Andromeda retrieves relevant ads; later models rank candidates for delivery. GEM improves recommendation models through shared learning. For advertisers, the practical work is to supply distinct, persuasive creative, a strong offer and reliable conversion signals.
01 / How an ad reaches a person
Meta describes retrieval as the first stage of a multi-stage recommendation system. Andromeda narrows tens of millions of eligible ads to a few thousand relevant candidates. More sophisticated models then predict value and relevance in later stages.

- Eligible ads provide the candidate pool. Your creative is one input alongside campaign settings and other signals.
- Andromeda retrieves candidates relevant to the person and context.
- Ranking and auction stages evaluate the candidates before delivery.
- Interactions and conversion signals inform subsequent learning. GEM supports recommendation models through knowledge transfer; it is not a separate button or a mandatory next step after Andromeda.
02 / What the published research actually says
Creative matters, but the official papers do not say that targeting, bids, measurement or the offer no longer matter. They also do not establish a universal “creative contributes 70%” rule or a fixed number of videos every account must launch.
| Published work | What it explains | What to do with it |
|---|---|---|
| Andromeda · Dec 2024 | More personalized, scalable ad retrieval as the creative pool grows. | Develop distinct messages and demonstrations for different customer needs. |
| GEM · Nov 2025 | A foundation model transfers learning to other ad recommendation models. | Keep ad content, the offer and conversion measurement consistent. |
| GEM training · Aug 2026 | Learning from ad content and engagement at much greater training scale. | Treat this as infrastructure research, not a new universal budget or targeting rule. |
03 / Test different reasons to buy
Creative diversity is useful when it tests a different argument, context or demonstration. A new caption color can be a variation, but it does not by itself test a new reason to buy. The following is a RemakeAds testing example, not a Meta requirement.
| One product | A different creative hypothesis |
|---|---|
| Problem → demonstration | A close-up of a brush reaching a difficult corner makes the benefit immediately visible. |
| Routine → convenience | Show how the product fits into an ordinary cleaning routine. |
| Question → proof | Answer a real product objection with an accurate demonstration. |
References[5]
04 / A practical creative-testing workflow
- Define the business result. For ecommerce, establish the purchase event, target acquisition cost and margin before judging creative.
- Check measurement. Validate Pixel events and, where used, Conversions API deduplication. Keep currency, purchase value and reporting windows consistent.
- Choose a small set of distinct creative hypotheses that your budget can actually test. Keep a clear naming system for each angle.
- Use Advantage+ automation where it fits your objective and available account controls. Keep required market, location and business constraints explicit.
- Avoid changing the offer, landing page, audience and creative all at once. Use Meta’s experiment tools when you need a controlled comparison.
- Review after enough delivery and conversion lag for your business. Use purchases, CPA and attributable revenue to judge outcomes, then use viewing and click metrics to diagnose why.
References[5]
05 / Read the whole journey before blaming the video
Use these as diagnostic questions, not automatic pause or scale commands. A single metric rarely identifies the cause on its own.
| What you observe | What to investigate next |
|---|---|
| Views, but few clicks | Product clarity, purchase motivation and the call to action. |
| Clicks, but few purchases | Landing-page speed, product/offer match, shipping, checkout and event recording. |
| Purchases, but weak profitability | Acquisition cost, gross margin, returns and the attribution window. |
| A former winner is declining | Delivery mix, competition, seasonality and creative fatigue before choosing the next test. |
Find a reference in Discover. Select a product if needed. Click Remake.
Questions, answered.
Is Andromeda the same thing as Advantage+?
No. Andromeda is part of the recommendation infrastructure. Advantage+ is a suite of advertiser-facing automation products. You configure campaigns; you do not switch Andromeda on as a separate campaign setting.
Does a successful TikTok video automatically work on Meta?
No. The purchase argument may transfer, but the audience context, placements, offer and destination differ. Adapt the creative and evaluate it with Meta campaign data.
Sources
Official facts, vendor descriptions and community tools are identified separately. Practical testing examples are RemakeAds editorial recommendations.
- Meta Andromeda: personalized ads retrieval engine · 2 Dec 2024Meta · official documentation
- Meta’s Generative Ads Model (GEM) · 10 Nov 2025Meta · official documentation
- Training GEM at LLM scale · 3 Aug 2026Meta · official documentation
- AI innovations in Meta’s ads ranking · 27 Mar 2025Meta · official documentation
- Meta Advantage+ explained in two minutes · 22 Apr 2025Meta · official documentation