What Andromeda Actually Changed and What Your Account Structure Should Look Like Now
Andromeda rewired what Meta optimises for. Most accounts still fight the algorithm. Here is the structural response that actually works in 2026.
The accounts still running interest-stacked campaigns with six creatives per adset are not just leaving performance on the table. They are actively working against the system that now decides whether their ads are seen at all. Meta's Andromeda retrieval engine, confirmed in the company's own December 2024 engineering post, did not tweak delivery at the margins. It replaced the fundamental unit of optimisation, shifting from individual ad performance to ad sequence performance, and most practitioners have not rebuilt their structures to match.
The numbers from Meta's engineering team are not subtle. A 10,000x increase in model complexity versus the prior system, a +6% improvement in retrieval recall, and a +8% boost in ad quality on selected segments. Underneath those figures is a change in what the algorithm is actually solving for: not "which single ad wins this impression" but "which sequence of ad exposures maximises value for this person over time". That is a different problem, and it requires a different structure to serve it well.
The good news is that the structural response is not complicated. One CBO, one or two adsets with broad or no targeting, and a deep creative stack of 15 to 50 genuinely different executions. The practitioners who made that shift early are reporting compounding gains. Those who did not are wondering why their scaling attempts keep stalling.
What Andromeda Is and Why the Engineering Blog Explanation Matters to Media Buyers
Most algorithm change coverage stops at "things will be different now". The Meta engineering post from December 2024 is worth reading properly because it explains the mechanism, and the mechanism has direct implications for what you build.
The post describes Andromeda as a retrieval and ranking system that replaced Meta's previous ad delivery stack. The key architectural change is the scale at which it operates: the system now evaluates candidate ads across a far larger pool than before, using two-tower neural retrieval models that can match ad content to user state in real time. The 3x improvement in inference throughput and 100x improvement in feature extraction latency that Meta reported are what enable this. The system can now consider what a person has already been shown, and when, as part of ranking the next impression.
That is the sequence problem made operational. The algorithm is no longer asking "what is the best ad for this person right now in isolation". It is asking "what is the best ad to show this person given what they have already seen, across the whole auction". Winning that question requires variety in your creative pool. If you have six ads, the sequencing logic has very little to work with.
For media buyers, the practical implication is that the campaign that "lost" the A/B test may have been providing the context that made the winning ad work. Looking at individual ad performance in isolation is reading the wrong signal.
The Sequence Problem: Why Individual Ad Performance Is the Wrong Read
The habit of ranking ads by ROAS or CPA and pausing the bottom half is deeply embedded in paid social practice. It was a reasonable heuristic under the old system. Under Andromeda it is counterproductive.
When you pause low-performing ads, you are removing variety from the pool that the sequence engine has available. An ad that converts poorly on first exposure may be doing real work as a reactivation touch, or as the second exposure that contextualises a harder conversion message. You cannot see that in the standard reporting view, because standard reporting attributes by last click or last impression, not by sequence contribution.
Andromeda delivered a +6% improvement in retrieval recall and a +8% boost in ad quality on selected segments, against a 10,000x increase in model complexity versus the prior system.
Meta engineering blog, December 2024
The implication for account management is uncomfortable: pausing ads based on per-ad ROAS is probably removing sequence signal from a system that needs it. The safer heuristic is to let the system run the full pool, suppress genuinely broken creative (wrong price, expired offer, clear brand violations), and evaluate the adset as a whole against incremental measurement rather than thinning the pool based on individual ad metrics.
This is not an argument for keeping bad creative indefinitely. It is an argument for a higher evidentiary bar before you pause, and for understanding that "this ad has a low reported ROAS" is not the same as "this ad is contributing nothing to the sequence".
Why Interest Targeting Now Restricts Rather Than Refines
Interest layers have been the default signal of audience sophistication in paid social for a decade. The logic was intuitive: if you are selling running shoes, layering in "marathon training" and "fitness enthusiast" interests narrows the pool to people who are more likely to want the product. That logic relied on the algorithm's audience modelling being worse than your targeting judgement. It no longer is.
Meta stated this directly in its March 2025 product update, quoted via Dataslayer's December 2025 analysis of 83 Meta advertising changes: "With AI-enabled advertising tools, the focus has shifted from niche targeting to creative diversification as the best lever to find relevant audiences." That is the platform formally signalling that interest stacking is the wrong approach.
The mechanism behind this is important. Andromeda's audience modelling is fed by behavioural and purchase signal across the entire Meta ecosystem, at a scale and recency that no interest taxonomy can match. An interest layer does not add signal. It removes reach. You are telling the algorithm to ignore a portion of the audience pool, including people who would convert but do not happen to have the relevant interest flag attached to their profile.
By Q1 2026, Meta had begun removing the standalone interest targeting narrow option from more accounts, with broad targeting becoming effectively mandatory based on practitioner-audited account observations documented in Skaleit Agency's April 2026 blog. For most advertisers this is not a choice that needs making. It is a direction of travel that has already arrived.
The correct response to "how do I reach the right people" in 2026 is not interest refinement. It is creative that signals clearly to the algorithm who the product is for. A creative that opens with "for runners training their first marathon" does more targeting work than an interest layer, with none of the reach restriction.
Consolidation in Practice: One CBO, One Adset, Many Creatives
The structural recommendation that follows from Andromeda's architecture is consolidation. Not the vague "fewer campaigns is better" advice that has circulated for years, but a specific configuration: one CBO campaign, one adset per objective or audience segment, with the full creative stack loaded into that adset.
The Skaleit Agency case study from April 2026 is instructive here. An ecommerce account running fragmented campaigns recovered from 2.0x to 3.89x ROAS, a 93% lift on roughly the same monthly spend, after consolidating to a single CBO, removing underperforming campaigns based on incremental attribution, and rebuilding creative coverage across five awareness stages. The same analysis documented a separate client lifting ROAS from 4.30x to 6.08x on a single campaign by adding 10 new creatives to an existing winning adset rather than launching a new campaign.
The second case is important because it isolates the mechanism. The spend did not change. The targeting did not change. The campaign was already working. Adding more creative variety into the same consolidated structure produced a compound improvement, consistent with a system that benefits from a wider sequence pool.
What this looks like operationally: one CBO with a daily or lifetime budget you are comfortable holding for at least four weeks. Inside it, one adset running broad or Advantage+ audience with no interest restrictions. All your active creative inside that one adset. The budget allocation and creative rotation is handled by the algorithm, not by your campaign architecture.
One CBO per objective. One adset per segment, broad audience, no interest layers. A minimum of 15 live creatives covering multiple angles and formats. Budget consolidated rather than spread across multiple campaigns. No manual rotation or bid overrides unless you have a specific justified reason.
How Many Ads Is Enough: The 15 to 50 Creative Stack Explained
The informal six-ads-per-adset recommendation that Meta maintained for years disappeared from documentation in 2025. According to Jon Loomer's analysis of 83 Meta changes, documented by Dataslayer in December 2025, top advertisers are now routinely running 15 to 50 ads per adset. That range is not arbitrary: it reflects the minimum variety needed for Andromeda's sequence engine to have meaningful options across different user states and journey stages.
Fifteen ads is a floor, not a target. At fifteen you have enough variety that the algorithm can construct different sequences for different people. At fifty you have enough variety that the algorithm can tailor sequences to specific journey stage, recency, and creative fatigue patterns. The practical constraint is usually creative production capacity, not platform limits.
What matters more than the number is the genuine variety within the stack. Fifteen near-identical variants with different copy lines does not give the algorithm fifteen distinct sequence options. It gives it one creative idea in fifteen formats. The stack needs to cover different angles, formats, emotional registers, and awareness stages to produce real sequence variety.
A useful planning framework for the creative stack:
| Awareness stage | Creative angle | Formats to include | Minimum ads |
|---|---|---|---|
| Unaware | Problem identification, category education | Short video, single image | 4 |
| Problem-aware | Solution framing, brand introduction | Video, carousel | 4 |
| Solution-aware | Product differentiation, comparison | Single image, video | 4 |
| Product-aware | Offer, proof, social validation | Carousel, testimonial video | 4 |
| Ready to convert | Direct response, urgency, guarantee | Single image, DPA | 4 |
Twenty creatives built across five awareness stages gives the sequence engine genuine variety to work with, and ensures that a user at any journey point is likely to receive contextually relevant material. This is the creative architecture that replaces interest targeting as a precision mechanism.
Iterating on Winners Without Resetting Signal: The Visual-and-Angle Rule
The learning phase concern has been the structural argument for keeping creative separate. If adding a new ad resets the learning phase, you want to limit additions. If consolidating creative into one adset means every new test disrupts all existing signal, the fragmented multi-campaign structure starts to look rational despite its cost.
Jon Loomer tested this directly in 2025, adding one new ad to a running adset of 22 ads. The learning phase did not reset, contradicting Meta's documented rules at the time. The system's handling of creative additions has changed, and the disruption that justified the old isolation strategy is no longer reliably present.
The practical implication: you can iterate into a winning adset rather than building a parallel campaign for each new test. Add new creative to the running adset. The algorithm will explore it without penalising the existing performers. If it catches signal, it will scale into the rotation naturally.
The rule for iterating on winners is to change one dimension at a time: the visual, or the angle, but not both simultaneously. If a UGC testimonial video is performing, test a different testimonial with the same creative treatment, or test the same testimonial recut with a different opening frame. Changing the visual and the messaging angle at the same time gives you no read on which change drove any difference.
This applies to copy too. If a particular proof point in the headline is working, test it with a different visual rather than building a new concept. The winner-iteration workflow inside a single consolidated adset, informed by incremental measurement rather than platform ROAS, is the operational model that Andromeda rewards.
Reading Incremental ROAS Instead of Platform-Reported ROAS
The consolidation recommendation collides directly with the most common objection: "but my platform ROAS went down when I consolidated". This is expected, and it is not the signal it appears to be.
Platform-reported ROAS measures attributed conversions against spend. Attribution in the consolidating phase tends to show a dip because you have removed campaigns that were getting credit for conversions, often conversions that would have happened anyway. The multi-touch attribution overlap disappears, the assisted attribution disappears, and the number looks worse before it looks better.
Incremental ROAS measures whether your spend is actually causing conversions that would not have occurred without it. The methodologies range from geo-based holdout tests (dark markets) to matched market experiments to Meta's own Conversion Lift tool. None of them are perfect, but any of them is more honest than comparing attributed ROAS across time periods during a structural change.
Meta reported that advertisers who had not previously used Advantage+ Creative and switched to AI-driven targeting features saw a 22% increase in ROAS, with an estimated +7% increase in conversions for businesses using its image generation tools, per the December 2024 engineering post. Those figures are from Meta, so treat them with appropriate scepticism about methodology. But they are consistent with the direction of practitioner-observed results: the system rewards consolidated, broad, creative-rich structure when measured against incremental outcomes.
The practical measurement approach during a consolidation: run an incremental test for four to six weeks before concluding anything. If you have enough volume, a geo holdout split is cleanest. If not, Meta's Conversion Lift with a holdout group gives you a directional read. Do not judge the consolidation by the platform ROAS curve in weeks one and two.
What to Expect as Andromeda Tightens Through the Rest of 2026
The trajectory from Meta's product decisions is clear. Broad targeting becoming default is not a temporary experiment. Creative quality and variety being the primary advertiser lever is not a positioning statement. The removal of interest-narrowing options from more accounts in Q1 2026 is a structural change, not a test.
The next wave of Andromeda improvements will likely push further in the same direction: more automated audience expansion, more model complexity in sequence building, and less room for manual structural intervention to outperform the algorithm's own decisions. The advertisers who build consolidated, creative-rich structures now are building in the direction of travel.
For accounts still running fragmented structures, the consolidation case is not "consolidate and ROAS goes up immediately". It is "consolidation removes the structural friction that prevents the algorithm from doing what it is now capable of doing". The accounts fighting the system with narrow targeting and thin creative stacks are not getting the worst of Meta's capabilities. They are getting a constrained version of them.
The accounts reporting compound gains, like the Skaleit case studies, are doing so because they removed the structural constraints and let Andromeda work with the full surface area it was built for. That is the practical argument for consolidation: not that the algorithm is infallible, but that it is now substantially better than legacy structural assumptions, and those assumptions have a cost.
Frequently Asked Questions
- Will consolidating to one CBO hurt performance during the transition?
- Platform-reported ROAS often dips during the first two to three weeks of consolidation as attribution overlap resolves. This is normal and does not reflect actual business performance. Run an incremental measurement test alongside the consolidation, ideally a geo holdout, and evaluate against that rather than the platform attribution curve. Most accounts stabilise and improve past baseline within four to six weeks.
- Should I still test creative in a separate campaign to avoid disrupting learners?
- The evidence from 2025 testing, including Jon Loomer's documented experiment of adding ads to a live 22-ad adset without triggering a learning phase reset, suggests that Andromeda handles creative additions without the disruption the old system imposed. Adding new creative directly to the running consolidated adset is now the recommended approach, as it preserves the existing signal while giving the algorithm new sequence options to explore.
- How do I handle multiple products or audiences if I am only running one adset?
- The one-adset rule applies within an objective and a broad audience segment. If you have genuinely different audiences, such as prospecting versus retargeting, or meaningfully different product lines with different economics, separate campaigns per objective are still appropriate. What the consolidation argument is against is the fragmentation within a single audience objective across multiple campaigns and adsets, which fractures the signal pool without adding targeting value.
- My account is small. Does Andromeda apply at low spend levels?
- Andromeda runs across Meta's full ad inventory, but the sequence-optimisation benefits become more pronounced with more data. At low volumes, the structural principles still apply: broad targeting, creative variety, consolidated budget. But you may see less dramatic impact than high-spend accounts where the sequence engine has sufficient impression volume to build meaningful sequence patterns. The floor for the structure to function well is roughly 50 conversions per week at the adset level.
Sources & Further Reading
- Meta Engineering Blog: Andromeda retrieval engine, December 2024https://engineering.fb.com/2024/12/02/production-engineering/meta-andromeda-advantage-automation-next-gen-personalized-ads-retrieval-engine/
- Skaleit Agency: Meta Andromeda complete guide for ecommerce, April 2026https://skaleit.agency/blog/meta-andromeda-complete-guide-ecommerce/
- Dataslayer: 83 Meta Ads changes that changed Facebook advertising, December 2025https://www.dataslayer.ai/blog/meta-ads-changes-2025-83-updates-that-changed-facebook-advertising-forever
- 1ClickReport: Meta Andromeda update 2025 guidehttps://www.1clickreport.com/blog/meta-andromeda-update-2025-guide
- Confect: Meta Andromeda tactics 2026https://confect.io/tactics/meta-andromeda-2026
- Affect Group: Andromeda 2026, how Meta Ads algorithms now deliver adshttps://affectgroup.com/blog/andromeda-2026-how-meta-ads-algorithms-now-deliver-our-ads/