What is Meta's Andromeda

In December 2024, Meta announced a next-generation personalized ad retrieval system called Andromeda.

Meta continuously tweaks its ad-serving algorithm. Performance marketers adapt to it and eventually find ways to game it. This cat-and-mouse game has been playing out for many years after Apple’s ATT changes back in 2021. The dark art of understanding and gaming Meta ads has been evolving.

The latest chapter in this dark art was flooding the Meta ad-account with so many creatives that ad creative volume became the name of the game. Because making creatives has become so cheap and because Meta is pushing for automated media buying, creative volume is increasing in orders of magnitude.

Creative Volume: Problem & Opportunity

To make ads more relavant, Meta needs advertisers to make a high volume of creaitve. However, high volume of ad creatives makes it difficult for Meta to offer better and more personalized ad recommendations. Meta’s ad system needs to evolve to find and score the most relevant ads for the user when they open the app. And Meta’s ad system must do it very quickly (in less than 300ms).

Meta needed a better way to select the "right" creatives, given the vastly increased scale of creatives. It also wanted to incentivize advertisers to make creatives that are more diverse and appealing to the users.

How Meta's Ad System Works?

Meta's ad system has two main stages:

Stage 1 : Retrieval

Meta has a massive inventory of hundreds of millions of ads. When a user loads their Facebook or Instagram feed, Meta’s ad system cannot possibly score every single ad for the user in real-time. The retrieval system's job is to instantly scan this entire inventory and retrieve a smaller, more manageable subset (e.g., the 1,000 most likely relevant ads) from the millions available. Kinda whittle down the list from 100s of millions of ads to a few 1000.

Stage 2: Ranking/Auction:

A second, more sophisticated ranking model predicts user and advertiser value to determine the final set of ads to be shown to the user. It takes that smaller subset of ads, say 1000, and carefully scores and ranks them to decide which single ad is the most relevant to show the user (in a specific ad slot).

What is Andromeda?

Andromeda is the next-generation ad retrieval system (stage 1).  Meta designed a neural net to work specifically on NVIDIA's Grace Hopper Superchip. It is a state of the art engineering of hardware and software on the latest chips to produce an ad retrieval system.

They trained this neural net on user data, higher-order interactions from people and ads data.

Andromeda has two goals:

1. Narrow down ad candidates: From millions of ads, whittle down to a much smaller number, like 1000, of relevant creatives for the user. This number could be anyone's guess, but 1000 seems reasonable.

2. Speed: Narrow down to that smaller list of ads very quickly. Likely in under 200ms or something ridiculous like this. Andromeda should work very fast in real-time to be useful to users of Facebook and Instagram.

How Andromeda Works?

Andromeda first organizes ads in a unique tree-like structure, and then traverses this structure based on the user’s preferences and interactions from the past.

Organizing: When you upload a new ad creative, Meta analyzes various themes, sub-themes and organized the creative a tree based graph. The leaf node in this tree is the ad creative itself. Andromeda creates a hierarchical index of ads. It is essentially a sophisticated filing system for millions of ads. Comparing every ad against a user's profile would be very slow and expensive. That's why Andromeda organizes ads into a multi-layered structure, much like a family tree.

Ads that are visually and thematically very similar are grouped together and placed in the same node. For example, ten ads that use the same image but have slightly different headlines would be clustered tightly together in a low-level node, typically leaf nodes. This node, in turn, might be part of a larger branch representing a broader creative theme, like social proof.

Retrieval: When a user opens their app, Andromeda doesn't need to analyze every single ad. Instead, it navigates the hierarchy. It can quickly discard entire top-level branches (e.g., "humor-based ads") if they are not relevant to the user, thereby ignoring thousands of ads in a single step. It then drills down into more promising branches, eventually reaching the most relevant nodes.

Why Is Creative Similarity Important?

If you provide ten nearly identical ads, they are all filed under the same node. The retrieval model sees this node as representing a single creative concept. All visually similar ad creatives with the same thematic nodes are grouped together as a single entity. Meta has even called it Entity Id .

Creative similarity is the organizing principle for Andromeda's hierarchical index. By pre-sorting and clustering ads based on similarity, the system can focus and can become very fast. This is why providing visually and conceptually distinct ads is so critical; it creates new, valuable nodes in the index for the AI to work with.

What Should Advertisers Do?

Advertisers should make creatives that helps Meta generate more unique Entity Ids. They should do two things:

1. Maximize the number of thematic nodes: During creative strategy, apply more themes and concepts while evaluating ad creatives against existing themes and concepts. The more variety you have, the better your ads are going to perform. Why ? More concepts means more nodes which in turn means that Andromeda will retrieve more of their ads. Hence, more of theri ads will participate in the auction and therefore have a chance of being shown to the user.

2. Reduce the number of visually similar creatives in a thematic node: Don't make too many ads that visually look the same for the same concept. Make ads that could be visually similar but for