On 28 October, Meta published a short note on the three AI systems that now decide which ads people see on Facebook and Instagram, called Andromeda, GEM and GenRecs. It contains one line every advertiser will like. There is “no setting to switch on, no new campaign type to adopt, no migration to plan”, because the changes already run in your account.
That is also why the note deserves ten minutes of your time. When nothing needs switching on, the only levers left are the ones you already had, and Meta lists them itself. They are the result you ask it to optimise for, the quality of the data you send back, and how much of the optimisation you let its models handle. Here is how the three systems fit together, how to read the numbers Meta quotes, and what we are changing in the accounts we run.
Key takeaways
- Andromeda shortlists your ads, GEM ranks them, and GenRecs reads what each ad shows.
- There is nothing to activate. Your optimisation event, your data and your constraints decide the outcome.
- Meta’s gains are self-reported, over different periods and scopes, and they do not add up.
- In our reading, distinct creative angles now count for more than near-identical variants.
How Meta picks an ad, in three stages
Each time someone opens Facebook or Instagram, Meta has to choose a handful of ads from a pool it puts at tens of millions. It does this in stages, and each of the three systems owns one of them.
Andromeda draws up the shortlist. It is the retrieval stage, launched in 2024 on Meta’s own MTIA chips and NVIDIA Grace Hopper hardware. Its job is to cut the pool down to a few candidates per person, after applying the targeting rules you have set. The new version, built with NVIDIA, sends each ad to specialised “expert” models instead of one general model, and scores every eligible ad. Meta’s engineering team puts its throughput at over one billion ads per second, 100 times the 2024 system.
GEM ranks the shortlist. GEM, the Generative Ads Recommendation Model, was first presented in 2025. Meta now calls it the foundation model behind ad recommendations on Facebook, Instagram, Reels and conversions that happen off Meta. It is too large to run on every request, so it passes what it learns to smaller ranking models that answer in real time. The new generation reads twice as much engagement history, with a model 2.5 times more complex.
GenRecs reads the ad itself. This is the newest piece. Language models read the text and images of each ad and describe them in what Meta calls semantic tokens, covering things like brand, colour, cut and style. GenRecs then works the other way round. For a given person, it generates the description of the ideal ad, then looks for the real ads that match it. Meta extended this generative retrieval to the Facebook Feed in the third quarter.
| System | Its role | Meta’s headline figure |
|---|---|---|
| Andromeda | Shortlists candidate ads for each person | +2.4% conversions on Facebook since 2025 |
| GEM | Ranks ads and trains the real-time ranking models | Up to +18.9% conversions and +12% click-through rate since 2025 |
| GenRecs | Describes each ad and matches it to the person | Nearly +2% conversions in targeted Feed segments |
How to read Meta’s numbers
Meta adds that improvements to its ranking system brought 8.3% more ad clicks and 15.7% more conversions on Facebook in the second quarter of 2026. Those are large figures, and they come with three caveats.
All of them are reported by Meta, from internal data, without a published method. The periods differ, from “since 2025” to the second or third quarter, and February 2026 for the GenRecs tokenisation result (+0.35% across surfaces). The scopes differ as well. Some figures cover Facebook only, one covers Instagram Reels (+0.24% for the latest Andromeda upgrade), others cover “targeted users”. They cannot be added together, and none of them is a forecast for your account.
They do show a direction. A growing share of Meta’s progress comes from its ranking machinery rather than from new campaign settings. If your results moved this autumn while nobody touched the campaigns, part of the reason may sit here. The only way to know is to compare with your own sales or CRM data over the same weeks.
First lever, the result you ask Meta to optimise for
Meta’s note ends with a sentence we would happily print on the office wall. “It still needs you to tell it what winning looks like for your business.” The optimisation event is that definition. If a campaign optimises for landing page views, GEM will find people who view landing pages, and it will do it very efficiently.
GEM does help with an old problem. Meta’s engineers write that what the model learns from click data, which is plentiful, carries over to conversion tasks where data is sparse. In practice, a campaign with few purchases or leads should struggle less than before when it optimises for that event. We take it as a reason to move some accounts from a soft event, such as a form opened or a page viewed, to the event closest to revenue, after a test. It does not make a weak signal reliable, though, and an account with three leads a month still needs a proxy event.
Second lever, the data you send back
Meta’s second condition is data quality, “including how you measure performance”. The ranking models learn from the conversions you report. Missing conversions, duplicated ones or a purchase sent without its value teach the model the wrong lesson, and they now do it faster.
The checks we run first on a Meta account are dull, and they pay off. The Pixel and the Conversions API should both send the key event, deduplicated with the same event ID. Value and currency should arrive with every purchase. Lead campaigns should send back which leads became customers, when the CRM allows it. If you want a second pair of eyes on this, it is the daily work of our tracking and reporting team.
Third lever, how much room you give the system
Andromeda’s engineering post confirms that retrieval applies “advertisers’ targeting rules” before any scoring. Every constraint you add removes candidates before the models even see them. Some constraints describe a fact about the business, like a delivery area, a legal age or existing customers to exclude. Keep those. Others are guesses made years ago, such as a stack of interests or an age band nobody can explain any more. Those are the ones worth testing without.
We were already moving that way with Advantage+ campaigns, as we explained when Meta retired Advantage+ Shopping and App campaigns. The October note does not mention Advantage+ by name, so treat the link between the two as our reading, not as Meta’s advice.
Paid Social
Give Meta’s models the right signal
We review your optimisation events, Conversions API and targeting constraints, then test each change against your real sales.
Fourth lever, creative the model can tell apart
GenRecs changes how the system sees your creative. Language models read each ad’s text and images and file it under descriptions such as brand, colour or style. Meta says new ads, refreshed creative and new products get “on a faster track”, because they inherit what the model already knows about similar items. Its engineers add that the effect is largest on “tail” ads with few impressions, which are often the ones smaller advertisers run.
Meta gives no rule on how many ads to run or how different they should be, so what follows is our reading. Ten versions of one image with ten headlines will probably be described in almost the same terms, which gives retrieval few genuinely different options. Five ads that show different products, settings, people or arguments give it more to match against different people. The same engineering post notes that generative AI tools have helped advertisers produce “hundreds of millions of new ads”, so volume on its own will not set you apart.
We now brief creative on distinct angles first and on variants second. When a client cannot shoot new material every month, our creative design team builds those angles from existing assets.
”If Meta does everything, what is left for us?”
It is the question we hear most after an announcement like this one, and we answered it at length when Meta announced fully automated campaigns. The October note makes the answer more concrete. The machine now decides who sees which ad. You still decide what counts as a win, which data proves it, which constraints are real and what the ads say. Those four decisions are harder to get right than bid settings ever were, and a mistake in any of them now spreads faster through the account.
What we would do this month
Nothing urgent, and nothing to rebuild. Black Friday is three weeks away, and restarting campaigns because of a ranking update would cost more learning than it brings. We would check the optimisation event of each campaign and run a short audit of the Conversions API and purchase values. We would also list the targeting constraints and mark the ones nobody can justify, then plan January’s creative around three or four distinct angles.
After that, we would read the results against sales data rather than against the figures in Meta’s note. They describe Meta’s average across billions of auctions, and your account is one of them.
Sources Meta for Business, 28 October 2026 - Meta Engineering, Andromeda, 28 October 2026 - Meta Engineering, GEM, 13 October 2026 - Meta Engineering, GenRecs Tokenization, 19 October 2026
