What Meta's ad system actually changed, and why most advertisers got it backwards
Twelve thousand years ago, humanity stopped chasing animals and started planting seeds.
Anthropologists call it the Neolithic Revolution. In practical terms we went from hunters to farmers, and it changed everything about what success required.
The hunter depended on his ability to track, pursue and capture. His success was individual, immediate, and demanded absolute control over every movement. The farmer learned to prepare ground, plant, and trust forces he did not control: weather, seasons, the soil itself. His success was systemic, delayed, and demanded patience and variety.
The hunter asked: where is my prey?
The farmer asked: what should I plant?
Digital advertising has just completed the same transition. Most of the people inside it are still hunting.
The engine was replaced and almost nobody noticed
In December 2024, Meta published a technical engineering post about a new ad retrieval system. They named it after the galaxy currently approaching our own, the one that will eventually merge with the Milky Way rather than destroy it. Whether the choice was deliberate or accidental, it turned out to be accurate.
There was no announcement, no email, no notice inside Ads Manager. By October 2025 the global rollout was complete. By early 2026, every advertiser on Facebook and Instagram was operating inside it, regardless of budget, vertical or sophistication.
Nothing changed on the surface. Same Ads Manager, same buttons, same reports. Underneath, the engine had been swapped.
Here is what it actually does, because this gets misreported constantly.
Meta's ad delivery runs in stages. Retrieval narrows tens of millions of eligible ads down to a few thousand candidates. Ranking then decides which of those candidates wins the impression. Andromeda is the retrieval stage. It does not choose the winning ad. It chooses the shortlist. Above it sits Lattice, the ranking and learning system that replaced Meta's previous collection of hundreds of siloed models, and GEM, the generative recommendation model now operating in that ranking layer.
That distinction matters more than it sounds. Andromeda does not decide who wins. It decides who is allowed to compete.
Built on NVIDIA Grace Hopper Superchips alongside Meta's own silicon, it raised model capacity by a factor of ten thousand. The system can now evaluate up to a hundred thousand ads per person per auction. Meta's internal testing reported an eight percent improvement in ad quality and a six percent improvement in recall.
Which sounds like unambiguous good news. For most advertisers, it was not.
Two sets of numbers, both true
Meta's own figures show Advantage+ Shopping campaigns delivering an average return on ad spend of 4.52 against 3.70 for manual campaigns, a twenty two percent improvement.
Meanwhile, an independent study of 3,014 e-commerce advertisers, covering 834 million dollars in spend, one million ads and 115.7 billion impressions across 73 countries during the full rollout year, found return on ad spend declining seven percent on average. Landing page conversion rates fell industry wide from 3.5 percent to 2.9 percent. High ticket products saw returns collapse by thirty one percent. Lower priced products fared worse still, dropping thirty five percent.
Both sets of numbers are accurate. The gap between them is the entire story.
And inside that gap sits the detail that should stop anyone who reads it carefully. The advertisers hit hardest were the ones who had been performing best before.
That is not a malfunction. That is the shape of every technological transition ever recorded. The people who suffer most are rarely the incompetent. They are the ones who mastered the previous discipline most completely, because mastery of a system is worthless the moment the system is replaced, and it is expensive to unlearn.
Something separated the advertisers who gained from the advertisers who bled. It was not budget. It was not vertical. It was a single misunderstanding about what the new system rewards.
The mistake almost everyone made
The old system rewarded finding one winner and scaling it. You built an ad that worked, produced twenty slight variations of it, and put your money behind the best performer. That was competent practice for fifteen years.
Under Andromeda it is close to self sabotage.
Each genuinely distinct creative receives its own entity identifier, its own ticket into the auction. Twenty near identical versions of the same concept do not receive twenty tickets. They collapse toward one. An advertiser running twenty variations of a single idea is competing with roughly one entry. An advertiser running twenty different ideas is competing with twenty.
The system rewards diversity and penalizes repetition, and it does so at the level of concept, not file.
This is where the farming metaphor stops being decorative and turns literal. A farmer who plants one crop across every field is exposed to a single blight. A farmer who plants variety survives conditions he could not have predicted. The advertisers who suffered most were rarely producing too little. They were producing plenty, and all of it was the same.
What the evidence now supports
Meta's testing found seventeen percent more conversions at sixteen percent lower cost when running one ad set containing twenty five creatives, compared with five ad sets containing five each. Consolidation beat fragmentation, because larger pools let the system learn faster.
A separate study across more than a million creatives found that brands launching twenty or more new ads per month achieved sixty five percent higher return on ad spend than brands testing fewer than ten.
Read those two findings together and the instruction is unambiguous. Fewer campaigns. More concepts. Broader targeting. And a production cadence most marketing teams are not currently built to sustain.
Where the money moves
The spend does not disappear. It relocates.
What used to be invested in finding the right audience now has to be invested in producing enough genuinely different things to say. Audience research, which once consumed a quarter of the effort in a serious campaign, becomes a minor line. Manual segmentation approaches zero, because the system infers the audience from the creative itself. Creative production, which was often the smallest budget line, becomes the largest.
This is why so many agencies struggled. An agency whose differentiator was targeting expertise sold a skill the platform now performs better and for free. An agency whose differentiator is creative range is selling exactly the scarce input.
Who gains and who loses
Gaining: brands with real creative capacity, products that are visually differentiable, catalogs deep enough to generate genuine variety, and businesses with an actual story, because a real story produces many true angles.
Losing: advertisers running a single creative concept, because there is nothing for the system to choose between. Generic products without differentiation, because the system cannot find angles that do not exist. Very narrow niches, where the retrieval pool is too small to learn from quickly. And anyone whose competitive advantage was manual targeting.
The paradox worth naming
For a decade, digital advertising trained us to see people as data to be exploited. We called them targets. We grouped them into cohorts. We reduced them to demographic categories. It was a profoundly dehumanizing view wearing the costume of scientific precision.
Andromeda, paradoxically, concedes that we are more complicated than our categories. The system does not ask whether you are a thirty five year old woman interested in entrepreneurship. It observes behavior, micro signals, resonances that we do not consciously recognize in ourselves.
In a strange sense, the machine reads human complexity better than the forms we used to fill out.
And that creates an opening.
If your message is authentic, if your offer comes from somewhere real, if what you sell genuinely changes something, the system will find the people it belongs to and amplify it. Truth no longer has to disguise itself as a demographic segment.
But if your message is thin, if your offer is generic, if you are extracting value rather than creating it, no targeting will rescue you. The system will expose the absence of substance with ruthless efficiency, and it will do it faster than the old one did.
What to do now
Produce concepts, not variants. Ten to twenty genuinely different ideas per campaign, not twenty crops of the same image. If two ads could be described in the same sentence, they are one ad.
Design natively for each format. A square creative forced into a vertical placement gets cropped, performs worse, and teaches the system that your creative is weak.
Consolidate. Fewer campaigns, fewer ad sets, larger budget pools. Fragmented budgets starve the learning.
Broaden targeting. The creative is now the targeting. Narrowing the audience on top of it fights the system that is trying to help you.
Stop optimizing daily. Give the system at least a week before you judge anything, and stop killing ads after twenty four hours of weak results.
Measure differently. Stop hunting for the perfect audience and start tracking which concepts resonate, at the creative level rather than the campaign level.
Your creative is the seed. Your budget is the ground. The system is the weather.
You cannot control the weather. You can choose what to plant, and how many different things.
And in an era where the machine understands the user better than you do, the only durable advantage left is the truth of what you are actually offering, expressed in as many honest forms as you can produce.
Myriam Vanegas
Sources: Meta Engineering, "Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine," December 2024. The Confect Andromeda Study, 3,014 e-commerce advertisers across 73 countries, 834 million dollars in spend, one million ads and 115.7 billion impressions, covering the 2025 rollout period. Meta internal testing data on ad set consolidation and Advantage+ Shopping campaign performance.
