Meta · Updated · 12 min
Meta Andromeda: why creative volume and variety now count
Since Andromeda, the creative is the strongest lever you still hold in a Meta campaign. The audience setting only draws the frame. Anyone still running three variants of the same idea a month is leaving most of the system unused.
Alexander HertiFounder of Business Growth Solutions
In short
Meta Andromeda is the AI system Meta uses to pick, out of tens of millions of possible ads, the few thousand worth considering for a particular person. Because Meta treats ads that look too alike as variations of the same creative, the winners are advertisers who run several clearly different concepts in a few broad campaigns, not many colour variants of one idea.
What Meta Andromeda is, technically
Before an ad reaches the auction it has to make the shortlist at all. Meta calls that step retrieval, and it is exactly the step Andromeda rebuilt. According to Meta's engineering blog from December 2024, this stage narrows tens of millions of ad candidates down to a few thousand relevant ones, processing about a thousand times more ads than the stages after it.
Technically, Andromeda is a large neural network running on NVIDIA Grace Hopper Superchips and Meta's own MTIA chips. Ads are organised in a multi-layer hierarchical index that is trained jointly with the retrieval model. Every ad has precomputed features ready, which the system matches against a person's behaviour and interests.
Meta gives concrete numbers: plus 6 per cent recall in the retrieval system across Facebook and Instagram, plus 8 per cent ads quality on selected segments and a 10,000-fold increase in model capacity. Meta explicitly names the growing number of creatives as the reason for the rebuild, driven partly by its own AI tools: more than a million advertisers used them to create over 15 million ads in a single month.
These numbers describe the system's average improvement, not the result for your account. Andromeda makes the selection more precise; it does not guarantee lower costs to any single advertiser.
Retrieval and ranking: how the stages fit together
Delivery at Meta runs in several stages. First Andromeda selects the candidates. Ranking models then evaluate those few thousand ads in more detail, estimating for example the probability of a click or a conversion, and only then does the auction decide which ad actually appears.
The logic that follows is simple: an ad that is not picked in retrieval cannot win in ranking, however high the bid. Meta does not disclose exactly which features of a creative Andromeda evaluates. What Meta did state clearly in its creative diversification guide from December 2025: ads that look or feel alike are seen as variations of the same creative, and learnings and delivery optimisation are shared at that level.
That brings a sentence advertisers have repeated for years much closer to the truth: the creative is the targeting. Which motif with which message you run decides, more than any interest selection, which people the ad ends up in front of.
Why cosmetic variants do nothing
The most common mistake after the change is well meant: twenty ads get uploaded, but they are twenty versions of the same idea, with a different button colour, a different font size, a different crop.
To a system that recognises similarity, those are not twenty chances but essentially one. The ads share their learnings instead of each opening up new people. Twenty times the effort, barely more effect. Meta itself recommends ads that are truly different in look, feel, storyline and message rather than small tweaks.
Different here means: a different angle, a different person on screen, a different format, a different claim. A testimonial and a product shot are two ads. The same testimonial in two colours is one.
What that means for account structure
The tight audience settings that used to fill half a strategy meeting have lost weight. In its Help Center, Meta recommends merging ad sets with similar goals, because an audience split across many ad sets gives each one less data and prolongs the learning phase. According to Meta, an ad set usually exits the learning phase after about 50 results in the week following its last significant edit.
For Advantage+ audience, where Meta's AI builds the audience itself, Meta cites possible cost advantages, for example 9.7 per cent lower cost per result for the traffic, engagement and leads objectives, and recommends an A/B test for almost all campaign types except retargeting. Hard limits such as location, minimum age or exclusions remain available.
| Area | Common before Andromeda | Sensible today |
|---|---|---|
| Campaigns | Many, split by audience | Few, split by objective or offer |
| Ad sets | One per interest, lookalike or placement | One or a few broad ad sets, analysed through breakdowns |
| Targeting | Narrow interests, exclusions, age bands | Advantage+ audience or broad, only hard limits set |
| Placements | Selected manually | Advantage+ placements with assets per placement |
| Creatives | A few variants of one idea | Several clearly different concepts per ad set |
| Main work | Building audiences and fine-tuning bids | Developing and producing concepts, securing signal quality |
The effort therefore shifts from account settings into production. Three hours that used to go into interests and exclusions are better spent on three additional motifs today.
What creative diversity means in practice
Meta defines creative diversification as distinct assets tailored to different personas or use cases, as opposed to iteration, meaning the testing of small changes. In practice it helps to plan diversity along a few dimensions. A new concept should differ clearly from what is already running in at least one of them.
| Dimension | Genuine diversity | Not genuine diversity |
|---|---|---|
| Angle (core message) | Save cost, reduce risk, gain time, overtake competitors | The same claim in other words |
| Persona | Managing director, head of sales, procurement, engineer | The same person in a different outfit |
| Format | Static image, talking-head video, screen recording, carousel | The same video in three aspect ratios |
| Hook (first seconds) | Provocative question, hard number, live demo, problem scene | Same opening, different font |
| Proof | Customer voice, case study, product demo, figures | The same quote on a different background |
| Offer and next step | Intro call, checklist, calculator, webinar | Different text on the same button |
Aspect ratios and copy variants are still useful, just as adjustments within one concept. Meta offers dedicated features for that, such as asset customisation per placement and multiple text options in one ad. They do not need an ad of their own.
The numbers going around the market
Since the change, agencies and tool vendors report clear differences between accounts with many different motifs and accounts with few. They quote fixed minimum numbers of new ads per month and an advantage in bundling many different motifs in one ad set rather than spreading them across several.
These figures come neither from Meta nor from our own accounts. Treat them as a direction, not a promise. Meta itself names no minimum, and its Help Center even warns against running too many ads at once: each one then gets less delivery and fewer ads exit the learning phase. The direction is still clear: more genuine variety beats more fine-tuning, but more sheer mass does not.
A weekly testing rhythm
Andromeda rewards accounts that deliver new concepts regularly. A fixed weekly rhythm stops tests from getting lost in day-to-day work. This is how we work in practice, as experience rather than a Meta rule:
- Monday, review: look at last week's results per concept, not per individual ad. Which angles carry, which do not?
- Tuesday, brief: define two to four new concepts, each clearly different from what is running in at least one dimension.
- Wednesday and Thursday, produce: the required formats per concept, for video with two or three different hooks.
- Friday, launch in one batch: add new ads once a week together instead of one at a time every day. Every significant edit can restart the learning phase.
- Ongoing, clean up: switch off ads that clearly trail after enough delivery instead of patching them. Winners keep running.
Three levels help when reading results. At the top is the metric the business pays for: cost per qualified lead or per meeting, not just cost per lead. Below that, link click-through rate and landing page conversion rate. For video, two diagnostic values join in: how many people watch the first three seconds and how many keep watching. A weak start is a hook problem, an early drop-off after that is a content problem.
In our experience a concept should run for at least seven days and receive a budget in the order of a few target leads before you judge it. Meta explicitly advises against editing ad sets during the learning phase.
Creative fatigue: when motifs wear out
Meta describes ad fatigue as the point where an audience has seen the same ad too often, interacts less and cost per result rises. In Ads Manager, Meta shows two levels for ad sets with a single creative: if cost per result is above that of your previous ads but below double, the status reads creative limited. From double upwards, it reads creative fatigue.
One detail matters: Meta also counts deliveries of the same image or video in other campaigns from your page. Running one motif in five campaigns at once wears it out faster. As a remedy, Meta recommends a new creative that differs substantially from the original and suggests you do not necessarily switch the old one off.
Many practitioners report that motifs tire faster since the change. Meta has published no hard numbers on that. Our rule of thumb: refresh by signal, not by calendar. Rising frequency combined with falling click-through rate and rising cost per result over two weeks is the moment for new concepts.
Andromeda in B2B: what is different
Most Andromeda guides are written for online shops. In B2B the same mechanisms apply, but under different conditions: the relevant audience is smaller, the sales cycle runs for weeks or months, and a cheap lead is worthless if nobody buys.
- The signal decides quality: if a campaign only optimises for completed forms, the system finds people who like filling in forms. If you want quality, you have to tell Meta which leads later turned into meetings and deals.
- CRM feedback through the Conversions API: Meta lets you send events from your CRM back via the Conversions API. For the conversion leads performance goal, Meta lists requirements such as at least 200 leads a month, daily uploads and a conversion rate of 1 to 40 per cent for the target stage. It currently works only with lead ads, meaning instant forms.
- Small accounts need earlier events: if you do not reach 50 meetings a week, optimise for an earlier, more frequent event and check quality yourself in the CRM. Consolidating instead of fragmenting matters even more here.
- In B2B, diversity often means roles: management, the specialist department and procurement have different objections. One concept per role is a natural form of creative diversity.
The metric you steer by should therefore not be the price per lead but the price per qualified conversation. A concept with double the lead price and triple the meeting rate is the better one.
Common mistakes after the change
- Selling twenty colour or copy variants of one idea as a test.
- Keeping the old structure of many small ad sets and then complaining about unstable results.
- Adding ads, changing budgets and editing winners every day, so that nothing ever leaves the learning phase.
- Optimising for cost per lead and never feeding quality back from the CRM.
- Running one successful motif in many campaigns at once and wearing it out faster.
- Judging after two days although there has been hardly any delivery.
What that means for production
Several genuinely different concepts a week can hardly be produced the classic way. A shoot day delivers one motif and a few cuts, and the next shoot day is six weeks out.
That is exactly where AI in creative production pays off, not because it looks better but because it delivers the volume the system now demands. What actually comes out of it we wrote down here: AI creatives, what that actually means.
Since August 2026, transparency obligations under the EU AI Act apply to AI-generated content. What that means for ads specifically is covered in a separate piece: labelling AI content.
How to tell it affects you
Three patterns keep showing up since the change. First: an ad that ran for months loses its effect within days, with nothing having changed. Second: you upload new motifs and performance still does not move.
Third: some of your ads get almost no delivery although the budget is there. That is rarely a budget problem. Often those ads resemble a motif already running so closely that the system sees no reason to serve them as well.
The test is simple: lay your last ten ads side by side and ask someone who does not know the offer how many different ideas they see. If the answer is two or three, you have your explanation.
What has not changed
Andromeda picks the ad, it does not sell for you. A weak offer stays a weak offer, and a page that hides the next step still loses the click.
The maths behind it also stays the same: what a meeting may cost is decided by your contribution margin, not by the platform. How to work that out is in our piece what may a B2B meeting cost.
What we took from it
We produce several angles per campaign instead of one idea in variants, we test hooks separately from the rest of the video, and we switch losers off instead of patching them. That was our way of working before; since Andromeda it is no longer an opinion but the condition.
If you are currently seeing your ads tire faster than before and more motifs no longer helping, the reason is often that the motifs resemble each other too closely. That is quickly spotted in a conversation. We handle concepts, production and campaign management as part of our performance creatives for Meta and Google, or you can reach us directly: contact.
Frequently asked questions
What is Meta Andromeda?
Andromeda is the AI system Meta uses to preselect ads on Facebook and Instagram. Out of tens of millions of possible ads it picks a few thousand candidates for a person, which then go into ranking and the auction. Meta introduced it in December 2024 and cites, among other things, 6 per cent more recall and 8 per cent higher ads quality on selected segments. For advertisers, it means the creative decides more than before who sees an ad.
How many creatives do I need for Andromeda?
Meta names no minimum and even warns against running too many ads at once, because each then gets less delivery. What matters is the number of clearly different concepts, not the number of files. In our practice, a rhythm of two to four new concepts per week and ad set works well, fewer on small budgets. Twenty colour variants of one idea count as one concept.
Do I still need interest targeting?
Mostly not as a tight limit. Meta recommends A/B testing Advantage+ audience for almost all campaign types except retargeting and advises merging fragmented ad sets. Hard limits such as location, minimum age or excluding existing customers still make sense. You can pass interests or custom audiences as a suggestion, and the system may then go beyond them. The actual audience targeting today sits in the creative.
Does Andromeda work for B2B?
Yes, the same mechanisms apply, but the conditions are tougher: smaller audiences, longer sales cycles and a wider gap between lead and deal. What matters most is feeding quality back to the system, for example via the Conversions API with data from your CRM. For the conversion leads performance goal Meta requires, among other things, at least 200 leads a month. Smaller accounts consolidate campaigns and check lead quality themselves.
How often should I refresh creatives?
By signal, not by calendar. In Ads Manager Meta shows the status creative limited when cost per result is above that of earlier ads, and creative fatigue from double upwards. Rising frequency with falling click-through rate is also worth watching. In our experience it makes sense to feed in new concepts continuously, so that replacements are already tested when winners start to fade. Running winners do not have to be switched off for that.
Does it matter whether creatives are AI-made?
For selection by Andromeda, based on everything Meta has published, what counts is how relevant and how distinct a creative is, not how it was made. Meta offers AI tools itself and names the resulting growth in ads as one reason for Andromeda. AI mainly helps produce the required number of different concepts. Legally, transparency obligations for certain AI content have applied in the EU since August 2026.
Sources
- Meta Engineering: Meta Andromeda, next-gen personalized ads retrieval engine (December 2024)
- Meta for Business: Demystifying creative diversification (December 2025)
- Meta Business Help Center: About Advantage+ audience
- Meta Business Help Center: About the learning phase
- Meta Business Help Center: Creative fatigue recommendations
- Meta for Developers: Conversions API for CRM integration