Three months ago, every performance marketer I know was arguing about whether Advantage+ was worth running. Today, the argument is different: Meta’s Ranking Engineer Agent (REA) is building new ranking models autonomously, Andromeda is using creative as the primary targeting signal, and Advantage+ keeps swallowing ad sets. That’s three different AI systems — and nobody has explained how they work together, which ones you actually control, and what order to deploy them in. So here’s the playbook.
TL;DR: Meta now runs three layered AI systems: REA (builds the ranking models), Andromeda (serves the creative), and Advantage+ (the advertiser UI). Meta’s own 31-test study shows Advantage+ Shopping delivers a 32% ROAS lift vs. business-as-usual (Social Media Today, 2023). Start at Advantage+, feed Andromeda diverse creative, and stop fighting REA.
What is the Meta AI agent stack?
The Meta AI agent stack is three layered systems: REA builds and improves the underlying ranking models autonomously, Andromeda retrieves and scores candidate ads in real time using creative features, and Advantage+ is the advertiser-facing UI that lets Meta’s AI decide targeting, placements, budgets, and creative rotation. Meta Engineering disclosed on March 17, 2026 that REA “doubled average model accuracy over baseline across six models,” with three engineers delivering proposals for eight models historically handled by two engineers per model (Engineering at Meta, 2026).
Most media buyers only think about the Advantage+ layer because that’s where the buttons live. That’s the mistake. The decisions you’re making inside Advantage+ are constrained by what Andromeda can retrieve, which is constrained by what REA has learned to rank. If you don’t know the stack, you’re pulling levers without knowing which layer you’re actually on.
Here’s the simplest way I frame it with clients: Meta built the AI stack bottom-up, but advertisers interact with it top-down. The button you clicked in Ads Manager is the last three millimeters of a stack that’s 100 miles tall. Treat it that way.
How do REA, Advantage+, and Andromeda fit together?
They stack vertically. REA sits at the foundation — an autonomous AI system that proposes, tests, and validates improvements to the ranking models Meta’s ads auction uses. Andromeda sits above REA as the retrieval engine — it reads creative features (video pace, copy embeddings, format) and matches them to user embeddings to decide which ads are even eligible for the auction. Advantage+ sits on top as the API surface advertisers actually touch. Andromeda alone delivered “+6% recall improvement and +8% ad quality on selected segments” when Meta rolled it out (Engineering at Meta, 2024).
The chart is the whole argument. Your lever budget shrinks as you move down the stack. If you want leverage, you have to spend it at the layer where you actually have influence — which is almost entirely the creative you feed in, the audience suggestion (not constraint) you set, and the budget you commit. Everything else is happening somewhere you can’t see.
I’ve written the full decoder for how Andromeda reads creative in my decoder for how Meta’s Andromeda algorithm reads creative, and a longer explainer on how the Andromeda algorithm works. If this is the first time you’re hearing any of this — start there.
When should I use each layer?
You don’t pick REA or Andromeda — they run whether you want them to or not. The only real “when” question is at the Advantage+ layer. Meta’s own 31-test global study found Advantage+ Shopping delivered a 32% increase in ROAS and a 17% decrease in CPA versus business-as-usual campaigns (Social Media Today, 2023, citing Meta’s internal study). That’s a floor, not a ceiling — but it sets the burden of proof at the right place.
My rule: default to Advantage+ Shopping for most DTC accounts doing $1M-$50M, and require your manual setup to justify itself against the Advantage+ baseline. Not the other way around. If you can’t beat Advantage+ with your own audience craft, you’re adding human labor for negative output. That’s not media buying — that’s resistance.
In my accounts, the decision I re-run every quarter isn’t “Advantage+ or manual” — it’s “how much of the budget belongs in Advantage+ Shopping vs. Advantage+ Audience on broader prospecting campaigns.” That’s a real operator question. The “should I use Advantage+” argument is over; that was 2024.
My full setup walkthrough for the campaign layer is in the complete Advantage+ Shopping Campaigns guide. For the audience layer, see broad targeting + Advantage+ Audience strategies for 2026.
Where do the three layers conflict?
They don’t conflict with each other — they conflict with the advertiser. The friction is almost always the same: operators trying to assert control at a layer where Meta has already taken it. Foxwell Digital’s analysis of Motion’s 550K-ad study found “about 5% of ads drive the majority of performance” across the dataset (Foxwell Digital, 2026; Motion, 2026). That’s your actual job: feeding Andromeda the kind of creative diversity it needs to find the 5%.
The conflict shows up in three predictable places. First, ad-set targeting: buyers try to narrow what Andromeda already knows how to expand. Second, ad-level rotation: buyers turn off the 4 “losers” in a 5-ad ad set before Andromeda has enough data to converge. Third, measurement: buyers attribute performance to the ad-set name instead of the creative, missing what Andromeda is actually doing. The answer to all three is the same — give the stack more creative, not more controls.
What’s the decision order for DTC accounts?
The playbook I run (and recommend) follows a strict order. Most accounts skip steps 2 and 3 and start optimizing at step 5, which is why they keep fighting the stack. Stop at the first step that actually gates your performance — that’s where your leverage lives.
- Creative diversity check. Do you have 6-10 genuinely different creatives across 3-5 distinct concepts in market? If no, stop here. The stack cannot find the 5% winners from 10 variants of one concept.
- Advantage+ Shopping on. Default for retargeting and prospecting at the purchase event. If you’re still running manual sales campaigns, the 32% ROAS and 17% CPA benchmark is the burden your manual setup has to beat.
- Advantage+ Audience on broad prospecting. Feed it a single broad audience with light demographic guardrails, not interest stacks. Let Meta narrow, not you.
- Signal hygiene. Conversion API, value optimization, 7-day click if your volume supports it. Andromeda ranks on signal quality — feed it good signal.
- Campaign structure. Consolidate. One Advantage+ Shopping campaign for purchase, one Advantage+ Audience for prospecting, one retargeting layer. Not 14 ad sets with overlapping audiences.
- Budget leverage. Put the scaling question at the account or Advantage+ campaign level, not the ad set. Meta’s AI agents inside Ads Manager will push budget where Andromeda signals strength — your job is sizing the pot.
- Creative velocity. Two to four new concepts every two weeks, feeding fresh inputs to the retrieval layer. This is the compound-interest lever.
Notice what’s not on the list: manual audience curation, detailed placement controls, dayparting schedules, and interest-layered ad sets. Those are all moves at layers where Meta no longer needs your input. That’s not me being lazy — that’s me reading what the stack is telling us. For a deeper walkthrough of the creative-side system, see my AI-native creative testing system for Meta ads.
What changes for media buyers in 2026?
Zuckerberg and Meta leadership have repeatedly said the goal is “fully automated AI ads by the end of 2026” — advertisers input a URL and a budget, Meta’s AI handles creative, targeting, and measurement (Marketing Dive, 2025). That’s the direction of travel. Meanwhile, the ChatGPT Ads launch in February 2026 at a reported $60 CPM has already eroded to closer to $25 CPM plus a CPC pivot ($3-5/click), telling us exactly how AI-buying auctions price-discover in the wild (Adweek, 2026; The Next Web, 2026).
What that means for your job: the work moves further upstream from the auction. You’re hired to have a creative point of view, a customer-insight point of view, and a measurement point of view — none of which Meta can automate from a URL and a budget, at least not this year. The buyers who’ll win the transition aren’t the ones defending their audience stacks; they’re the ones briefing three diverse concepts every two weeks and letting the stack do the rest.
This is why I’ve written the full cluster strategy in my DTC Meta ads strategy for 2026. The playbook there is the same shape as the one above — it just zooms out one level.
Frequently asked questions
Is REA something I can turn on or off?
No. REA is a Meta-internal AI system — an autonomous agent that proposes and validates improvements to the ranking models behind the ads auction. Meta Engineering published in March 2026 that REA doubled average model accuracy across six ranking models (Engineering at Meta, 2026). Advertisers have no switch; it runs whether you notice or not.
Does Andromeda replace targeting, or does it just change how it works?
It changes how it works. Andromeda is a retrieval engine that uses creative features as primary signals alongside audience signals — Meta Engineering reported +6% recall improvement and +8% ad quality on selected segments (Engineering at Meta, 2024). Your audience setting still scopes the candidate pool. The matching within that pool happens in embedding space.
Should I run Advantage+ Shopping or manual campaigns?
Run Advantage+ Shopping by default. Meta’s 31-test global A/B study showed a 32% ROAS increase and 17% CPA reduction vs. business-as-usual (Social Media Today, 2023). Make your manual setup justify itself against that benchmark, not the other way around.
How many creatives do I need for the stack to work?
Six to ten genuinely different concepts across 3-5 themes — not variants of one concept. Motion’s 550K-ad study found about 5% of ads drive the majority of performance (Foxwell Digital, 2026). If your inputs are uniform, Andromeda can’t find the winners.
What happens when Meta hits “fully automated” in late 2026?
The buyer’s job moves upstream. Meta has signaled a URL-plus-budget automation target by end of 2026 (Marketing Dive, 2025). Your leverage becomes creative, customer insight, and measurement — the things a URL doesn’t give Meta. In-platform knob-pulling will be worth less.
For a deeper dive, see my guide on how to use manus ai in meta ads manager: a practitioner’s 2026 setup guide (and what the china block changes).
For a deeper dive, see my guide on should you opt out of meta’s advantage+ ai creative auto-tweaks? a 2026 practitioner decision framework.
The bottom line
Meta’s AI stack isn’t a single product — it’s three layers stacked on top of each other, and your control surface shrinks as you move down. REA builds the models, Andromeda retrieves on creative, and Advantage+ gives you the handful of levers you actually get to pull. Every quarter from here, the set of things Meta automates gets bigger, and the set of things you control gets smaller. That’s the trend, not a prediction.
The buyers who’ll still have jobs in 2027 are the ones who figured that out in 2026 and spent their energy on the parts of the job the stack can’t do — creative briefs, customer insight, measurement, and portfolio-level budget calls. Everything else is Meta’s problem now. Stop fighting it.