What Andromeda does
Every time someone opens Facebook or Instagram, Meta has a fraction of a second to choose a few ads from tens of millions of candidates. The first step, called retrieval, narrows that pool down to the few thousand ads worth ranking for that person. Andromeda is the system Meta built for that step.
Meta introduced it in December 2024 and updated it in October 2026. Today, Meta says, Andromeda “scores every eligible ad” for each person, and “ad content fused with real-time serving context” is what drives the choice. Since 2025 it has delivered a 2.4% increase in conversions on Facebook.
Meta was also clear about why it had to be rebuilt: generative AI creative tools “have helped advertisers rapidly create hundreds of millions of new ads.” More ads and more variations, and a system that now looks at every one of them and decides who each one is for.
Why creative became the targeting
Retrieval is where an ad wins or loses its audience. If Andromeda doesn't match your ad to a person at that first step, no bid and no audience setting brings it back. And what it matches on is, increasingly, the ad itself: what it shows, who appears in it, what it says and how people respond to it.
Two things from Meta make the direction clear:
- Retrieval reads the creative. Meta describes Andromeda's ad input as the ad's content combined with the moment it would be served. The audience is no longer something you pick first and then fill with an ad. The ad's content does the picking.
- Ranking learns from the creative too. Meta's GEM model, which it calls the “central brain” of ad ranking, uses creative representation as one of its inputs. Meta says GEM lifted ad conversions by 5% on Instagram and 3% on Facebook Feed in Q2 2025.
Put together, the platform now does the targeting automatically, and the main input it targets with is your creative.
Pick the audience, then find an ad
- Detailed interest and lookalike targeting
- Many ad sets, each with a narrow audience
- Creative chosen by opinion, tested in market
- Learn what failed after the budget is spent
Make the ad, and the ad finds the audience
- Broad targeting and automated campaigns
- Fewer campaigns, more genuinely different creatives
- Each creative is a hypothesis about who it is for
- Weak creative means fewer people, the wrong people or a higher cost, and you pay to find out
What changes for advertisers
Creative quality now decides reach, not just response. A weak hook or a late brand reveal used to cost you clicks. Now it also decides who the system shows the ad to, and how much it costs to reach them.
Different is not the same as more. Ten versions of the same idea give the system one signal, not ten. What helps is creatives that differ in concept, format, person on screen and message, each strong enough to earn delivery on its own.
Testing in market got more expensive. Every creative that enters an automated campaign needs spend before the system learns what it is good for. Launch six creatives and find out two were weak, and your budget has paid for the algorithm's learning.
It isn't only Meta. TikTok and Google's automated campaigns work the same way: you supply the creative and the goal, and the system chooses the people. The same video often runs on all three, each with its own safe zones, pacing and sound habits.
Why pre-validation across channels is the only way
If the algorithm owns targeting, the creative is the one input you fully control. So the most important question moves to before launch: is this creative strong enough, and right for this objective and this platform, to deserve the spend?
There are only two ways to answer it. Let the algorithm answer with your media budget, one campaign and one platform at a time. Or score the creative before it goes live, fix what is weak, and only feed the machine creative that has earned its place.
- Score before spend. Rate every creative on what the platforms reward: hook, early branding, pacing, message, call to action and platform fit, weighted for the campaign objective.
- Fix it for each channel. Check safe zones, length and sound for TikTok, Reels, Shorts and YouTube, so one idea works in every feed it lands in.
- Validate against live data. Link each creative to its real results across platforms, so the next brief starts from evidence, not opinion.
That is how KRITIQ works: predict creative performance before spend, then check the prediction against real-time data from every channel the creative ran on.
The evidence from our campaigns
We test KRITIQ's pre-spend scores against real results. A creative KRITIQ scored 5.0 drove a 21.3% ad recall lift on TikTok. A fintech creative revised on KRITIQ's advice cut cost per install by 62%. And across five banking creator videos, KRITIQ's ranking matched real brand exposure with a 0.9 rank correlation.
Sources
- Engineering at Meta, “Meta Andromeda: Powering the next generation of ads retrieval through model-system co-design”, 28 October 2026.
- Engineering at Meta, “Meta Andromeda: Supercharging Advantage+ automation with the next-gen personalized ads retrieval engine”, 2 December 2024.
- Engineering at Meta, “Meta's Generative Ads Model (GEM): the central brain accelerating ads recommendation AI innovation”, 10 November 2025.
- KRITIQ case studies: FMCG brand lift, fintech cost per install, banking score vs results and ROI on the platform fee.
KRITIQ is independent and is not affiliated with or endorsed by Meta, TikTok or Google.