If you're running separate ad sets for Feed, Reels, and Stories, you're paying for learning that Meta's system no longer needs to do separately. Meta Lattice — the unified ML ranking architecture now running across Meta's entire ad delivery pipeline — transfers performance signals across placements automatically. The placement-split structure that made sense in 2022 now works against you.
In this post:
- What Meta Lattice is and how it replaced hundreds of siloed models
- How it differs from Andromeda and GEM in Meta's AI stack
- Why placement-specific campaigns create redundant learning phases you're paying for
- What to rebuild for the Lattice era
What Meta Lattice Is
Before Lattice, Meta ran hundreds of separate prediction models — one per combination of placement (Feed, Reels, Stories, Search) and objective (clicks, video views, purchases, leads). Each model trained independently. A creative that performed well on Reels shared nothing with the Feed model. Every new surface had to accumulate its own training data from scratch.
Lattice is a single Multi-Domain, Multi-Objective architecture that replaced all of them. When a creative drives conversions on Reels, Lattice immediately incorporates that signal into its Feed and Stories predictions. Cross-placement learning is continuous and simultaneous. There's no warmup period per placement because there's no longer a per-placement model.
The numbers from Meta's engineering paper, accepted at KDD 2026: Lattice delivered a 10% improvement in revenue-driving metrics, a 6% increase in conversion rates, and a 20% reduction in compute capacity needed. These are platform-wide production figures from a completed rollout, not lab benchmarks.
Where Lattice Fits: The Three-Layer Stack
Lattice is one of three distinct AI systems now running in sequence for every ad auction.
Andromeda (retrieval) acts first. With potentially millions of eligible ads for any given user, Andromeda narrows this to a viable candidate set — deciding which ads are even considered for the auction. Meta's AI Blog describes it as a co-designed hardware-software-AI system that delivered an 8% ad quality improvement when introduced.
Lattice (ranking) acts second. It takes Andromeda's shortlist and determines which ad wins — ranking candidates across all objectives simultaneously using a single model rather than objective-specific models for each outcome. This is where the cross-placement signal transfer happens in real time.
GEM (teacher model) runs in the background. As covered in Meta GEM: What It Means for Creative, GEM is too computationally heavy to serve ads in real time. Instead it functions as a teacher: trained at scale across Meta's full ecosystem, it distills knowledge into both Andromeda and Lattice continuously.
The pipeline:
GEM (background) → Andromeda (retrieval) → Lattice (ranking) → Ad servedUnderstanding these as separate systems matters because they have different implications for how you should structure accounts. Lattice — specifically its cross-placement signal sharing — is what makes placement-specific campaigns a structural liability.
Why Placement-Specific Campaigns Now Hurt You
The old logic was defensible: different placements have different audience behaviors, so segment them and optimize separately. That was correct when separate models existed. It no longer is.
Running placement-restricted ad sets today creates problems that Lattice was built to eliminate:
Redundant learning phases. Each ad set restricted to a single placement must independently accumulate enough conversion signal to exit learning. You're paying for that data acquisition repeatedly, surface by surface — when Lattice is already sharing signals across all of them automatically.
Data fragmentation. Splitting spend across placement-specific ad sets divides your conversion data into smaller streams. Lattice learns best from pooled signals. Fragmentation means each individual stream is weaker — the opposite of what drives the system's optimization.
Placement lock-in where the system wants flexibility. Lattice's cross-surface allocation finds where your spend performs best in real time. Manual placement restrictions force it to optimize within an artificial constraint rather than using the full distribution it's designed to manage.
Brainlabs confirmed this pattern from client accounts in early 2026: "Performance on Meta is no longer primarily driven by what you tweak in Ads Manager. It's driven by what the system can learn from you."
What to Rebuild
The structure that works with Lattice rather than against it is simpler than what most teams are running.
Consolidate placement-specific ad sets. Merge them into consolidated campaigns with Advantage+ placement enabled. The system reallocates spend across Feed, Reels, and Stories in real time based on Lattice's cross-surface signals — better than any manual split you can configure statically.
Shift focus to inputs, not levers. Three inputs move the system now: signal quality (first-party data, clean Conversions API implementation, strong event match quality), creative diversity (multiple formats, hooks, and angles so the system can identify what resonates across surfaces), and measurement setup. The structure becomes the constraint only when the inputs are weak.
Prioritize Reels-native creative. Because Lattice transfers learning across surfaces, a creative that wins on Reels signals back to Feed and Stories predictions. Vertical, hook-first creative built for Reels now benefits broader delivery — it's feeding the cross-placement model, not just a single surface.
Measure at account or campaign level. Lattice reallocates spend across ad sets continuously. Ad-set-level ROAS figures are increasingly misleading when the system is moving budget underneath your reports. The meaningful measurement frame is account-level performance, ideally validated with incrementality testing rather than last-click attribution.
| Old structure | Lattice-era structure |
|---|---|
| Placement-specific ad sets | Consolidated campaigns, Advantage+ placement |
| Objective-per-ad-set optimization | Single campaign across objectives |
| Ad-set-level ROAS monitoring | Account-level performance + incrementality |
| Manual placement weighting | System-managed allocation |
| Surface-specific creative briefs | Reels-first creative that transfers across surfaces |
This is where bulk fits into the restructuring workflow. bulk reads your live Meta account — campaigns, creatives, current structure — and surfaces the placement-fragmented setups limiting what Lattice can learn. Rebuilding campaigns is the mechanical part; the value is doing it across an account without touching each one by hand.
The Rollout Is Already Complete
Lattice isn't a roadmap item. The KDD 2026 paper documents a completed production rollout, and its effects are already reflected in platform-wide performance figures. Advertisers running consolidated campaigns and Advantage+ placement are operating inside the Lattice environment. Those running placement-split structures are paying a learning tax on a system that no longer requires that separation.
The adjustment is straightforward. Lattice-era campaigns are simpler to manage, not more complex — fewer ad sets to monitor, fewer placement-specific budgets to babysit. The system handles the allocation. Your job is to give it better inputs and stop getting in its way.
bulk handles campaign management for Meta ads teams — reading your account structure, proposing consolidations, and executing the rebuild with your approval. See how it works →