Ad ranking has always been a two-stage problem. There are far too many eligible ads to score every one of them for every impression, so the system first retrieves a manageable candidate set from an enormous pool, then ranks that set carefully. For most of the history of paid social, almost all of the sophistication lived in the second stage. Retrieval was comparatively crude, and advertisers compensated by doing the narrowing themselves through detailed audience targeting.
Andromeda is Meta's rebuild of the first stage: a machine-learning retrieval engine, developed with NVIDIA and running on accelerated hardware, that applies far more model capacity to the job of selecting candidates from a pool numbering in the tens of millions. Meta has described it as delivering meaningful gains in ad quality and personalization. The strategic consequence for advertisers is more interesting than the architecture, and it is this: the system is now much better at finding the right ad for a person, provided the right ad exists in the pool.
Give retrieval something to choose from
siriusly.ai turns one brief into a matrix of genuinely different reels - different personas, hooks, creators and languages - rather than thirty rewordings of the same ad.
Try it freeThe Shift in Plain Terms
| The old model | The current model | |
|---|---|---|
| Who narrows the audience | You, through targeting | The system, through retrieval |
| Your main lever | Interest and demographic stacks | Creative supply and signal quality |
| Ad sets per campaign | Many, tightly defined | Few, broad |
| What limits performance | Finding the audience | Having an ad worth showing them |
| What a "test" means | Audience A vs audience B | Creative A vs creative B |
None of this is unique to Meta - the same consolidation happened across every major ad platform as automated bidding and broad targeting matured. What Andromeda does is sharpen the point. If retrieval can genuinely find the person your ad is for, then hand-drawn audience segments are mostly a constraint you are placing on a system that would do better without it. And if the system is choosing between your ads on behalf of each individual person, then the number and variety of ads you supply is the thing that determines how good a choice it can make.
Volume Is Not the Same as Variety
The advice "make more creative" gets misread constantly. Thirty ads that are the same video with different opening captions is not thirty candidates - to a retrieval system judging on content, it is close to one candidate submitted thirty times. It does not widen the range of people your account can serve well. It just splits your budget and slows learning.
The distinction worth internalizing is between surface variants and argument variants.
Surface variants: same script, different caption font. Same footage, different music. Same claim, reworded. Same creator, different thumbnail. These are cheap to make, which is exactly why accounts fill up with them.
Argument variants: a different reason to buy, said by a different kind of person, to someone at a different stage of knowing about you. These change who the ad works on.
Surface variants are worth making, but they are optimization, and optimization only pays off after you have found an argument that works. Argument variants are discovery, and discovery is what an under-supplied account is actually short of.
The Four Axes Worth Varying
A creative matrix that produces real variety varies things that change the ad's meaning, not its styling. Four axes cover most of the useful space:
1. Persona
Not a demographic bucket - a specific person with a specific situation. "A parent of a toddler who has given up on getting eight hours" is a persona. "Women 25 to 44" is a targeting parameter, and the system does not need you to supply it any more. The persona changes the vocabulary, the objection being handled and the proof that lands.
2. Awareness stage
The same product needs completely different first sentences depending on how much the viewer already knows:
| Stage | What the hook does | Example shape |
|---|---|---|
| Unaware | Names a symptom, not a cause | "I thought this was just normal." |
| Problem aware | Names the problem exactly | "If you have tried X and it did not stick..." |
| Solution aware | Contrasts categories | "I stopped doing X and did this instead." |
| Product aware | Handles the blocking objection | "I assumed it was overpriced. Here is the math." |
| Most aware | Leads with the offer | "They are doing 30% off and I finally caved." |
Five stages times three personas is fifteen genuinely distinct ads before you have varied anything visual.
3. Creator
Age range, gender, ethnicity, setting and register. This axis matters more than it looks, because a viewer's read on "is this person like me" happens before they process a single word. It is also the axis most accounts never vary, because with human creators each new one is a new contract, a new brief and another two weeks. Generated creators remove that constraint entirely - the reasoning is in the UGC production guide.
4. Hook mechanism
Independent of what the hook says, there is how it opens: a direct-to-camera claim, a visual reveal, a pattern interrupt, a question, a mid-conversation cold open. The same script with two different hook mechanisms genuinely retains different people, and this is the axis with the highest ratio of impact to production cost.
Building the Matrix
Worked example - a first flight
3 personas x 3 awareness stages x 2 hook mechanisms = 18 argument-level creatives. Then, for the winners only, add creator and language variants on top.
The order matters: discovering which argument works is cheap at low resolution, and only the survivors deserve a full-quality render and a localized set.
Two rules keep a matrix from degenerating:
- Never vary two axes to produce the same ad twice. If persona A at problem-aware and persona B at problem-aware produce near-identical scripts, one of your personas is not actually distinct.
- Kill on the hook, not on the ad. Most of a creative's fate is decided in the first three seconds. Judging a full 30-second creative on a metric that is dominated by its first second confuses two separate decisions.
What Still Matters Besides Creative
Retrieval being smarter does not make the rest of the account irrelevant. Two things remain load-bearing:
Conversion signal quality. Every automated system is downstream of the events you send it. Server-side conversion tracking, correct event deduplication and accurate value parameters do more for delivery than any creative decision, because they define what the system is optimizing toward in the first place. An account with broken signal will not be rescued by good ads.
Structural simplicity. Fragmenting spend across many small ad sets divides the data that automated bidding needs to learn from. Consolidated structure with broad targeting gives the system more evidence per decision. This is not a new recommendation, but it compounds with everything above: broad structure plus varied creative is the combination that gives retrieval both the freedom and the raw material to work.
How to Read the Results
A varied matrix produces a messier report than a set of near-duplicates, and it is worth knowing what to look for:
- Expect uneven delivery. The system will concentrate spend on a subset quickly. That is the mechanism working, not a bug, but it does mean an ad that got 2% of spend was not really tested.
- Compare within an axis, not across the matrix. "Persona A beat persona B, holding stage and hook constant" is a usable finding. "Ad 14 beat ad 3" usually is not.
- Watch for a winner that only wins on cheap attention. A hook that spikes three-second views and collapses at add-to-cart is finding the wrong people efficiently.
- Refresh on fatigue signals, not on a calendar. Rising frequency with falling click-through is the trigger. Replacing creative on a fixed schedule throws away ads that were still working.
Common Mistakes
Reading "creative volume matters" as "publish more assets". Thirty variations of one argument gives retrieval one thing to choose, thirty times.
Fix: build the matrix along persona, awareness stage, creator and hook mechanism. Vary the argument before you vary the styling.
Keeping tight interest stacks alongside broad automated campaigns. You are re-imposing the constraint the system was rebuilt to remove, on strictly less information than it has.
Fix: go broad and let creative do the segmenting. If a segment genuinely needs different treatment, express that as a different creative, not a different audience.
Rendering every variant at maximum quality before knowing which argument works. Most of that spend is discovering that a first line was weak.
Fix: test hooks cheap, re-render winners at full quality. See the model comparison for what each tier actually costs per second.
Fixing creative in an account with broken conversion signal. Better ads optimized toward the wrong event still lose.
Fix: verify event quality, deduplication and value parameters first. Creative is the biggest lever only once the signal underneath it is sound.
FAQs
Does Andromeda change anything I do in Ads Manager directly?
No. It is infrastructure inside the ads system, not a setting you configure. What changes is which of your existing levers is worth pulling: creative supply and signal quality, rather than manual audience narrowing.
How many creatives per campaign is enough?
There is no universal number, and it scales with budget - an ad set needs enough spend per creative to learn anything. As a working heuristic, aim for enough genuinely distinct creatives that each can plausibly receive a meaningful share of the budget, and add more as budget grows rather than splitting a fixed budget thinner.
Is this an argument against ever using detailed targeting?
It is an argument against using it by default. There are still legitimate cases: regulatory constraints, genuine geographic or language boundaries, exclusion lists, and retargeting pools where the audience definition is the product of your own data rather than a guess about interests. The change is that narrow interest targeting is no longer the default first move.
Does the creative have to be video?
No, but short-form video is where attention and inventory are, and it carries far more information per impression than a static image - which is exactly what a content-aware retrieval system has to work with. It is also the format where variant production has become cheap enough to actually build a matrix.
Where does siriusly.ai fit?
It produces the matrix. Every script is built around a persona, a desire and an awareness stage rather than being a reworded copy, creators are cast per variant, and each language is generated natively rather than dubbed - so what reaches the platform is a set of genuinely different candidates instead of one ad in thirty outfits.