Meta's Q4 2025 earnings call contained one sentence most advertisers didn't notice: "We doubled the size of the GPU cluster we used to train GEM." That doubling produced a 3.5% lift in ad clicks on Facebook and a 1%+ gain in conversions on Instagram — inside a single quarter. GEM is why those numbers moved.
In this post:
- What GEM is and how it works at the architecture level
- How Andromeda and GEM divide the ranking work
- Why creative is now the primary targeting signal
- What the volume and testing implications are for your account
- What data quality has to do with GEM performance
What GEM Is
GEM stands for Generative Ads Recommendation Model. Meta published the technical details in November 2025 — and the headline is that it's the largest foundation model for recommendation systems in the industry, trained at the same scale as large language models.
That's not a marketing claim. Meta's engineering team describes a system trained across thousands of GPUs using hybrid sharded parallelism, with a 23x increase in effective training FLOPs compared to the prior architecture. The jump in compute isn't background engineering news. It directly determines what signals the model can learn and how precisely it can predict which ad a given user will convert on.
What makes GEM architecturally different from previous ad ranking systems is what it trains on. It doesn't just learn from ad clicks. It trains on organic user behavior across Facebook and Instagram simultaneously — reading how users scroll, engage, comment, and share across both surfaces, then using that signal pool to predict ad relevance. The model processes thousands of behavioral events per user in sequence, which lets it understand context, not just intent.
The result: GEM ranks ads by predicting the entire action sequence a user is likely to take — not just whether they'll click, but whether they'll click, convert, and return.
How Andromeda and GEM Work Together
GEM doesn't see every ad. Meta runs tens of millions of active ads at any moment, and scoring all of them for every impression would be computationally impossible. That's where Andromeda comes in.
Andromeda is Meta's ad retrieval system — the upstream filter. It narrows the full ad pool from millions down to thousands of eligible candidates using a retrieval architecture with 10,000x more capacity than its predecessor. The result is a much richer shortlist for GEM to rank.
GEM then reads each candidate ad — analyzing creative content, format, tone, copy — alongside the user's behavioral history and predicts the most relevant match. Creative quality is an explicit input to that ranking. An ad that reads as low-signal or mismatched to the user's demonstrated behavior doesn't make it through, regardless of bid.
This is the pipeline: Andromeda retrieves, GEM ranks. The creative is read at the ranking step.
Creative Is Now the Targeting Signal
The practical implication follows directly from the architecture. If GEM reads the creative to predict audience fit, then your creative defines who sees your ad — not your audience settings.
Meta's Andromeda rollout made broad targeting viable for most accounts. GEM extends that principle deeper: the system now routes ads toward the right users by analyzing what the ad communicates, not what segment the advertiser targeted. Narrow custom audiences and lookalikes don't add precision in this model — they limit the surface area GEM has to find fit.
Nielsen's marketing science research puts the creative contribution to sales ROI at 56%. In the pre-GEM world, that number reflected creative's role in persuasion after the right person had already been reached. In the GEM world, creative also does the targeting work. The 56% figure understates the leverage.
What does "high-quality creative" mean in terms GEM can read? Format signal (vertical video, native-looking content), behavioral alignment (ads that match the type of content a user already engages with), and conversion-event relevance (creative that represents what the user has shown intent to buy). Authenticity signals — real products, real contexts, real outcomes — outperform polished brand content because they more closely resemble the organic content GEM was trained on.
Volume and Testing Implications
If creative is the primary signal, the throughput of creative production becomes a direct performance variable. Accounts that test more creative surfaces per week give GEM more signal to optimize against.
Practitioners who have restructured accounts around creative volume report lower acquisition cost volatility. The mechanism is straightforward: more creative diversity means the model has more to route, which means it can find better matches for more users across more moments. Solving the creative volume problem is no longer a production question — it's an optimization strategy.
The stability window is also longer under GEM. The model needs 7+ days and 50–75 conversion events to learn optimal routing for a new ad set. Touching campaigns before that window closes interrupts the sequencing model. Patience here is a competitive advantage — most teams still optimize on shorter windows.
For teams using bulk to manage creative uploads, the production-to-live cycle shrinks enough that maintaining a larger creative pool becomes operationally viable. When uploading and launching ads takes hours instead of days, testing cadence can keep pace with what GEM needs to optimize.
Data Quality as a Compounding Moat
GEM learns from conversion events. The richer the conversion signal flowing back through your Meta pixel and Conversions API, the more precisely GEM can optimize for your actual business outcomes.
An account sending purchase and subscription events to Meta gives GEM fundamentally better training signal than one sending only page-view or add-to-cart events. The improvement compounds — each conversion trains the model to find better next conversions. Accounts with clean, deep-funnel CAPI tracking accumulate an optimization advantage that grows over time.
This is why the Advantage+ automation push from Meta isn't just product strategy — it's GEM-native. Advantage+ campaigns send Meta the full decision loop (the model proposes, you approve, it learns from outcome), which aligns perfectly with how GEM builds its targeting predictions.
The teams that outperform in this environment aren't the ones with bigger budgets or tighter audience segments. They're the ones producing more creative, tracking deeper conversion events, and letting GEM route. The model is the strategy now. Execution speed and data quality determine who wins.
bulk handles the creative upload and campaign execution layer for Meta ads teams — keeping the production loop moving so GEM has what it needs to optimize. See how it works →