Meta’s New AI Ad Stack: Andromeda, GEM, and GenRecs Explained
Introduction
If your Meta campaigns have felt a bit all over the place lately, you’re not imagining it. In October 2026 Meta confirmed upgrades to the three AI systems that now more or less decide which ads turn up on Facebook and Instagram: Andromeda, GEM and GenRecs. Together they make up what we’ve started calling Meta’s AI ad stack. And they shift where advertisers should be putting their time.
Quick answer: Meta’s AI ad stack has three layers. Andromeda pulls a shortlist of relevant ads out of tens of millions of options. GEM is the foundation model that trains the ranking models (the ones that actually pick the winning ad). GenRecs describes every ad using shared semantic tokens, so the system gets its head around new ads faster. The upshot for advertisers? Creative variety and clean conversion data now matter more than fiddling with audience targeting.
What Is Meta Andromeda?
Andromeda is Meta’s ad retrieval engine. Put simply, it decides which ads even get a shot at someone’s attention. Every time a person opens Facebook or Instagram there are tens of millions of ads Meta could show them, and Andromeda whittles that down to a few thousand candidates. In milliseconds. All of this happens before any ranking or auction kicks in.
So think of retrieval as the first gate. If your ad doesn’t make the shortlist, it never gets to the auction. Doesn’t matter how good your bid is.
How Andromeda has changed since 2024
Meta first announced Andromeda back in December 2024. It swapped out a patchwork of separate models and rule-based filters for one deep-learning system. Meta’s numbers at the time: model capacity up 10,000x, recall up 6%, and an 8% lift in ad quality on selected segments.
The October 2026 version takes it a lot further. It was built with NVIDIA and now scores every eligible ad for every impression, at over one billion ads per second on a single GPU server. That’s roughly 100 times the throughput of the original. It also sends each ad to specialised “expert” models, which makes the shortlist more personal.
Why this matters for your ads
A bigger, smarter retrieval layer can look at far more creative per person. That’s why the old “keep it to a handful of ads per ad set” rule has quietly faded away. What Andromeda really rewards is creative that’s actually different, not ten near-identical edits of the same idea. If your team is already stretched, this is usually where outside help with video and creative production pays for itself.
What Is Meta GEM (Generative Ads Recommendation Model)?
GEM is Meta’s ads foundation model. You’ll often see it called the “central brain” of the whole ad system. Meta says it’s the largest recommendation foundation model in the industry, trained across thousands of GPUs, and the approach borrows a fair bit from how large language models are built.
Here’s the bit people tend to miss, though. GEM doesn’t pick ads itself. It teaches the smaller ranking models that run on each surface (Facebook Feed, Instagram Reels and so on) using techniques like knowledge distillation. And because it learns across all of Meta’s apps, something it picks up from Instagram video engagement can improve predictions over on Facebook Feed.
GEM’s results so far
- Q2 2025: roughly 5% more ad conversions on Instagram, and 3% on Facebook Feed.
- Q2 2026: Meta put 8.3% more clicks and a 15.7% conversion lift on Facebook down to its ranking improvements, GEM included.
- October 2026: the newest GEM reads twice as much of each user’s history. Meta says that’s delivered around 19% better conversions since last year.
Worth remembering these are Meta’s own numbers, so take them as directional. What you actually see will depend on your account, your offer and how good your data is.
What Is Meta GenRecs?
GenRecs (short for generative recommendations) is the newest piece, and Meta only went into detail on it in October 2026. What it changes is pretty fundamental: how the system describes an ad in the first place.
Up until now, Meta’s models tracked performance against individual IDs for each campaign, creative and product. Problem is, those IDs change all the time. So a brand-new ad started out with next to nothing to learn from. GenRecs swaps IDs for concepts instead. Large language models read an ad’s text and images, then describe it with shared building blocks called semantic tokens. Things like brand, colour, fit, style.
Solving the cold-start problem
Because every ad gets described in the same vocabulary, the system can spot what a new ad has in common with ads it already knows. Say you launch a new pair of trail-running shoes. It can borrow what it’s learnt from similar products rather than starting cold. Meta says this shortens the learning phase. It doesn’t get rid of it, mind you.
Generative retrieval
GenRecs also runs a new generative retrieval model. Easiest way to picture it is a small LLM built just for ads. It looks at the posts and ads someone engages with and predicts which ads they’ll probably find relevant. Those picks sit alongside Andromeda’s, so the ranking stage gets a second source of strong candidates.
How Andromeda, GEM and GenRecs Work Together
Each layer feeds into the next one. Roughly, this is what happens in the split second after someone opens Instagram:
- GenRecs has already described every eligible ad in shared semantic tokens.
- Andromeda and the GenRecs generative retrieval model both shortlist ads for that person, side by side.
- Ranking models, trained by GEM, predict clicks and conversions for each ad on the shortlist.
- The auction weighs those predictions against your bid and budget, and picks the ad that gets shown.
Notice what isn’t on that list? Detailed audience targeting. Matt Steiner, Meta’s VP of ads infrastructure, made the point himself: targeting matters less and less as the ranking models get better at working out who an ad is right for.
What Meta’s AI Ad Stack Means for Australian Advertisers in 2026
Good news first: you don’t have to change any settings to get the benefit of these upgrades. What you do need to change is what you feed the system. If it were our account, this is where we’d start.
1. Diversify your creative concepts
Retrieval likes variety now. Aim for somewhere around 10 to 20 genuinely different ads per campaign, more or fewer depending on budget. Different angles, different formats, different people in mind. Mix it up with static images, short-form video, carousels and UGC-style stuff. (Ten colour swaps of one image? That’s one idea, not ten.)
2. Make every ad easy to understand
GenRecs is reading your text and visuals to work out what you’re actually selling. So show the product clearly. Call it what it is. Keep your catalogue titles and descriptions accurate (if you sell online, that mostly means tidying up the product data in your ecommerce website). The more specific the creative, the faster it gets matched to the right people.
3. Sort out your conversion data
GEM can only learn from the signals you send it. Run the Meta Pixel and Conversions API together, deduplicate your purchase and lead events, and have a look at your Event Match Quality in Events Manager. Good digital marketing foundations start with clean tracking, boring as that sounds.
4. Go broader with targeting
Loosen up those tight interest stacks and let Advantage+ audiences and placements do the heavy lifting. Put that energy into your offers, messaging and landing pages instead. At the end of the day a fast, conversion-focused website still decides whether a click turns into a customer. And if those same pages are set up properly for search (that’s where a good SEO service earns its keep), they’ll keep pulling in organic traffic long after the campaign wraps up.
5. Measure what actually matters
Meta’s reported lifts are, well, measured by Meta. Back them up with your own read on incrementality. Compare blended cost per acquisition, run conversion lift tests, and keep an eye on revenue in your own analytics, not just Ads Manager. Running Google Ads as well? Then it really helps to have an SEM agency looking at search and social side by side, so you can see what each channel is genuinely doing.
Frequently Asked Questions
What is the difference between Andromeda and GEM?
Andromeda does retrieval, so it picks a shortlist of ads that might suit someone. GEM sits at the ranking stage. It trains the models that predict which of those shortlisted ads is most likely to get a click or a conversion.
Do I need to change my Ads Manager settings for Andromeda, GEM or GenRecs?
Nope. Meta rolls these upgrades out automatically. The part you control is the quality of what goes in: creative variety, clear product info and accurate conversion tracking.
Is audience targeting dead on Meta?
Not completely. But it’s carrying a lot less weight than it used to. These days broad targeting with strong creative tends to beat narrow interest stacks, mostly because the AI models are just better at finding the right people.
How many ads should I run per campaign in 2026?
There isn’t a magic number. That said, 10 to 20 distinct creative concepts is a sensible place to start for most budgets. Go for different ideas over small tweaks.
Does GenRecs help new ads exit the learning phase faster?
Yes, that’s really the whole point of it. Because a new ad gets described in shared semantic tokens, Meta can lean on what it already knows about similar ads, and the learning phase gets shorter.
Do these changes affect small Australian businesses?
They do. The same AI ad stack sits behind every Facebook and Instagram campaign, whatever the budget, wherever you are.
Conclusion: Feed the Machine, Then Steer It
Meta’s AI ad stack has moved the levers that drive performance. Andromeda decides which ads get to compete. GenRecs helps the system understand them. GEM sharpens the predictions that pick the winner. The advertisers doing well in 2026 aren’t trying to out-guess the algorithm. They’re giving it better creative, better data and a clearer strategy, and then letting it get on with the job.
Keen to get your Meta campaigns working with Andromeda, GEM and GenRecs rather than against them? Have a chat with the Planet Media team about a social media advertising strategy built for how Meta actually works now.
