On this pageJump to a section
Four days ago Meta's engineers published something most advertisers will never read, and it describes the part of the system I think about most.
The post introduces Andromeda. It does not talk about budgets or targeting tricks. It talks about what happens before any of that: the step where tens of millions of ads get narrowed to a few thousand that are even allowed to compete for a person's attention.1
I have bought Meta ads for ten years, and I am not going to pretend one engineering post tells me everything. This is a first look. What I can tell you is what Meta said, what it left out, and why I think the ad has to survive a step you never see.
What did Meta actually announce?
A new way to do the first and widest step of ad selection.
The system in one paragraph
Meta describes Andromeda as its proprietary machine learning system design for retrieval in ad recommendation. It runs on NVIDIA's Grace Hopper Superchip hardware, with Meta's own MTIA chips also named in the post.1 The post says it has been deployed across Instagram and Facebook.1
Why Meta built it
The post gives two reasons. The volume of ad candidates is growing, because Advantage+ keeps automating audience creation, budget allocation, placements and creative generation. And new generative AI tools for creating and optimizing ad creative mean the number of creatives in Meta's systems is expected to grow significantly.1 The challenge is to choose among all of them quickly, because delays hurt the viewer's experience.
What is retrieval, and why should an advertiser care?
It is the step that decides whether your ad gets a chance at all.
The pipeline in plain English
Meta says retrieval is the first step in its multi-stage ads recommendation system. In the following stage, larger and more sophisticated ranking models predict people and advertiser value to determine the final set of ads shown.1
Retrieval is the first step in our multi-stage ads recommendation system.
Think of a hiring process. Retrieval is the stage where a few thousand résumés are pulled from tens of millions. Ranking is the interview. You can be a great candidate and never get interviewed if your résumé looks the same as a thousand others, or says nothing the system can use.
Why this changes the question
For years, advertisers asked how to win the auction. Andromeda makes me ask an earlier question first: does the system have a reason to put this ad in front of this person at all? That is a question about what your ad says, who it is for and what the platform has learned about people who respond to it.
What does Meta claim it delivered?
Some specific numbers, all from Meta.
The reported results
The post says Andromeda's deployment achieved a +6% recall improvement to the retrieval system, delivering a +8% ads quality improvement on selected segments.1
It also reports that advertisers who had not previously used Advantage+ creative saw a 22% increase in ROAS from Meta's ads when they turned on its AI-driven targeting features, that businesses using image generation are seeing an estimated +7% increase in conversions, and that more than a million advertisers used its generative AI tools to create more than 15 million ads in a month.1
How to read them
Treat these as claims by the platform about its own product. The post does not name the advertisers, the time period or the test design behind the 22% figure, and the 8% quality gain is on selected segments, not on every account. I read them as a sign of direction and effort. I do not read them as a promise about your account.
What does this mean for a small account?
Here is my read. It is my interpretation of one engineering post, not guidance from Meta.
Creative becomes the entry ticket
If the pool of ads is growing quickly and retrieval has to pick a few thousand, an ad that is a near copy of another ad in your account is not adding much. Ten versions of one idea give the system one idea to work with. Ten different ideas, such as a problem-led angle, a proof-led angle, a story, a how-to and a comparison, give it ten reasons to show you to different people. My weekly creative testing cadence is built around exactly that.
Signal quality matters more, not less
The ranking stage predicts people and advertiser value.1 How the platform understands "value" for your business depends on what you tell it counts. If your account only reports a page view, it learns to find page viewers (the fix is in my Conversions API baseline). If it hears about booked calls and paid clients, it has something real to learn from. Meta's own Q3 results show the company calling AI the driver of its quarter,2 a quarter I broke down in Q3 2024 and the election squeeze. The advertiser's job is to give that AI better information.
Consolidate, do not scatter
Meta's Help Center advises avoiding high ad volumes in the learning phase, because when you create many ads and ad sets the delivery system learns less about each one, and says combining similar ad sets combines their learning.3 That fits what Andromeda implies: put your varied concepts inside a simpler structure, rather than spreading small budgets across many separate ad sets.
Nick Shackelford, talking to Foundr about a beverage brand he helped launch, described his account setup in a video posted on May 31, 2024: "it's simple as possible as can be and it's just an emphasis on diversity of creative."4 That was months before Meta's Andromeda announcement, so I would not read it as a reaction to it. It does line up with the direction Meta described: a small structure and a wide range of distinct ads.
How does this connect to Advantage+?
Directly. The post's title ties Andromeda to Advantage+ automation.
What Meta says Advantage+ does
The post says Advantage+ automates budget allocation, audience targeting and bid adjustments, streamlining campaign management and boosting performance through more ads in the system for different audiences.1 Notice the last phrase. Meta is describing a world where one advertiser's creative reaches many different audiences, and the system, not you, matches them.
What that leaves for you
If the matching is automated, the human work shifts to two inputs: what you put in the creative and what you tell the platform counts as success. Everything else is increasingly handled upstream. That is why I keep saying concepts, not tweaks. A concept is a reason for a particular person to stop scrolling. A tweak is a different outfit on the same reason.
How can you tell a tweak from a concept?
With a simple test I call the swap test. Take two of your ads. Swap the first line of one with the first line of the other. If the ads still say the same thing to the same person, they are tweaks of one idea. If the swap makes no sense, because one is about a problem and the other is about a result, they are different concepts.
Do this for your top five ads. You may find only two or three real concepts hiding in what looked like a varied account. The gap between "ads I have" and "ideas I have" is your first creative to-do list.
What should you do this month?
Keep it practical and small.
- Audit what you have. List your active ads. Group them by the idea behind them. If you have twelve ads and two ideas, you have two ads.
- Write five distinct concepts. Different problem statements, different proof, different formats such as a person talking, a short demo and a text-led image.
- Fix the signal first. Check what events your ad account is told to optimize for. If it is anything above a lead or a booking, plan to improve it.
- Keep the structure simple. Fewer ad sets with more distinct creative beats many ad sets with near copies.
- Leave it alone long enough to learn. Meta says edits made during learning start it over.3 Make decisions on a schedule, not on a bad morning.
A word on generative AI creative
Meta is pushing generative tools, and they can help you produce more concepts, faster. But the tool does not choose the idea. If you let it produce ten near-identical images from one prompt, you have made ten tweaks very quickly. The thinking about problem, promise and proof still has to come from someone who understands your buyer.
A note on timing
It is the first week of December. Holiday auctions are busy, and I would not tear up a working account this month because of one blog post. Use the next few weeks to plan: write the concepts, shoot or draft the creative and decide how you will test it. Then launch the new concepts when the holiday noise settles and you can read results cleanly. A thoughtful January start beats a panicked December rebuild.
What do we not know yet?
A lot, and it is better to say so.
The honest unknowns
- How many ads per account is too many or too few. The post does not say.
- How Meta judges whether two creatives are meaningfully different.
- How much of the benefit flows to small advertisers versus large ones.
- Whether account setup advice from Meta will change in the coming months.
Anyone telling you they know these answers after a single engineering post is selling something. Meta also describes its own plans in the post, including a further large increase in model complexity to come, which tells me this system will keep changing. Build habits that survive changes, like honest signals and real creative thinking, instead of settings that only work this month. I will update this piece as Meta publishes more and as I see what actually happens in real accounts.
You cannot trick a system that is deciding what to show before it ranks anything. You can only give it something worth showing.
Andromeda is the clearest sign yet that the work has moved upstream. Less cleverness in targeting. More thought in the idea, the offer and the signal. That is a good trend for operators who understand their buyers and a hard one for anyone leaning on a trick. It also rewards patience: a small account that adds one genuinely new concept a week, keeps its signal honest and resists the urge to rebuild every Monday will look very different in March than one that reacts to every headline.
If you want a second opinion on your creative mix or your conversion signals, you can reach me at ekaistone.com.
Keep building.Ekai
Questions people ask
What is Meta Andromeda?
Andromeda is what Meta calls its proprietary machine learning system for retrieval in ad recommendation. Retrieval is the first step in Meta's multi-stage ad system and selects a few thousand relevant ads from tens of millions of candidates.
When did Meta announce Andromeda?
Meta's engineering team described Andromeda in a post dated December 2, 2024, titled Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine.
What results does Meta claim for Andromeda?
Meta says its deployment across Instagram and Facebook achieved a 6% recall improvement in the retrieval system and an 8% ads quality improvement on selected segments. These are Meta's own figures.
Does Andromeda mean I should run more ads?
Not more of the same ad. The post says the number of ad creatives in Meta's systems is expected to grow significantly. My read is that distinct concepts matter more than tweaks of one ad, but Meta has not published guidance on how many ads to run.
Does Andromeda change how I set up my account?
Meta has not published setup instructions for advertisers in the post. The sensible response is to improve creative variety, send the platform accurate conversion signals and avoid unnecessary edits that reset learning.
Sources
- 1
- 2
- 3
- 4
