Three Systems Now Decide Your Meta Ads: Andromeda, GEM, Lattice

Meta Lattice decided your last auction, and you never touched a setting. Same budget, same creative, same audience — different result, because three separate AI systems now sit between your campaign and every impression, and each one does a different job. None of them are optional, none of them show up in the dashboard by name, and most buyers have never had a reason to tell them apart until delivery started acting strange for no visible reason.

Meta ad delivery in 2026 runs through three stages, not one algorithm: Andromeda retrieves the eligible ad shortlist, GEM scores predicted engagement, and Lattice ranks and prices the auction. Every advertiser has run on this stack since October 2025, whether they opted into Advantage+ or not.

That is the complete answer. What follows is what each stage does, and which lever, creative volume, signal quality, or bid strategy, you should pull depending on where the actual bottleneck sits.

Three systems now decide what your ad shows

Buyers talk about "the algorithm" as one thing. It isn't. Meta's current ad-serving stack, covered in more depth in how the automation stack actually works, is built from three named, separately documented systems, each handling a distinct stage of the pipeline that runs before an ad ever renders in a feed.

Reporting from Search Engine Land names them as Andromeda, GEM, and Lattice — retrieval, ranking-signal, and ranking/auction, in that order. Meta has not published a detailed diagram of exactly where one system's job ends and the next one's begins, so treat this as three stages rather than a fully documented pipeline.

Andromeda decides what is even eligible to compete

Andromeda is the retrieval stage — the gatekeeper. Before any ranking or bidding happens, Andromeda pulls the shortlist of ads even eligible to be considered for a given impression, out of the full advertiser catalog. It completed its global rollout in October 2025.

If your ads aren't showing up at all for a segment you'd expect them to reach, Andromeda's retrieval pass is the more likely first suspect — not a bid that's too low. This is the stage our Andromeda explainer covers in more depth; worth reading if retrieval, not ranking, is your actual problem.

GEM scores engagement before anyone bids

Once Andromeda has a shortlist, GEM (the Generative Engagement Model) scores it. GEM is a ranking-signal model: it predicts how a person is likely to engage with each candidate ad and feeds that prediction downstream. Meta has described it as roughly 4x more efficient than the system it replaced. GEM began rolling out in Q2 2025, and Meta published a technical paper on the model on 2025-11-10.

"More efficient" here is Meta's own framing of GEM's compute-to-signal ratio, not an independently verified benchmark — worth saying plainly, since a lot of the trade coverage repeats the figure without that caveat.

What Meta Lattice actually does

Lattice is the ranking and auction layer. It's the stage that takes GEM's engagement-scored shortlist and actually runs the auction — ranking candidates and settling on price. Where Andromeda decides who's in the room and GEM scores how interesting each ad probably is, Meta Lattice is where the room gets sorted and the bill gets written.

It's the newest of the three names to show up in changelogs and conference recaps, and it's also the stage most directly downstream of your bid strategy. A bid change doesn't touch Andromeda's retrieval or GEM's engagement prediction — it changes what happens once Lattice runs the auction. If you've been treating "the algorithm" as a single black box, Lattice is usually the part you're actually negotiating with.

How Andromeda, GEM, and Meta Lattice hand off

In sequence: Andromeda retrieves the eligible shortlist for an impression, GEM scores predicted engagement across that shortlist, and Meta Lattice ranks the scored candidates and runs the auction that decides who wins and what they pay. Each stage narrows or reorders what the previous one produced.

Every advertiser runs through all three regardless of whether they've opted into Advantage+. There's no toggle that routes you around Andromeda, GEM, or Lattice. The opt-in choice in Advantage+ campaign setup, explained in when AI buying earns budget, controls how much decision-making Meta's AI takes over on top of that serving stack: targeting, placement, budget. It never controls whether the stack itself runs.

That distinction matters for the broader convergence happening across platforms right now — Meta isn't alone in moving ad serving onto layered retrieval-and-ranking systems, but its version is the one with three publicly named stages.

What this stack implies for your account

Three systems means three different diagnostic questions, and mixing them up wastes testing budget.

• If ads aren't delivering at all: Andromeda's retrieval pass is filtering you out before ranking even starts. Check catalog eligibility, policy flags, and whether your Advantage+ Audience settings are excluding the segment, before touching bids. It's also worth reviewing how Advantage+ actually allocates budget — a retrieval problem can look like a budget problem from the reporting screen.
• If delivery is inconsistent or engagement-weighted oddly: that's GEM's territory. Creative volume and variety feed GEM's engagement predictions directly, and a thin creative set gives it less to score against. Our creative volume playbook covers building that variety without diluting quality; adlibrary's ad enrichment makes it faster to see which formats competitors are already rotating through at volume. A creative refresh can also trip a fresh learning phase — if delivery stalls right after a swap, our diagnostic steps are the next stop.
• If you're winning impressions but losing the auction on cost: that's Lattice. Signal quality and bid strategy are the two levers that actually move Lattice's output. On signal, Lattice's auction runs on the conversion signal you're feeding it, not just the engagement score GEM predicted, which is why your Conversions API setup, walked through in our implementation guide, and an EMQ score check tend to move the needle faster than a bid change. On bid, our Meta bid strategy guide breaks down which controls still do something meaningful once Lattice is doing the ranking.

Structurally, this is also why the Marketing API deprecation that finished on 2026-05-19 mattered beyond a tooling inconvenience. Meta blocked new ASC/AAC campaign creation starting with v24.0 (2025-10-08), then blocked it across all API versions in v25.0 (Q1 2026).

Agencies and tools had to rebuild on Advantage+ campaign objects, including Meta Advantage+ Sales, because those are the objects that route into the Andromeda/GEM/Lattice stack Meta now runs everything through — our Meta Ads API tools comparison covers which vendors made that rebuild cleanly. adlibrary's own API access sits outside that dependency entirely, since it reads the public ad corpus rather than an advertiser's Marketing API account.

What you still control

The three-system stack narrows some levers and leaves others fully in your hands.

You still control creative supply: how many distinct ad variants you feed the system, and whether they're genuinely differentiated rather than four crops of the same asset. Our use-case walkthrough for creative iteration covers building that loop without burning a production team.

You still control the broad targeting versus custom-audience split, even after Meta pulled a number of detailed-interest categories in January 2026 and narrowed Advantage+ Shopping Campaigns, or ASC+, to retargeting plus AI discovery, per eMarketer. And you still control signal quality end to end. Conversions API setup, event deduplication and match rate are arguably the most underused lever of the three, since they're the one input GEM and Lattice can't compensate for on their own.

What you've lost some of is creative control at the margins. Agency complaints, reported by the FT and covered by Mamba Digital, center on Advantage+ Creative silently modifying submitted assets: cropping logos, adding music, rearranging text, without a clear opt-out at the asset level. That's a real complaint from real agencies, not a hypothetical. If you've noticed a submitted creative rendering differently than you built it, this is the likely cause, not a bug on your end.

I've watched this play out on accounts where the account team assumed a CPA drift was a bid problem, spent two weeks adjusting bid caps, and it turned out to be a stale pixel feeding degraded signal into what is now Meta Lattice's auction. The fix was a CAPI dedup pass, not a bid change — a five-minute diagnosis once you know which of the three systems you're actually looking at.

This three-system stack isn't a reason to stop testing bid strategy — it's a reason to test the right thing first. Figure out which stage is actually bottlenecking your delivery before you spend budget on the wrong lever.

Meta Lattice: frequently asked questions

What is Meta Lattice? Lattice is the ranking and auction layer in Meta's ad-serving stack — the system that takes the engagement-scored shortlist from GEM and runs the auction that decides which ad wins an impression and at what price.

Is Lattice the same thing as Andromeda? No. Andromeda is the retrieval stage that decides which ads are even eligible to compete for an impression. GEM scores predicted engagement across that shortlist. Lattice ranks the scored candidates and runs the auction. They're three distinct, sequential systems.

Do I have to opt in to Andromeda, GEM, or Lattice? No. Every advertiser has run through this serving stack since Andromeda's global rollout completed in October 2025, regardless of whether Advantage+ automation is toggled on. Opting into Advantage+ changes how much AI decides on top of the stack, not whether the stack runs.

What does GEM stand for? GEM stands for Generative Engagement Model. It's the ranking-signal system that predicts engagement for each ad candidate before Lattice's auction runs, and Meta describes it as roughly 4x more efficient than its predecessor — a figure that comes from Meta itself, not an independent benchmark.

How do I know which system is limiting my results? Work backward from the symptom: no delivery points to Andromeda's retrieval, uneven or engagement-skewed delivery points to GEM, and winning impressions but losing on cost points to Lattice's auction. Each has a different fix — eligibility and targeting for the first, creative volume for the second, signal quality and bid strategy for the third.

Three systems, three different problems, one account. Meta Lattice is usually the fastest one to act on once you've ruled out the other two. Knowing which of the three you're fighting is the whole diagnosis, and everything after that is picking the lever that actually moves the number you're staring at.

Want to see how competitors are structuring creative volume for this stack? Start a free trial or check pricing.