Meta rebuilt the system that decides which ads even get a chance to show. Here's what their engineering post actually says, in plain English, with examples.
Meta engineering post · 28 Oct 2026 · explained for media buyers Prepared by Sami Siddique · November 2026
<aside> 🏷️
How to read this page
<aside> ⚡
The short version
Every time someone opens Facebook or Instagram, Meta has to pick an ad from millions of options in a few milliseconds. It can't study every ad carefully, so it works in stages.
| Stage | What happens |
|---|---|
| 0 · All ads | Every live ad from every advertiser. |
| 1 · Targeting filter | Removes ads whose targeting rules exclude this person. |
| 2 · Shortlist (Andromeda) | Scores every eligible ad fast and keeps the best few. |
| 3 · Ranking | Studies the shortlist in detail. |
| 4 · The ad shown | One winner gets the slot. |
Think of hiring
Andromeda is the recruiter. A million resumes come in. The recruiter screens them fast and passes a short pile to the hiring manager.
The old weakness
The recruiter used a simple checklist because there were too many resumes. If a great candidate got cut there, the hiring manager never saw them. Same for your ads.
[Meta says] The new retrieval model reaches quality that "approaches ranking-level." In hiring terms, the recruiter is now nearly as sharp as the hiring manager, and still just as fast.
[Meta says] Generative AI creative tools helped advertisers make "hundreds of millions of new ads." The pool of ads exploded.
More ads means more work for the shortlist stage. The 2024 system checked tens of millions of ads per second per chip. Meta needed billions.
[Meta says] "Small improvements in retrieval model performance can drive large gains in advertiser value." Getting the right ad onto the shortlist matters a lot.
[My read] Your ads now compete against far more ads than a year ago, a lot of them AI-made. Generic ads are easier to replace.
Meta calls the new model a "Mixture of Experts." Instead of one generalist judging everything, there's a panel of specialists, and a router that calls in only the relevant ones for each ad. Large AI language models use the same trick.
To make it concrete, we'll follow one person.