Meta Rebuilt Its Ad Brain Again: GEM and Andromeda in 2026

In February I wrote It’s Not Andromeda for Triple Whale. It covered four AI systems, one ad stack, and why nobody has cracked it.

This month Meta published two big engineering posts, GEM on Oct 13 and Andromeda on Oct 28. The same day as the Andromeda post, Meta for Business put up two plain-English explainers (one, two).

My team and I went through all of it against what I wrote.

Most of it holds, but five things changed. A few things I wrote back then were also wrong, so I’ll own those too.

What was already true and what’s new

A few people on X pointed out that Meta’s Nov 2025 GEM post already talked about engagement feeding predictions. They’re right about the idea. The Oct 13 post takes it a step further.

Already true in Meta’s Nov 2025 GEM post

GEM learned from engagement across surfaces. Meta’s example was Instagram video engagement improving Facebook Feed ad predictions.
Each surface still had its own prediction system, “optimizing each domain’s predictions for its specific objective (such as clicks or conversions).” GEM taught those separate models.

New in Meta’s Oct 13, 2026 GEM post

One central model now predicts clicks across all surfaces and conversions across the major goals. In Meta’s words, “Previously, each surface and each prediction objective had its own system.”
Clicks now train conversion predictions. Meta says “the generalizable representations learned on data-rich click tasks transfer into data-sparse conversion tasks.” Clicks are everywhere and conversions are rare, so what people click now has a bigger say in who Meta thinks will buy.
GEM reads twice as much of each person’s engagement history in production, on a new design called Kunlun that replaced InterFormer.
A new system called Solaris gets GEM’s predictions ready right before each auction.
Meta reports up to 18.9% more conversions and 12% higher click-through rate across surfaces over the past year.

The Oct 13 post writes up gains Meta built over the past year. It doesn’t say anything launched that day.

The short version

GEM still doesn’t score ads live. Its read on you now reaches the live decision.
GEM is now one model predicting across surfaces and objectives.
Meta reports bigger GEM lifts, up to 18.9% more conversions.
Andromeda now scores every eligible ad before it builds the shortlist.
AI-made ads added hundreds of millions of new ads to the pool.

GEM is still the teacher, and Andromeda is still the casting director. What I’d tell you to do in your account barely moves.

1. GEM’s read on you now reaches the live decision

In February I wrote that GEM doesn’t serve ads directly.

That’s still true for live scoring. In its Oct 13 post, Meta says running a model this big inside a live ad request costs too much.

What’s new is a system Meta calls Solaris. In background jobs, it works out GEM’s read on the user and ad pairs a request is likely to need. Then it hands those to the models that do serve ads, right when they need them. If an exact match is missing, it falls back to your recent history or to people like you.

GEM still isn’t scoring ads live. Its notes now reach the models that are.

2. GEM is now one model across surfaces and objectives

GEM learning across surfaces isn’t new. Meta’s Nov 2025 GEM post already described that.

The change is on the prediction side. Meta’s Oct 13 post describes a move from separate systems for each surface and objective to one central model. That one model predicts clicks across surfaces and conversions across the major objectives.

Meta also says what GEM learns from engagement now carries into conversion prediction.

My Disney Corporate analogy fits even better now, with one corporate brain learning from every property and predicting for all of them.

3. Meta reports bigger GEM numbers

These are from Meta’s Oct 13 post. Meta didn’t give a baseline for any of them.

Up to 18.9% more conversions across its ad surfaces and objectives.
12% higher click-through rate across its ad surfaces and objectives.
For targeted users over the past year, conversion lifts of 5.8% on Facebook, 4.1% on Instagram and 1.6% for off-site conversions.
The engagement history GEM reads in production doubled. Meta also mentions 25x. That’s what the infrastructure can handle, which is different from what’s running today.

Why you can’t compare these to my February numbers

The GEM numbers in my Triple Whale article came from Meta’s Nov 2025 post. They were 5% on Instagram and 3% on Facebook Feed, both from Q2 2025.

Those cover a different window and different users, and Facebook Feed is narrower than all of Facebook.

So 4.1% on Instagram now is NOT a drop from 5%. And 5.8% on Facebook isn’t growth from 3%. They measure different things.

4. Andromeda now scores every eligible ad

In February I said Andromeda narrows tens of millions of ads down to a few thousand worth considering for you. That part still stands, and there’s still a shortlist.

What changed is how it gets there. Meta’s Oct 28 post says the new Andromeda scores every eligible ad on the way in, at close to ranking-level quality.

Meta reports:

Over 1 billion ads scored per second on a single GPU server, 100x the 2024 system.
About 300x less compute than the 2024 retrieval model.
2.4% more Facebook conversions since 2025.

One more detail from that post. Advertiser targeting rules now run on the same GPU as the scoring. Hold onto that for the hype section below.

My casting director analogy still works. The casting director just gets through a much bigger pile of tapes before handing over the shortlist.

5. AI-made ads grew the pool

Meta says advertisers used its generative AI creative tools to make hundreds of millions of new ads. That grew the pool of candidate ads to what Meta calls an unprecedented scale. Meta also says Advantage+ automation and generative creative will keep growing it.

My read… a pool that big makes lazy variation even easier to ignore. Three hooks on one body of content get lost faster when hundreds of millions of new ads are in the mix.

What still holds from February

GEM is still the teacher. Meta says GEM teaches the “hundreds of vertical models” that serve ads.
Nobody has cracked it. One central model learning across surfaces and funnel stages makes a one-silo hack even less likely to last.
Neither October post mentions Lattice. Meta still named it as a current system in its Jan 28, 2026 newsroom post.
Tactics that get passed around still stop working fast. In the accounts I’ve worked in, I’ve seen it happen in days, sometimes hours. That’s my experience. Meta hasn’t published a number on it.

What I got wrong in February

A few things in my Triple Whale article don’t hold up. I’d rather tell you myself.

UTIS. I wrote that UTIS calibrates Lattice’s ranking. Meta’s UTIS post is about organic Facebook Reels recommendations. It doesn’t mention ads, so I can’t stand behind that line.
GEM teaching Lattice. I can’t source that one either. Meta says GEM teaches “hundreds of vertical models,” and the Lattice paper never names GEM.
“Hundreds of models into one.” That was wrong. The Lattice paper says it replaced numerous smaller, siloed models with a few larger shared ones.
WhatsApp. I listed it as a surface GEM trains on. Meta names Facebook, Instagram and Business Messaging.
Andromeda’s compute efficiency. I called “tripling” 2026 guidance. On Meta’s Q4 2025 earnings call, CFO Susan Li said extending Andromeda to run on Nvidia, AMD and MTIA chips, along with model innovations, let Meta “nearly triple Andromeda’s compute efficiency.” That was a Q4 2025 result.
LoopFM. Meta says LoopFM roughly doubles how well GEM’s knowledge transfers compared with plain distillation. I wrote it up as plain knowledge distillation and left LoopFM out. Meta had already made that 2x claim in its Nov 2025 post, so it isn’t one of the five changes.

The main argument still holds. These details just needed fixing.

What the hype posts get wrong

I’ve read a lot of threads on these posts this week. The same claims keep coming up.

People say it’s live in every account. Meta does say the gains reach every account. But Meta measures the lifts on targeted segments, and the new way it describes ads doesn’t cover every item yet. Your account isn’t promised the average.
People say brand new ads start with data from day one. What Meta describes is softer. New ads borrow what Meta learned from similar ads before their first impression. Meta reports a 0.35% conversion lift from that work in its Oct 19 GenRecs post. A head start helps. Your new ad still has to win on its own.
People say creative is 100% of your targeting. That’s overstated. Andromeda still runs advertiser targeting rules, on the same GPU that scores the ads. Creative matters a ton. Your targeting inputs still count.
People say Meta’s sequence learning means you should sequence your ads. Those are two different things. Meta’s sequence learning is the model reading the order of one person’s actions. Ad sequencing is a buying setup you control. Meta’s posts don’t say one drives the other.
People quote Meta’s numbers as if they came with a baseline. Meta didn’t give one, so anyone promising you 18.9% is guessing.

And a lot of these threads end in a course or a tool. Read them with that in mind.

What to do in your account

I’m still not giving you a 7-step playbook to beat Andromeda.

Give the system real creative variety. Use different angles, formats and creators. Three hooks on one body doesn’t count.
Use creators to reach new people. Watch Joe Marston’s First-Time Impression Ratio (FTI), the share of impressions going to new people versus repeat viewers. His target is 10 to 20% of impressions being truly incremental each month.
Give new ads a fair read. Judge them on your own numbers once they’ve had real spend.
Measure against your own baseline. Net profit first, then blended MER, then new customer numbers. Check the platform numbers against those.
“Go broad” still isn’t universal. Test it for YOUR business.

Meta rebuilt the machine again, and you’re still the marketer.