage+ campaign. It pollutes the engine's read. Exactly why Campaign 02 is separate.
From zero to rhythm in four weeks.
A clean launch sequence. Foundations first, core campaigns next, funnel coverage once the engine is out of learning.
Foundations
- Confirm CAPI server-side + dedup
- Domain verification
- AEM events, Purchase #1
- Catalog feed health
- Lock the naming convention
Launch core
- Campaign 01, Advantage+ Sales, broad, cap 25–30%, 8+ ads
- Campaign 02, Testing, first 2–3 concept batches
- Don't touch for 3–4 days
Add funnel coverage
- Launch Campaign 03, retargeting
- Launch Campaign 04, retention
- First graduation reviews from Testing
Establish rhythm
- Run the weekly loop
- Read at campaign + creative level
- Promote / cut · queue next batches
- Refresh ads nearing fatigue
This is how Alpaca Ads
runs Meta.
Use it as the default structure for every ecommerce account. Adapt only with a reason, and write the reason down.
Paid Media Methodologies · Meta Ads · v1.0