When Lattice changed everything about creative strategy

Lattice is the system Meta used to unify ads ranking: one model in place of hundreds, split by objective and by surface. Cross-domain learning means a creative that performs on Feed learns from Reels, and that changes how you think about message consistency across placements.

Note on sources, added 29 August 2026. This note is dated 1 February 2025, and the first version said I had studied the Lattice paper. That wasn't possible: the Meta for Business announcement on ads ranking is from March 2025, and the paper on arXiv was filed on 9 December 2025. I fixed the sentence and rewrote the section that credited the paper with things it doesn't contain. From here on, what the paper says and what I take from it stay separate.

What Lattice replaced and why it matters

Before Lattice, Meta ran separate models for each surface: Feed, Reels, Stories, Marketplace. Each one optimized independently. The result was that a creative could excel on Feed and end up invisible on Reels without the system understanding why.

Lattice unifies everything under a shared ranking system. Engagement signals on one surface feed the model that serves the others. A video that generates high watch-through on Reels passes information to the model that decides what to show on Feed — and vice versa.

The practical implication: there's no longer a "Feed creative" and a "Reels creative" as separate categories. There's a creative that works or doesn't work, and the system figures that out across every surface in parallel.

The practical implications for creative production

Message consistency across placements isn't an aesthetic question anymore. It's a signal question.

If your hook on Feed says "save 30%" and the Reels version of the same product goes after identity ("for people who don't want to look like everyone else"), you're sending Lattice conflicting signals about what that creative stands for. The system doesn't know which angle to attribute to the audience that converts.

The rule I apply: one narrative angle per set of variants. Formats change — ratio, length, text overlay — but the psychological trigger and the promise stay consistent. That lets Lattice accumulate consistent signals on that angle instead of scattering them across incompatible narratives.

Creative production no longer starts from "what do I make for Feed" but from "what's the angle, and how do I adapt it to each surface without betraying it."

What the paper says, and what it doesn't

The paper is Meta Lattice: Model Space Redesign for Cost-Effective Industry-Scale Ads Recommendations, arXiv 2512.09200, filed on 9 December 2025 and accepted at KDD 2026. It describes a redesign of the model space that puts together "cross-domain knowledge sharing, data consolidation, model unification, distillation, and system optimizations", and it reports the numbers from the deployment inside Meta: 10% on revenue-driving top-line metrics, 11.5% on user satisfaction, 6% on conversion rate, 20% capacity saved.

What the paper leaves out counts as much as what it contains. The word creative never appears, not once in the whole document. Neither do diversity, redundancy or entity. Where the paper says diverse, it's talking about input formats and domains, not about ads. It's a piece of work about architecture and infrastructure cost, signed by forty-two authors and written for people who build recommendation systems.

The Lattice paper never uses the word creative. Everything you read about what Lattice does to creatives is somebody's reading, not documentation.

So from here on I'm the one drawing operational consequences out of an architecture, and I give them for what they are. First: if the system builds a shared representation of creative and audience that holds across surfaces, then the material you hand it in the early window counts more than it used to. On the accounts I run, sending variants with different angles, contradictory tones and incompatible promises into that window slows the exit from the learning phase. That's an observation, not a measurement: I have no publishable data series that proves it.

Second consequence, also mine: two semantically identical creatives don't add up in reach. They compete for the same slot and the system distributes one of them. That's the reasoning behind what I call Entity ID diversity, which I covered separately — with the same warning about sources.

How to audit your creative strategy against Lattice principles

The audit starts from a simple question: how many distinct narrative angles do you have live right now?

Not variants. Not formats. Angles. An angle is a combination of psychological trigger + promise + product frame. "Saving money" and "social status" are different angles. "Saving money with social proof" and "saving money with urgency" are variants of the same angle.

Do this mapping across all your live creatives. If you have 15 ads and 2 angles, you have a semantic diversity problem, not a volume problem. The system treats those 15 ads as 2 creatives with many instances — and it charges you for the redundancy in CPM.

The benchmark I use: at least 3-4 distinct angles for every stage of the funnel, with a minimum of 2-3 format variants per angle. Below that threshold, you're limiting Lattice's exploration.

How campaign structure changed with Advantage+ and broad audiences

The way I read it, Lattice is also why Advantage+ Shopping does better with high creative diversity: the system needs heterogeneous material to match precisely, because different audiences respond to different angles. Meta doesn't write it anywhere in those terms, so take it as a working hypothesis that has held up for me, not as a fact.

With broad audiences — no interest targeting, no narrow custom audience — the only targeting lever left in your hands is the creative itself. The angle you pick determines who you attract. Lattice uses the content of the creative as an implicit targeting signal.

In practice: in an Advantage+ Shopping campaign with a broad audience, having 4 distinct angles is the same as having 4 different implicit targeting setups. The system distributes each angle to the audience that responds. This is targeting by creative, not targeting by parameter — and it's the direction Meta is pushing everything in.

I applied this system on a fashion project with documented results at 5.9x ROAS. Creative diversity wasn't an add-on — it was the central performance mechanism.

Changelog

29 August 2026: fixed the sentence that claimed I had studied the paper, published ten months after the date on this note. Added the primary sources: Meta's announcement of March 2025 and arXiv paper 2512.09200 of 9 December 2025. Rewrote the section on the paper, keeping what the document states apart from what I take from it, and flagged that the word creative doesn't appear in the paper.

---

If you want to work on your creative strategy taking Meta's real architecture into account — Lattice, Andromeda, Advantage+ — book a 1-to-1 strategy session. We analyze your account and build a system that works with the system, not against it.