Next-generation Meta Andromeda: what really changes for your campaigns in Morocco.
On October 28, 2026, Meta presented the next generation of Andromeda, the system that decides which ads can enter the auction. There is no setting to turn on. For an advertiser in Morocco, the change plays out elsewhere: the real variety of your creative, the quality of your conversion signal and the way you measure.
What Meta announced on October 28, 2026
Andromeda is the first stage of Meta's ad delivery: retrieval. For each person, this system narrows tens of millions of ads to a few thousand candidates. Only those move on to detailed ranking and then the auction. An ad dropped here has no chance of reaching that person, whatever your budget. Launched in December 2024, the system has now moved to a new generation. According to Meta's engineering post, it scores every eligible ad at close to ranking-model quality, at over one billion ads per second on a single GPU server. The same day, Meta published an advertiser summary covering three building blocks: Andromeda, the GEM model and GenRecs. The name Andromeda 2.0 circulates widely, but Meta does not use it: the company speaks of a next generation. It gives no exact production date either. The post says the system is in production without saying since when. Be wary of any explanation that ties a drop in your results to one specific day.
- Retrieval: tens of millions of ads narrowed to a few thousand per person.
- New: every eligible ad is scored, at over one billion per second.
- Related news: GEM for ranking, GenRecs for understanding ads.
- Not stated by Meta: the production date of this version.
What changes technically, without the jargon
The system used to rely on one model to represent each user. It now uses several specialized models, called experts. A routing network sends each ad to the experts most relevant to that person, and only those do the work. Meta reports about 300 times less inference compute than the 2024 retrieval model, while approaching the quality of a ranking model. The other change is in the hardware. Checking your targeting rules, scoring and selecting the best ads now all run on the same GPU server, which cuts network traffic by about 100 times. That speed lets the system examine a larger pool of ads within the same few milliseconds. Why does Meta need this? Because the pool is exploding. Its generative creative tools have produced hundreds of millions of new ads, and nearly 10 million advertisers used them in the third quarter of 2026. For you, that means more competing ads at this first filter. GenRecs adds a reading of the content: language models describe the text and image of each ad, which helps recent ads with little history.
- Several expert models replace a single user model.
- About 300 times less inference compute than the 2024 model.
- Targeting, scoring and selection combined on one GPU server.
- GenRecs describes ad text and images, which helps new creative.
How to read Meta's numbers
Three figures are circulating, and they do not measure the same thing. The +2.4% in Facebook conversions covers the period since 2025, with no detailed method. The +1.37% on Facebook and +0.24% on Instagram Reels relate to third-quarter 2026 improvements, when model complexity rose tenfold with no added compute cost. The former was rounded to 1.4% in Meta's earnings presentation. Do not add them up: Meta does not say whether the second is included in the first. The often-quoted +15.7% in conversions refers to something else: the whole ranking system on Facebook in the second quarter of 2026. The 18.9% credited to GEM is a maximum, worded as up to, depending on surfaces and objectives. Two limits apply to all these results. They are averages published by Meta, unaudited, and none is broken down by country. Nothing supports claiming a gain in Morocco. The 100 times comparison also sets a per-GPU-card rate in 2024 against a per-server rate in 2026, and Meta does not say how many cards that server holds. Treat these figures as a direction, not a promise.
- +2.4%: Facebook, since 2025, method not detailed.
- +1.37% Facebook and +0.24% Reels: third-quarter 2026 improvements.
- +15.7%: all ranking on Facebook in the second quarter, not Andromeda alone.
- No country figures: no gain to extrapolate for Morocco.
What this changes for your creative
In its guide to how its auction works, Meta recommends supplying more creative variety so the system has material to test. It publishes no official number. The thresholds in circulation, such as a 20-ad minimum or a 60% similarity limit, come from agency blogs and tool vendors, and they contradict each other. We found no Meta documentation confirming them. The logic itself stays simple. If Meta reads the content of each ad, five versions of the same visual with different headlines do not open five doors. A real new concept changes the angle, the person addressed, the format or the stage of the buying journey. In Morocco, you have a natural lever: language and context. A talking-head video in Darija for a buyer who pays on delivery, a carousel in French for a B2B buyer and customer proof in Arabic all address different needs. Our working benchmark, which is not a Meta rule: start with 3 to 5 distinct concepts per campaign, then use our creative testing method to decide which ones to keep.
- Meta asks for more variety, with no official number of ads.
- The similarity thresholds in circulation are not confirmed by Meta.
- A distinct concept changes the angle, audience, format or buying stage.
- Webotic benchmark: 3 to 5 distinct concepts per campaign to start.
What stays in your hands: goal, data, targeting
Meta says it plainly: there is nothing to turn on, but the result depends on three things. The outcome you give the system, the quality of your data and the room you leave the models. The goal comes first. As Meta puts it, a lead is not a sale. If you optimize for the cheapest form fill, a more powerful system will simply find poor-quality form fills faster. Send qualification back through the Conversions API instead, as in our guide to lead quality scoring. On an e-commerce site, also send order value, separating delivered orders from those refused on delivery where you can. Then comes data. A properly installed Pixel and CAPI remain the prerequisite. An over-fragmented budget also starves the system of learning: Meta aims for about 50 conversions per week per ad set. At a CPA of 100 MAD, that is already 5,000 MAD per week. Targeting, finally, does not disappear. Andromeda still applies your rules before scoring ads: an area limited to Casablanca stays limited to Casablanca. Stacking restrictions, however, shrinks the pool of people you can reach. Our article on Advantage+ Audience explains what to keep.
- Optimize for the real outcome: sale, qualified lead, delivered order.
- Send value and qualification back through the Conversions API.
- Consolidate budget to aim for about 50 conversions per week per ad set.
- Keep useful limits, such as your delivery area, and drop the others.
A 30-day action plan for advertisers in Morocco
In week 1, check your measurement. Verify deduplication between Pixel and CAPI, event match quality and the real goal of each campaign. Record your CPA, ROAS and qualified lead rate for the last 30 days: this is your comparison baseline. In week 2, simplify. Merge ad sets aimed at the same audience, and keep a split only where it reflects a real difference in offer, margin or area. In week 3, produce distinct concepts. List your customer objections, then give each one an angle, a format and a language. For e-commerce, see also our Advantage+ Shopping guide. In week 4, measure on your own data: confirmed sales, leads called back, delivered orders. Meta recommends judging on cost per result or ROAS and testing incrementality with a Conversion Lift study when your account has access. Check what happens after the click too. Meta's ad quality score takes in shipping delays, unexpected charges and customer satisfaction, so a disappointing service weighs on delivery. To frame your spend, see our Facebook Ads pricing in Morocco.
- Week 1: audit tracking and lock a comparison baseline.
- Week 2: merge overlapping ad sets.
- Week 3: produce distinct concepts from customer objections.
- Week 4: measure on real sales and leads, not on Meta's average.
FREQUENTLY ASKED QUESTIONS
- What is Meta Andromeda?
- Meta Andromeda is the machine learning system that handles the first stage of Meta's ad delivery. For each person, it narrows tens of millions of ads to a few thousand candidates. Those ads then go through detailed ranking and the auction. Launched in December 2024, Andromeda received a next generation presented on October 28, 2026. It scores every eligible ad at over one billion ads per second on a GPU server. It is not a product you turn on in Ads Manager.
- Does Andromeda 2.0 really exist?
- The name Andromeda 2.0 comes mainly from the trade press and agencies. Meta speaks of the next generation of Andromeda in its October 28, 2026 publications. The update is real: specialized expert models, scoring of every eligible ad and a pipeline combined on one GPU server. Meta says the system is in production but gives no launch date. So you cannot reliably tie a change in your results to one specific day, especially without comparing your seasonality, creative and bids.
- Do I need to change my Meta campaigns because of Andromeda?
- Not to receive the gains: Meta says there is no setting to turn on, no new campaign type and no migration. You do if your account holds the system back. Check three points: a goal that matches a real outcome, a reliable conversion signal through the Pixel and Conversions API, and a structure simple enough for each ad set to get enough conversions. On an account that is already clean, refreshing creative concepts is often enough. Avoid rebuilding everything without a measured reason.
- How many creatives do you need for Andromeda?
- Meta publishes no official number. It recommends supplying more creative variety without setting a threshold. Figures such as a 20-ad minimum or a 60% maximum similarity come from third-party blogs and are not confirmed by Meta. Our working benchmark is to start with 3 to 5 genuinely distinct concepts per campaign, then adjust to your budget. A distinct concept changes the angle, audience, format or buying stage. Changing only a headline or a color does not create a new concept.
- Does targeting still matter with Andromeda?
- Yes. According to Meta's engineering post, Andromeda applies the advertiser's targeting rules before scoring ads. A geographic limit, a language or an exclusion is still respected. What loses value is stacking narrow interests, because the models infer much of the relevance from behavior and ad content. In Morocco, keep the constraints tied to your business, such as your delivery area, and leave the rest of the audience search to the models.
- Does the 2.4% gain apply to advertisers in Morocco?
- Nobody can say from the published data. The +2.4% figure concerns Facebook conversions since 2025, and Meta gives neither a detailed method nor results by country. It is an average at platform scale, not a forecast for your account. Your outcome depends on your offer, creative, tracking and the competition in your sector. To judge a real effect, compare your own sales or qualified leads before and after, or use a Conversion Lift study if your account has access.