A CMO I advise called last week with a familiar complaint: Meta ROAS dropped 18% quarter-over-quarter despite holding creative spend flat. Her team blamed the algorithm. The algorithm wasn't the problem. Her campaign architecture was built for a system that no longer exists.
Meta's Andromeda retrieval engine, which completed its global rollout in October 2025, represents the most significant change to how Meta serves ads in over a decade. The old system asked "who in this audience is most likely to convert?" Andromeda asks a fundamentally different question: "which ad should this specific person see right now?" That inversion breaks every campaign structure built on audience segmentation, funnel stages, and manual targeting controls.
The performance implications are measurable. Common Thread Collective's analysis of 3,014 advertisers found ROAS declined 7% on average during the rollout, with high-ticket products seeing a 31% collapse. But advertisers who restructured around Andromeda principles reported 17% more conversions at 16% lower cost. The gap between winners and losers is widening, and the variable isn't budget. It's architecture.
The Retrieval Layer Changes Everything
Before Andromeda, your targeting parameters were instructions. You defined an audience, Meta found people in that audience who might click. Targeting was the lever; creative was the execution.
Andromeda inverted this completely. The system now analyzes every element of your creative (visuals, copy, pacing, format, emotional tone) and uses those signals to find the right users. Meta's engineering team reports a 10,000x increase in model complexity for ads retrieval, with query speed improving 3x and feature extraction 100x faster.
What this means operationally: your targeting settings are now soft suggestions. The system will expand beyond your audience definitions if the creative signals suggest a better match elsewhere. As PPC Hero's recent analysis puts it, "our job is not to 'control' performance, but to design systems that generate strong, clean signals for it."
The practical consequence is that granular audience segmentation now works against you. Spreading budget across multiple campaigns and ad sets dilutes the signal Andromeda needs to detect patterns. The algorithm is searching through tens of millions of potential ad-to-user combinations; you want your data concentrated enough for it to learn clearly.
Consolidation: The New Campaign Math
The structural shift is straightforward: fewer campaigns, fewer ad sets, more budget per learning system, broader targeting inputs.
1ClickReport documented a case where restructuring one DTC client from 2 ad sets × 4 creatives to 1 ad set × 12 creatives moved ROAS from 1.9x to 2.7x in 11 days. Spend held constant. The only variable was architecture.
The old TOF/MOF/BOF funnel logic is particularly counterproductive now. Campaign structure should mirror business goals, not audience segments. Think acquisition vs. retention, lead vs. sales, geographic or operational constraints. Unify signals toward the final business outcome rather than fragmenting them across arbitrary funnel stages.
Creative volume matters more than ever. Accounts running 3 creatives per ad set consistently underdeliver compared to accounts running 8 to 15. Andromeda clusters visually similar ads under the same Entity ID, so ten ads with the same product photo and different headlines count as one ad in retrieval terms. Visual diversity isn't a nice-to-have; it's structural.
The CFO Conversation
Here's where this gets board-relevant. Meta's internal benchmarks show Advantage+ campaigns deliver an average 32% increase in ROAS and 17% lower CPA compared to manual-only campaigns. But those gains require proper setup: clean conversion tracking, sufficient creative volume, and consolidated campaign architecture.
The tracking foundation is non-negotiable. As one practitioner noted on LinkedIn, "media buying is effectively merging with data engineering. When the feedback loop is the only real optimization lever left, the integrity of your CAPI setup matters infinitely more than campaign structure." First-party data and custom audiences are now your primary targeting controls.
For B2B specifically, the implications are significant. Detailed targeting still makes sense for new accounts with fewer than 50 weekly conversions, very niche markets, hyper-local campaigns, or budgets under $30 per day where the AI doesn't have enough data to learn effectively. But once you cross those thresholds, fighting the system costs more than feeding it.
The Two-Week Pilot
If your Meta campaigns are still structured around audience segmentation and funnel stages, here's a low-risk test:
Take your highest-spend campaign. Consolidate ad sets into a single ad set with 10 to 15 visually distinct creatives. Enable Advantage+ audience with your current targeting as suggestions only. Hold budget constant. Run for 14 days.
Measure ROAS, CPA, and conversion volume against the previous 14-day period. If the consolidated structure underperforms, you've lost two weeks of learning. If it outperforms (as the data suggests it will for most accounts), you've found your new baseline.
The risk isn't running this test. The risk is continuing to optimize a campaign architecture built for a system that no longer exists while your competitors figure out the new one.
