How the Meta Ads algorithm picks ads: Andromeda, GEM and GenRecs

A year ago Meta described four elements of its ads ranking system. At the end of October 2026 it published a follow-up: three changes to Andromeda, GEM and GenRecs. According to Meta, ranking improvements drove 8.3% more clicks and 15.7% more conversions on Facebook in Q2 2026.

All three changes are already live on every account. Meta says you don't need to turn on any setting or switch to a new campaign type. The effect does depend on three things: the goal you optimize for, the quality of the data you share (including how you measure results), and how much optimization you hand over to the models.

Below we explain how each of the three elements works and what it means for creative, campaign goals and measurement. The numbers come from Meta. The practical conclusions are ours, and we label them as such.

How Meta picks an ad

When someone scrolls Facebook or Instagram, Meta's system picks an ad for them from tens of millions of active ads. It does this in stages.

• Retrieval. Andromeda narrows the pool down to a short list of candidates for that person. According to Meta, millions of ads become a few thousand.
• Ranking. Ranking models predict how likely that person is to click or convert on each ad on the list. These models are taught by GEM, Meta's large foundation model.
• Understanding content. GenRecs adds a description of what is in the ad, written by a language model. This lets the system know what an ad is about even before anyone has seen it.

GenRecs feeds its description of the ad's content into both retrieval and ranking.

Only ads from the short list make it to ranking. That is why Meta says better retrieval at the start leads to more clicks and conversions later on.

Andromeda: more of your ads make it into retrieval

Andromeda is the ad retrieval engine Meta launched in 2024. In 2026 it got a new architecture. Instead of one general model, it routes each ad to specialized expert models. Meta says this lets retrieval assess ads almost as accurately as the ranking stage and evaluate every eligible ad, at a speed of more than a billion ads per second.

Meta describes the change like this: retrieval used to look mainly at the ads the system knew best. Now it looks at the whole pool in the same amount of time. For advertisers, this means the system considers more of your ads, creative variants and products for each person.

Results according to Meta: since 2025 Andromeda has driven 2.4% more conversions on Facebook. In Q3 2026 Meta made the model ten times more complex without extra compute. This added +1.37% conversions on Facebook and +0.24% on Instagram Reels.

What this means in practice (our interpretation). Retrieval now also assesses ads the system knows less about, such as newer variants in an ad set. Ranking still decides whether they are shown. A wider range of creative and products raises the chance that something from it reaches the right person.

GEM: the model that teaches the other models

GEM (Generative Ads Recommendation Model) is an ad recommendation model trained at the scale of large language models. When Meta described it in 2025, it ran on Facebook Feed and Instagram. Today it is the foundation model for ad recommendations on Facebook, Instagram and Reels, and for conversions outside Meta's apps. It supports all major campaign objectives, from website conversions to app installs.

GEM passes what it has learned to smaller models that set rankings in real time on every surface and for every objective. An improvement to GEM therefore reaches all surfaces at once. The new version analyzes twice as long a history of a user's interactions with ads, and the model itself is 2.5 times more complex.

For advertisers, one sentence in the announcement matters most. Campaigns with few conversions benefit from what GEM has learned from click data in other campaigns. Meta says this helps advertisers who optimize for events deep in the funnel, such as purchases.

Results according to Meta: since 2025 GEM has driven up to 18.9% more conversions and a 12% higher CTR across Meta's surfaces and objectives. In selected user groups it drove +5.8% conversions on Facebook, +4.1% on Instagram and +1.6% for conversions outside Meta.

GenRecs: a language model reads your ad

GenRecs (Generative Recommender) adds large language models with knowledge of the world and of ads to the recommendation system. Such a model analyzes the content of the ad and the user's preferences at the same time.

The first element is tokenization. A language model reads the text and images of each ad and describes it with a set of semantic tokens, markers of its most important features. Meta gives examples: product brand, color, cut and style. The tokens form a shared vocabulary for the whole system. Thanks to it, a new ad with no delivery history lands right away next to similar ads and products the system knows a lot about.

The second element is generative retrieval. For a specific person, GenRecs describes in the same tokens what the ideal ad for them would look like. It then looks for the closest ads available in the system. This mechanism strengthens Andromeda at the retrieval stage.

Meta lists two effects for advertisers:

• New ads, refreshed creative and new products start performing sooner, because the system understands them through their similarity to what it already knows.
• The system can reach a person looking for what you sell, even if nothing in their past clicks resembles your ad.

Results according to Meta: in Q3 2026 generative retrieval on Facebook Feed drove almost 2% more conversions in selected segments. Tokenization alone, combined with GEM and the Adaptive Ranking model, drove +0.35% conversions across all surfaces.

The numbers Meta shared and how to read them

All the numbers come from Meta's announcement and come without a methodology. Some are "up to" values, some apply to selected user segments. Treat them as a description of where the system is heading. Your account's results depend on your goal, data and creative.

You can't add these percentages up, because they cover different periods, surfaces and groups. Data as of October 2026. We update the table when Meta shares new numbers. The only way to see what these changes do for your account is an incrementality test (more in the measurement section).

What to change in your creative

Meta gives no creative recommendations in the announcement. We draw the conclusions below from how GenRecs and Andromeda work.

Show the product and its features so they can be read. The language model describes an ad based on its text and image. When the creative shows the product and the copy says what it is and who it is for, the model has something to build a description from. A mood-driven creative that makes it hard to tell what you sell gives it less to work with. In services and B2B, this role falls to the name of the service, the customer's problem and the industry you are speaking to.

Make variants differ in what the model will record in its description. Tokens describe the features of the ad and the product. Two variants with the same product and a headline reworded will probably get a very similar description. A different product, a different person in the photo or a different problem in the headline is a difference the model can see.

Add new creative regularly. Meta says new ads and refreshed creative start performing sooner, because the system understands them through their similarity to ads it already knows. So you can refresh creative more often instead of waiting for the old one to wear out.

In e-commerce, look after the titles, descriptions and images in your catalog. Meta mentions new products alongside new ads. Catalog ads take their text and images from the product feed, so the feed is what the model builds the product description from. This is our interpretation, consistent with how Meta describes tokenization.

If you need more varied creative for testing, see how we make AI creatives and UGC videos.

Optimization goal and campaign structure

Meta closes the announcement by saying the system still needs you to tell it what a win looks like for your business. It says nothing about campaign structure. The conclusions below are ours.

Optimize for an event that has value for the business. If you make money selling online, the goal is a purchase. In lead gen, a qualified lead works best: one that your sales team has judged worth pursuing. Until now, a low volume of such events was a common reason to optimize for a shallower goal, such as a form submission. GEM reduces this problem, because campaigns with few conversions benefit from knowledge gathered on other data. Check it with a test: run a second campaign alongside your current one, optimized for an event deeper in the funnel, and compare the cost of acquiring a customer.

Leave the models room to optimize. This is the third item on Meta's list: how much optimization you hand over to the models. In practice, that means broad audiences or Advantage+ audience, automatic placements and fewer ad sets, each collecting more data. Narrow interests and hand-picked placements shrink the group of people for whom your ad can make it into retrieval at all.

Data and measurement

The second item on Meta's list is data quality, including how you measure results. GEM learns from signals across the whole platform and also covers conversions outside Meta's apps. The more complete the conversion data from your site and CRM, the better the model knows who buys.

What we check on a client's account:

• Conversions API runs alongside the pixel, and events from both sources are deduplicated.
• The purchase event passes the order value, so the system can optimize for revenue.
• In lead gen, the qualification result from the CRM goes back to Meta as a separate event.

We usually implement Conversions API through server-side GTM. How Conversions API itself works and how to set it up is covered in our article on Meta CAPI. To check whether the Meta pixel on your site sends more than page views, run our free GA4 audit, which also reads the pixel's events.

Measurement is a separate matter. Meta calculated the numbers in the announcement itself, and the reports in Ads Manager also show Meta's attribution. If you want to know how many sales your campaigns drive beyond what would have happened anyway, you need an incrementality test with a control group, such as Conversion Lift in Meta Ads. It is best to run one before the season in which you spend the most.

Want us to check your account for this? See how we run Meta Ads campaigns, or get in touch.