The short answer
Meta Andromeda is the retrieval stage of Meta's ad system: it narrows tens of millions of candidate ads down to a few thousand for each person before ranking begins. Because it chooses between creatives, not audiences, the brands that keep many winners live tend to spread them across angles and awareness levels. 57% of the brands with 5+ winners on our board run both video and image.
Andromeda in one paragraph
Andromeda is the machine learning system Meta uses for ad retrieval, the first step of its delivery pipeline. Meta's engineering team describes it as the stage that selects "from tens of millions of ad candidates into a few thousand relevant ad candidates," processing three orders of magnitude more ads than the stages after it. Meta reported a 6% recall improvement and an 8% ads quality improvement on selected segments when it shipped.
The reason it exists is volume. Meta built Andromeda because Advantage+ automation and generative AI tools were multiplying the number of creatives in the system, and the older retrieval setup could not score that many options per person. Practitioner write-ups put the end of the global rollout around October 2025. Meta never announced a single cutover day.
Where it sits in the pipeline
People often say "Andromeda" when they mean "the algorithm." It is one stage of several.
| Stage | What happens | What it means for you |
|---|---|---|
| Retrieval (Andromeda) | Tens of millions of eligible ads cut to a few thousand per person | If your ad is not pulled here, it never reaches the auction |
| Ranking | Heavier models predict actions for the shortlisted ads; GEM trains and informs these models | Predicted conversion rate and ad quality decide order |
| Auction | Bid, estimated action rate and quality combine into a total value | Budget and bid only matter for ads that survived the first two steps |
Meta's March 2025 business update on ads ranking frames it plainly: as businesses upload more creatives "to support a diversification strategy," Andromeda helps pick the right creative for each person. That is the shift. Retrieval used to lean on the audience you defined. Now it leans on matching people to creatives.
Why the creative does the targeting now
Meta has not published the full feature list for Andromeda. It has said what its ranking foundation model uses: the GEM post lists ad attributes including format and "creative representation" alongside each person's activity history. In practice, the system learns which people respond to which kinds of ads, and the ad itself becomes the strongest description of who it is for.
Two consequences follow.
- Near-duplicate ads compete with each other. If ten ads look and sound alike, the system has little reason to treat them as ten options. The industry calls this creative similarity, and grouped ads are often described as sharing an Entity ID. Neither is an official Meta metric, but the behavior they describe is what advertisers report.
- Different ads reach different people. An unaware story ad and an offer-first ad for the same product can be retrieved for completely different people. That is your reach, coming from the creative instead of the ad set.
What 2,443 winners look like under Andromeda
Our October 9, 2026 snapshot covers 2,443 live Meta ads from 749 brands, each with at least $50k in estimated spend. All of them have been delivering in the post-Andromeda system. We looked at how they spread across the dimensions the system can tell apart.
34%
of winners written for unaware readers
5
awareness levels all represented, none above 34%
57%
of brands with 5+ winners run both video and image
55%
median share of a brand's winners on its single top angle
Across the board, no single approach dominates. Unaware ads are the largest group (825 ads), but problem-aware (491) and solution-aware (496) are each about a fifth of the board. Root-cause is the most common narrative angle at 34%, and seven other angles share the rest.
The brand-level view is more useful. 125 brands have five or more winners live at once (1,458 ads, 60% of the board). Here is how they spread their winners:
| Brand spread (5+ winners live) | Brands | Ads | Share reaching top tier |
|---|---|---|---|
| One awareness level only | 25 | 216 | 7.4% |
| Two or more awareness levels | 100 | 1,242 | 11.7% |
| One format (video or image only) | 54 | 552 | 9.6% |
| Both video and image | 71 | 906 | 11.9% |
Top tier means the top 10% of the board by winner score, so about 10% is the baseline. The gaps are a few points, not a cliff, and the groups are small. Read it as a direction: brands whose winners speak to more than one stage of awareness get a slightly larger share into the top tier.
The typical brand in that group is spread, but not evenly. Its most common angle covers 55% of its winners, and its most common hook pattern about 62%. That looks like one core message plus several real alternatives, which is a sensible structure for a system that retrieves by creative.
Three brands, three patterns
The Farmer's Dog has 27 winners live, across five angles, five awareness levels and five hook patterns. Six of the 27 are in the top tier, 22% against a board baseline of about 10%. Two of its ads show how far apart one brand's concepts can sit.
Meet Beverly.
- Est. spend
- $5.4M
- Days live
- 89
- Format
- Video, 49s
A character-led opener aimed at people who are not thinking about dog food at all. Nothing in the first line names the category.
A few days ago, we celebrated his 19th birthday.
- Est. spend
- $1.2M
- Days live
- 152
- Format
- Video, 46s
Same product, different person on screen, different reader. It goes to a dedicated landing page instead of the product page, so the whole path differs too.
Earth Breeze runs 23 winners with seven different hook patterns, the widest spread of any brand with 20+ ads on the board. Five of the 23 are in the top tier. Its demo ads, founder reads, interview clips and offer statics would be hard for any system to confuse with one another.
Give me one minute and you'll never buy a laundry jug like this again.
- Est. spend
- $841K
- Days live
- 113
- Format
- Video, 94s
- Variants
- 2
One of three Earth Breeze openers built on "give me a minute," but each one attacks a different product (jugs, liquid detergent, dish pods), so the concepts stay distinct.
Stepprs is the opposite case. It has 29 winners, all video, with one tagged angle, one awareness level and one hook pattern, and the same headline on 28 of the 29 ads. Several scripts open with nearly the same sentence. The ads are long-lived (197 to 487 days), so they clearly sell, but none of the 29 reaches the top tier. That is what heavy similarity tends to look like on our board: steady, but each new ad adds little reach.
What to do with this
- Audit your live ads by concept, not by count. Write down the angle, awareness level, opener type and format for each active ad. If most rows look the same, you have fewer options in retrieval than your ad count suggests.
- Cover at least two awareness levels. One ad for people who have never thought about the problem, one for people comparing solutions. Our awareness levels guide has openers for each.
- Run both formats. 57% of multi-winner brands on our board do. A static offer ad and a 60-second story are read as different things.
- Change the idea when you want reach, change the hook when you want lift. A new opener on a proven body is a good refresh. It is not a new concept. Our creative diversification playbook shows how to plan both.
- Study spread, not single ads. When you research a competitor, look at the whole set of ads it keeps live. On Ad Radar, every winner carries the same ten tags, so you can see a brand's spread across angles and awareness levels in one view on the hook breakdown.
The ads that last are not the cleverest single ads. They are sets of different ideas, each giving the retrieval system a reason to show it to someone new. That is the practical meaning of Andromeda, and it matches how long winners survive on the board: the median ad has been live 120 days.
Figures marked as estimated spend come from Ad Radar's model of engagement on public Meta Ad Library ads. They are estimates, labeled as such, and are best used to rank ads against each other.