Meta's ads system now operates across four major systems:
- Andromeda finds a shortlist of ads for each impression. Think: “which ads are even worth considering for this person right now?” Engineering at Meta
- Lattice ranks that shortlist to pick the winner. Think: “which ad on this list is most likely to drive the goal?” ai.meta.com
- Sequence Learning changes how the system learns by looking at the order and recency of people’s actions, not just static profiles. Think: “what did they do first, then next, then recently?” Engineering at Meta
- GEM is a large model Meta references as part of the ranking stack. You can’t “toggle” it; you improve what it learns from by giving Meta clean signals and genuinely different creative concepts. Facebook+1
If there’s one takeaway from all of this: give Meta clean signals, clean data, and genuinely different creative concepts. Consolidate structure so each ad set gets enough conversions. Your “targeting” is now mostly your events and your creatives.
1) Andromeda: How Ads Get Considered
What it does: From millions of ads, it quickly pulls a relevant set for this person at this moment. If your ad never makes the shortlist, it never gets a chance to win. Engineering at Meta
Why:
- Creative diversity matters more than micro-targeting. If your 10 ads are near-duplicates, Andromeda treats them as the same idea and you limit your chances of reaching different buyers.
- Coverage across placements and formats increases chances to be retrieved.
Real-life example (India DTC, supplements):
You run a Magnesium product. Instead of 12 edits of the same UGC testimonial, run 5 truly different concepts:
- “Sleep before-and-after” routine demo.
- “Migraine day-in-life” manage-the-day story.
- “Gym recovery” with DOMS focus.
- “What a doctor checks” authority explainer.
- “Mom-life” stress calming use-case.
These are five genuinely different concepts. Expect more stable delivery than 12 clones with swapped captions. Practitioners reporting “random spend spikes” post-Andromeda usually fed Meta lookalike ads and near-duplicates. Give it different ideas. Engineering at Meta