Meta Andromeda: how copywriting changes ad delivery
Meta Andromeda: how copywriting changes ad delivery
Meta Andromeda is the artificial intelligence-powered ad retrieval system that Meta began operating in late 2024. Instead of starting with the advertiser-defined audience segmentation, Andromeda first analyzes the ad's copy, creative, and format to predict which user is most likely to engage and convert.
Moisés Schwan Bach
Operations Manager · First Brazil
Operations Manager
Introduction
Anyone who manages paid traffic campaigns on Meta Ads noticed, throughout 2025 and 2026, that ads with the same targeting structure started delivering differently. Part of this change has a name: Andromeda, the ad retrieval system that Meta rebuilt to decide, among tens of millions of active ads, which ones reach each person.
The central change is not subtle. Andromeda inverted the delivery logic: instead of using the audience configured by the advertiser as a starting point, the system first reads the ad content (copy, image, video, format) and uses this content to predict behavior. This shifts the center of gravity of the paid media strategy: targeting is no longer the main lever, and copy has become a delivery signal.
This article explains how Andromeda works, why copy has gained weight, what changes in the campaign structure, and which writing errors now harm ad distribution even before it reaches the auction stage.
Introduction
Anyone who manages paid traffic campaigns on Meta Ads noticed, throughout 2025 and 2026, that ads with the same targeting structure started delivering differently. Part of this change has a name: Andromeda, the ad retrieval system that Meta rebuilt to decide, among tens of millions of active ads, which ones reach each person.
The central change is not subtle. Andromeda inverted the delivery logic: instead of using the audience configured by the advertiser as a starting point, the system first reads the ad content (copy, image, video, format) and uses this content to predict behavior. This shifts the center of gravity of the paid media strategy: targeting is no longer the main lever, and copy has become a delivery signal.
This article explains how Andromeda works, why copy has gained weight, what changes in the campaign structure, and which writing errors now harm ad distribution even before it reaches the auction stage.
Introduction
Anyone who manages paid traffic campaigns on Meta Ads noticed, throughout 2025 and 2026, that ads with the same targeting structure started delivering differently. Part of this change has a name: Andromeda, the ad retrieval system that Meta rebuilt to decide, among tens of millions of active ads, which ones reach each person.
The central change is not subtle. Andromeda inverted the delivery logic: instead of using the audience configured by the advertiser as a starting point, the system first reads the ad content (copy, image, video, format) and uses this content to predict behavior. This shifts the center of gravity of the paid media strategy: targeting is no longer the main lever, and copy has become a delivery signal.
This article explains how Andromeda works, why copy has gained weight, what changes in the campaign structure, and which writing errors now harm ad distribution even before it reaches the auction stage.
How Meta Andromeda replaced audience targeting logic?
Meta Andromeda is a retrieval system based on artificial intelligence that, in milliseconds, matches eligible ads for each available impression. According to Meta's engineering team, the system was designed to scale personalization capacity up to 10,000 times compared to the previous model, while maintaining a sublinear inference cost.
Before Andromeda, the decision flow started with the audience: interests, demographics, and lookalike audiences defined who saw what. Andromeda inverted this order. It first evaluates the engagement history, copy, creative, and ad format, and from there predicts which users are most likely to react in a way that is relevant to the campaign goal.
In practice, this means that detailed targeting fields (interests, behaviors) began to function as suggestions, not as rigid restrictions. Since the February 2026 update, Meta treats these inputs as general guidance for the algorithm, rather than a definitive filter of who can or cannot receive the ad.
The system works in two complementary layers: Andromeda filters the universe of possible ads and produces a reduced set of plausible candidates; GEM (Generative Ads Recommendation Model), the next layer, decides which of these candidates should actually be displayed to each person at that moment. Andromeda decides what can appear; GEM decides what should appear.
How Meta Andromeda replaced audience targeting logic?
Meta Andromeda is a retrieval system based on artificial intelligence that, in milliseconds, matches eligible ads for each available impression. According to Meta's engineering team, the system was designed to scale personalization capacity up to 10,000 times compared to the previous model, while maintaining a sublinear inference cost.
Before Andromeda, the decision flow started with the audience: interests, demographics, and lookalike audiences defined who saw what. Andromeda inverted this order. It first evaluates the engagement history, copy, creative, and ad format, and from there predicts which users are most likely to react in a way that is relevant to the campaign goal.
In practice, this means that detailed targeting fields (interests, behaviors) began to function as suggestions, not as rigid restrictions. Since the February 2026 update, Meta treats these inputs as general guidance for the algorithm, rather than a definitive filter of who can or cannot receive the ad.
The system works in two complementary layers: Andromeda filters the universe of possible ads and produces a reduced set of plausible candidates; GEM (Generative Ads Recommendation Model), the next layer, decides which of these candidates should actually be displayed to each person at that moment. Andromeda decides what can appear; GEM decides what should appear.
How Meta Andromeda replaced audience targeting logic?
Meta Andromeda is a retrieval system based on artificial intelligence that, in milliseconds, matches eligible ads for each available impression. According to Meta's engineering team, the system was designed to scale personalization capacity up to 10,000 times compared to the previous model, while maintaining a sublinear inference cost.
Before Andromeda, the decision flow started with the audience: interests, demographics, and lookalike audiences defined who saw what. Andromeda inverted this order. It first evaluates the engagement history, copy, creative, and ad format, and from there predicts which users are most likely to react in a way that is relevant to the campaign goal.
In practice, this means that detailed targeting fields (interests, behaviors) began to function as suggestions, not as rigid restrictions. Since the February 2026 update, Meta treats these inputs as general guidance for the algorithm, rather than a definitive filter of who can or cannot receive the ad.
The system works in two complementary layers: Andromeda filters the universe of possible ads and produces a reduced set of plausible candidates; GEM (Generative Ads Recommendation Model), the next layer, decides which of these candidates should actually be displayed to each person at that moment. Andromeda decides what can appear; GEM decides what should appear.
Why has ad copy become a delivery signal, and not just a conversion one?
Copy has always influenced click-through and conversion rates. What changes with Andromeda is the moment the system begins to read it: before delivery, not just after. Ad copy is now one of the signals the model uses to decide whom to show the ad to, alongside the image, video, and chosen format.
This happens because Andromeda reads the creative asset as a whole, including the copy, as a semantic representation of the ad's proposition. It then crosses this representation with behavioral patterns of millions of users to predict affinity. Generic text, which could apply to any product in the same industry, generates a weak signal. Specific text, which names the pain point, the context of use, or the audience's decision stage, generates a stronger signal and helps the system find the right buyer more accurately.
This shift explains why campaigns with good copy, but little variation among the ads in the set, have been losing reach. Different copy over the same image, or just a new headline over the same visual template, is not read by Andromeda as a signals-wise new enough element to unlock new audiences.
Why has ad copy become a delivery signal, and not just a conversion one?
Copy has always influenced click-through and conversion rates. What changes with Andromeda is the moment the system begins to read it: before delivery, not just after. Ad copy is now one of the signals the model uses to decide whom to show the ad to, alongside the image, video, and chosen format.
This happens because Andromeda reads the creative asset as a whole, including the copy, as a semantic representation of the ad's proposition. It then crosses this representation with behavioral patterns of millions of users to predict affinity. Generic text, which could apply to any product in the same industry, generates a weak signal. Specific text, which names the pain point, the context of use, or the audience's decision stage, generates a stronger signal and helps the system find the right buyer more accurately.
This shift explains why campaigns with good copy, but little variation among the ads in the set, have been losing reach. Different copy over the same image, or just a new headline over the same visual template, is not read by Andromeda as a signals-wise new enough element to unlock new audiences.
Why has ad copy become a delivery signal, and not just a conversion one?
Copy has always influenced click-through and conversion rates. What changes with Andromeda is the moment the system begins to read it: before delivery, not just after. Ad copy is now one of the signals the model uses to decide whom to show the ad to, alongside the image, video, and chosen format.
This happens because Andromeda reads the creative asset as a whole, including the copy, as a semantic representation of the ad's proposition. It then crosses this representation with behavioral patterns of millions of users to predict affinity. Generic text, which could apply to any product in the same industry, generates a weak signal. Specific text, which names the pain point, the context of use, or the audience's decision stage, generates a stronger signal and helps the system find the right buyer more accurately.
This shift explains why campaigns with good copy, but little variation among the ads in the set, have been losing reach. Different copy over the same image, or just a new headline over the same visual template, is not read by Andromeda as a signals-wise new enough element to unlock new audiences.
How does creative diversity impact campaign performance in Andromeda?
Creative diversity, in the context of Andromeda, means ads that vary in a real way: different pain points, different personas, different emotional triggers, and different formats. It is not about minor budget color variations or headline tweaks on the exact same creative asset.
The practical recommendation that has been consolidating among paid media operators in 2026 is to build 8 to 12 conceptually distinct creative concepts per campaign, with 15 to 20 active ads simultaneously, each combining a different copy angle, format, and trigger. Running too few creatives, even if individually optimized, limits the behavioral data that Andromeda receives to refine delivery.
A practical way to organize production is to split the creative budget into two blocks: half in "anchor creatives," variations on visual and copy structures already validated, updated monthly; half in "exploratory creatives," deliberate tests with different photographers, cast, scenery, and narrative structures from those already used. Exploratory ones have a lower success rate, but they are the ones that tend to open up new pockets of audience, the real mechanism behind long-term account growth. This constant cycle of hypothesis, testing, and creative optimization is, in practice, an application of growth marketing to paid media.
How does creative diversity impact campaign performance in Andromeda?
Creative diversity, in the context of Andromeda, means ads that vary in a real way: different pain points, different personas, different emotional triggers, and different formats. It is not about minor budget color variations or headline tweaks on the exact same creative asset.
The practical recommendation that has been consolidating among paid media operators in 2026 is to build 8 to 12 conceptually distinct creative concepts per campaign, with 15 to 20 active ads simultaneously, each combining a different copy angle, format, and trigger. Running too few creatives, even if individually optimized, limits the behavioral data that Andromeda receives to refine delivery.
A practical way to organize production is to split the creative budget into two blocks: half in "anchor creatives," variations on visual and copy structures already validated, updated monthly; half in "exploratory creatives," deliberate tests with different photographers, cast, scenery, and narrative structures from those already used. Exploratory ones have a lower success rate, but they are the ones that tend to open up new pockets of audience, the real mechanism behind long-term account growth. This constant cycle of hypothesis, testing, and creative optimization is, in practice, an application of growth marketing to paid media.
How does creative diversity impact campaign performance in Andromeda?
Creative diversity, in the context of Andromeda, means ads that vary in a real way: different pain points, different personas, different emotional triggers, and different formats. It is not about minor budget color variations or headline tweaks on the exact same creative asset.
The practical recommendation that has been consolidating among paid media operators in 2026 is to build 8 to 12 conceptually distinct creative concepts per campaign, with 15 to 20 active ads simultaneously, each combining a different copy angle, format, and trigger. Running too few creatives, even if individually optimized, limits the behavioral data that Andromeda receives to refine delivery.
A practical way to organize production is to split the creative budget into two blocks: half in "anchor creatives," variations on visual and copy structures already validated, updated monthly; half in "exploratory creatives," deliberate tests with different photographers, cast, scenery, and narrative structures from those already used. Exploratory ones have a lower success rate, but they are the ones that tend to open up new pockets of audience, the real mechanism behind long-term account growth. This constant cycle of hypothesis, testing, and creative optimization is, in practice, an application of growth marketing to paid media.
What changes in the campaign structure with Andromeda and GEM?
With the delivery decision concentrated in two AI layers (Andromeda and GEM), excessively segmented campaign structures lose part of the purpose they had before - a movement similar to what already occurs in the Google Ads auction. Simpler structures, such as Advantage+ Shopping as a primary campaign, tend to give the algorithm more volume of conversion data per ad set, which improves the quality of optimization.
This does not eliminate the importance of the conversion signals sent by the advertiser. On the contrary: as manual targeting has lost weight, the quality of the conversion signal (Pixel and Conversions API running in parallel, with Event Match Quality consistently above 7) has become one of the few direct controls the advertiser still has over delivery. Clean conversion signals and diverse creatives are now the two most relevant performance levers.
Another practical effect is the speed of the delivery cycle. Ads are tested and worn out by the algorithm faster than in the previous model, which requires a higher rate of creative replacement than many operations were used to maintaining.
What changes in the campaign structure with Andromeda and GEM?
With the delivery decision concentrated in two AI layers (Andromeda and GEM), excessively segmented campaign structures lose part of the purpose they had before - a movement similar to what already occurs in the Google Ads auction. Simpler structures, such as Advantage+ Shopping as a primary campaign, tend to give the algorithm more volume of conversion data per ad set, which improves the quality of optimization.
This does not eliminate the importance of the conversion signals sent by the advertiser. On the contrary: as manual targeting has lost weight, the quality of the conversion signal (Pixel and Conversions API running in parallel, with Event Match Quality consistently above 7) has become one of the few direct controls the advertiser still has over delivery. Clean conversion signals and diverse creatives are now the two most relevant performance levers.
Another practical effect is the speed of the delivery cycle. Ads are tested and worn out by the algorithm faster than in the previous model, which requires a higher rate of creative replacement than many operations were used to maintaining.
What changes in the campaign structure with Andromeda and GEM?
With the delivery decision concentrated in two AI layers (Andromeda and GEM), excessively segmented campaign structures lose part of the purpose they had before - a movement similar to what already occurs in the Google Ads auction. Simpler structures, such as Advantage+ Shopping as a primary campaign, tend to give the algorithm more volume of conversion data per ad set, which improves the quality of optimization.
This does not eliminate the importance of the conversion signals sent by the advertiser. On the contrary: as manual targeting has lost weight, the quality of the conversion signal (Pixel and Conversions API running in parallel, with Event Match Quality consistently above 7) has become one of the few direct controls the advertiser still has over delivery. Clean conversion signals and diverse creatives are now the two most relevant performance levers.
Another practical effect is the speed of the delivery cycle. Ads are tested and worn out by the algorithm faster than in the previous model, which requires a higher rate of creative replacement than many operations were used to maintaining.
Which copy errors compromise ad delivery on Andromeda?
Three recurring errors reduce the effectiveness of copy under Andromeda. The first is generic copy, written for "everyone in the niche," without naming a specific pain point or use context; this type of text generates a weak affinity signal for the algorithm. The second is superficial variation: changing only the headline or CTA over the same image and calling it an A/B test; Andromeda does not interpret this as a new creative. The third is ignoring the hook of the ad: in video formats, the system gives disproportionate weight to the first few seconds, and strong copy over a weak opening loses most of its delivery potential.
Correcting these three points does not guarantee performance, but it removes barriers that currently prevent the algorithm itself from learning from the ad. Specific copy, genuinely different creatives, and strong openings form the minimum foundation for Andromeda to have what it needs to find the right audience.
Which copy errors compromise ad delivery on Andromeda?
Three recurring errors reduce the effectiveness of copy under Andromeda. The first is generic copy, written for "everyone in the niche," without naming a specific pain point or use context; this type of text generates a weak affinity signal for the algorithm. The second is superficial variation: changing only the headline or CTA over the same image and calling it an A/B test; Andromeda does not interpret this as a new creative. The third is ignoring the hook of the ad: in video formats, the system gives disproportionate weight to the first few seconds, and strong copy over a weak opening loses most of its delivery potential.
Correcting these three points does not guarantee performance, but it removes barriers that currently prevent the algorithm itself from learning from the ad. Specific copy, genuinely different creatives, and strong openings form the minimum foundation for Andromeda to have what it needs to find the right audience.
Which copy errors compromise ad delivery on Andromeda?
Three recurring errors reduce the effectiveness of copy under Andromeda. The first is generic copy, written for "everyone in the niche," without naming a specific pain point or use context; this type of text generates a weak affinity signal for the algorithm. The second is superficial variation: changing only the headline or CTA over the same image and calling it an A/B test; Andromeda does not interpret this as a new creative. The third is ignoring the hook of the ad: in video formats, the system gives disproportionate weight to the first few seconds, and strong copy over a weak opening loses most of its delivery potential.
Correcting these three points does not guarantee performance, but it removes barriers that currently prevent the algorithm itself from learning from the ad. Specific copy, genuinely different creatives, and strong openings form the minimum foundation for Andromeda to have what it needs to find the right audience.
Frequently asked questions
What is Meta Andromeda?
Meta Andromeda is the AI-based ad retrieval system that Meta introduced starting in late 2024. It decides which ads are eligible for each impression by reading the creative and copy, rather than from the audience targeting configured by the advertiser.
What is the difference between Andromeda and GEM?
Andromeda filters, from among millions of active ads, a reduced set of plausible candidates for each user. GEM (Generative Ads Recommendation Model) operates in the next layer and decides which of these candidates should be displayed at that specific moment.
Does audience segmentation still matter after Andromeda?
It matters, but with less weight. Since February 2026, Meta treats detailed targeting fields as suggestions for the algorithm, not as mandatory restrictions. The creative and the copy have started to carry more weight in the delivery decision than the manually configured audience.
How many different creatives does a campaign need to have in Andromeda?
The practice that has been consolidating in 2026 is to maintain from 8 to 12 distinct creative concepts per campaign, with 15 to 20 active ads simultaneously, varying copy, format, and emotional trigger among them, and not just small adjustments to the same creative.
First Brazil
First Brazil
First Brazil is a growth agency with operations in Brazil and the United States, specializing in civil construction and real estate, integrating digital marketing, sales structuring, and AI automation. It builds acquisition systems that connect paid traffic, CRM, web development, and sales processes for builders, developers, and modular construction companies.
Talk to the team