The Andromeda update: what changes for your Meta ads
Andromeda is Meta's ad retrieval engine: in a few hundred milliseconds, it decides which ads, out of tens of millions, are eligible for a given person. For the advertiser, the consequence is budgetary: the performance lever moves away from tuning your targeting and toward producing genuinely different messages and improving the quality of your conversion signals.
- Fine-grained targeting loses value, because the engine rebuilds the matches between a person and an ad on its own.
- Creative diversity becomes the main performance variable, ahead of the number of creatives.
- An account poorly equipped with conversion signals falls behind, no matter how good its targeting.
- Structures with ten ad sets fragment the learning rather than sharpening it.
On this page
Andromeda
Andromeda is Meta's ad retrieval system, deployed across Facebook and Instagram. It filters tens of millions of candidate ads down to a few hundred before the ranking stage. It relies on deep neural networks rather than rules, which lets it rebuild the matches between a person and an ad on its own instead of drawing on predefined segments.
How Andromeda chooses the ads a person will see
Ad delivery on Meta happens in three stages. Retrieval narrows a vast inventory down to a short list of candidates. Ranking orders that list. Delivery serves the selected ad. Andromeda operates at the first stage, the one that decides who even gets to enter the race.
That is what makes the change significant. An ad dropped at retrieval is never ranked, so never served, whatever its budget. The levers that acted on ranking now carry less weight than before. Those that act on eligibility carry more.
The old system filtered with a relatively coarse rule-based logic, built on features defined by humans. Andromeda uses a deep neural network and a hierarchical index trained jointly, which lets it consider a far larger pool of candidates without losing precision or speed.
The right question is no longer whether you are targeting the right people. It is whether your ads are eligible at the right moment, and whether the system understands what counts as a result for your business.
One question still decides every advertising budget: knowing what the campaign actually added. That is the subject of our incrementality tests.
What this change moves in your budget
The line item that decides performance has moved. Fine-tuning the targeting, which used to absorb most of the billed hours, returns less and less. What returns value is producing genuinely different messages, and that work costs shooting and writing time, not account-management hours. A business that does not shift this budget pays for work whose return keeps falling.
- How many genuinely different angles are running right now, not how many visuals?
- What share of the billed time goes to tuning the account, and what share to creative production?
- Who produces our messages, and at what pace can we supply new ones?
- Are our conversion events reliable, and since when?
- At what cost per sale would we stop funding this channel?
A solid answer proposes moving budget toward production and tells an angle apart from a variant. A hollow answer proposes refining the audiences, testing new settings, or raising the budget to let the system learn.
Deciding how to split a budget between account management and creative production is the trade-off this change forces. A 90-minute consultation settles it, with a written summary you can circulate in your organization.
What actually changes inside an ad account
Three shifts structure everything that follows. They reinforce one another, which is why accounts that fix only one of them see little difference.
| Element | Previous logic | Andromeda logic |
|---|---|---|
| Targeting | Interest segments, lookalikes, manual exclusions | Broad audiences, the engine rebuilds the matches |
| Structure | Five to ten ad sets per campaign | Few ad sets, consolidated budget, concentrated learning |
| Creative | Three to five near-identical variations | A portfolio of distinct angles, varied formats and messages |
| Signals | Pixel alone, events with no value | Pixel and Conversions API, events with values and deduplicated |
| Measurement | Cost per click, cost per thousand | Return on ad spend, cost of acquisition, margin on spend |
The creative row deserves a clarification, because it is the one most often misread. The engine does not reward the volume of creatives. It rewards the distance between them. Twenty variations of the same visual with a different headline occupy a single position in its representation space. Three genuinely distinct angles occupy three.
Producing more creatives without diversifying the angles is the most expensive mistake in this transition. You multiply the production hours and the account sees only a single proposition. The cost rises, the return does not move.
The four most common mistakes since the rollout
Keeping a fragmented structure
Ten ad sets split the conversions across ten learning pools that are too thin. The engine takes longer to understand what works, and on low-volume accounts sometimes never gets there.
Stacking exclusions out of habit
Each exclusion removes eligible candidates before evaluation even happens. On broad targeting, a stack of exclusions inherited from old campaigns amounts to throttling the engine without knowing it.
Treating variation as diversity
Changing a button's color or a headline's hook does not create a distinct signal. A different angle implies a different promise, a different format or a different audience in the narrative.
Neglecting the quality of the conversions sent
A purchase event with no monetary value, duplicates between the pixel and the API, a missing product identifier: the engine then optimizes toward a blurry target. It is the most widespread flaw and the easiest to fix.
Mapping your creative angles
This is the exercise that decides the return on your production budget, before the first ad even goes live.
This is where most accounts stall. Everyone understands that creative diversity is needed. Almost no one, looking at their own library of files, can say whether it is diverse or not. So they produce more, hoping volume will stand in for variety.
Here is the grid we use on engagements to settle it. An angle is defined by the combination of four dimensions. Two ads that share all four are variants. Two ads that differ on at least two dimensions are distinct angles.
| Dimension | Question | Possible values |
|---|---|---|
| Promise | Which benefit is put forward? | Gaining, avoiding a loss, saving time, cutting a cost, belonging |
| Entry | How does the narrative begin? | A lived problem, a demonstration, third-party proof, a comparison, behind the scenes, a head-on objection |
| Form | What is the reading rhythm? | Short video with an immediate hook, dense static, sequential carousel, filmed testimonial, text on a plain background |
| Speaker | Who speaks? | The brand, a customer, an employee, an independent third party, no one |
The two-dimension test
Take a concrete case. A services company sells management software. It has twelve files live. Run them through the grid and you get this: eleven of them say save time, open with a demonstration of the interface, are dense statics and let the brand speak. They share all four dimensions. The twelfth is a filmed testimonial from a customer telling the problem they had before. It differs on all four.
This account has two angles, not twelve. And the engine sees it that way, even though production cost twelve times the price of one file.
The test is about what the ad says and how it says it, never about the file. Two videos shot separately with the same script and the same angle remain a single proposition to the engine.
Finding the empty cells
The point of the grid is not to judge what exists, it is to reveal what is missing. Once the current ads are positioned, the empty cells jump out and become the production queue.
In the example above, no ad takes on an objection head-on, none lets an employee speak, none compares the solution to the alternative the customer uses today, namely a spreadsheet. Three obvious angles, none produced. That is almost always the result of the exercise: the empty cells are numerous and they were invisible before the grid was drawn.
How many angles, and at what pace
Four to six genuinely distinct angles are enough for most SMB accounts. Below four, the engine has nothing to compare. Above six, you scatter production without the account having the conversion volume needed to tell them apart.
The renewal pace matters as much as the number. An angle wears out, not because the audience tires of it, but because the engine has finished measuring its useful reach. You then replace the weakest angle rather than adding to the pile, which keeps the library legible and the budget concentrated.
Diversifying is not contradicting yourself. The four dimensions vary, the brand promise does not. An account whose angles tell four different companies gains in diversity what it loses in recognition, and the return rarely follows.
The exercise, in one hour
The system improves performance by relying on deep neural networks at the retrieval stage rather than on features defined by humans.
Meta EngineeringWhat stays in-house is the shooting capacity and the raw material: products in real situations, the people who use them, actual cases. It has become the limiting resource on this channel, and it cannot be bought externally without access to your operations. What can be delegated is the rest: account structure, conversion-signal hygiene, angle definition, editing, rotation and reading the cost per sale. The right question to ask a provider is no longer how many management hours, it is how many distinct angles they will help you produce per quarter.
How to adapt an account without breaking what works
The order matters. Fixing the creative before the signals amounts to feeding an engine that optimizes toward the wrong target. So you always start from measurement.
1. Clean up the conversion signals
Check that the Conversions API runs alongside the pixel, that purchase events carry a monetary value, that the product and currency parameters are present, and that no event is counted twice. It is the only step that improves results without changing anything in the campaigns.
2. Consolidate the structure
Consolidate toward one objective per campaign and a reduced number of ad sets. This step often lowers results for one to two weeks, while the learning rebuilds. It has to be announced before you do it, otherwise it gets judged on the wrong days.
3. Fill the empty cells of the grid
The mapping described above gives the production queue. You fill it with the missing angles rather than variants of existing ones, and you replace the weakest instead of stacking.
4. Shift the tracking indicators
Cost per thousand and cost per click become diagnostic indicators, not steering ones. Return on ad spend, cost of acquisition and margin on ad spend take their place. Sales metrics move ahead of traffic metrics.
To check before concluding that Andromeda is penalizing your account
The three conditions to meet before opening that channel are in Instagram advertising.
This channel is compared to the others in concentrating budget on one channel.
Meta advertising is one lever of the Optimize the profitability of your digital campaigns goal.
Already running a marketing team? See how we plug in as reinforcement on Meta advertising.
Frequently asked questions about Andromeda
What is Meta's Andromeda update?
Andromeda is Meta's ad retrieval system. It filters tens of millions of candidate ads down to a few hundred before ranking. It relies on deep neural networks rather than predefined rules.
Should you drop detailed targeting on Meta?
In most cases, yes. The engine rebuilds the matches between a person and an ad on its own, often better than a manually defined segment. Detailed targeting keeps its use on very narrow markets or when a regulatory constraint imposes a restriction.
How do you know if your ads are truly diverse?
By coding each ad on four dimensions: the promise put forward, the entry of the narrative, the form and the speaker. Two ads that share all four are variants of a single angle. At least two of them have to differ to speak of distinct angles.
How many creative angles should you produce?
Four to six genuinely distinct angles are enough for most SMB accounts. Below four, the engine lacks material to compare. Above six, you scatter production without the conversion volume needed to tell them apart.
Why did my results drop after the rollout?
Three causes recur constantly: a structure too fragmented that dilutes the learning, incomplete conversion signals that steer optimization toward a blurry target, and a homogeneous creative portfolio that gives the engine only a single proposition to test.
Does Andromeda also affect awareness campaigns?
Yes, since it acts at the retrieval stage, upstream of all objectives. The effect is simply less visible on campaigns without conversion, because signal quality plays a smaller role there than on sales objectives.
- Meta Engineering, Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine, accessed July 2026.
- NVIDIA, Grace Hopper Superchip, accessed July 2026.
- Meta for Business, Conversions API documentation, accessed July 2026.
Gabriel almost always takes your first call and carries out your audit. He builds the strategy starting from your growth goal: where to put your budget, which market to test and how to connect each lead to a real sale in your CRM. He mainly leads engagements for three goals: Optimize the profitability of your digital campaigns, Develop a new market, and Generate demand and growth. With Geneviève, he also works on organic search (SEO), AI visibility (GEO) and conversion rate optimization (CRO). The sales a Google Ads or Meta Ads campaign brings in depend on the page that receives the click. He writes mainly about marketing strategy, paid advertising and measurement.
About Falia →
SEO
Paid advertising
Conversion
AI visibility
Web design