If your Meta campaigns started behaving strangely sometime in late 2025 — broad targeting suddenly beating your best lookalikes, creative variance spiking, CPMs creeping up without a clear explanation — you weren't imagining it. This post breaks down what Meta's retrieval system actually does, what's solidly documented versus what's agency speculation, and what I'm doing inside real DTC accounts to adapt.
What is Project Andromeda, and when did it actually happen?
Andromeda is Meta's AI-driven ad retrieval system — the stage that narrows millions of candidate ads down to a shortlist before ranking and auction even happen. Worth correcting a framing that's spread in a lot of 2026 agency content, including an earlier version of this article: Andromeda isn't a "2026 system." Meta's own engineering team detailed it publicly in December 2024. What's true for 2026 is that advertisers kept feeling its effects more directly through 2025 as it extended across more objectives and placements — broad targeting outperforming lookalikes, creative driving more of the variance in performance — not that the system itself is new this year.
The mechanism is a real inversion worth understanding: previously, you told Meta who to show your ad to, and the creative was largely decoration. Andromeda reads your creative and behavioral signals to help decide which ads are even eligible to enter the auction in the first place. If your ad doesn't make it through retrieval, it never reaches ranking — regardless of budget or bid aggressiveness.
Meta's own December 2024 numbers for the initial launch: a roughly 10,000× increase in model capacity at the retrieval stage, and an 8% ad quality improvement on the segments it was measured against. Those are real, Meta-published figures — but they describe the original launch, not a fresh 2026 development, and the 8% figure was specifically on selected segments, not a blanket improvement across all ads.
Did Meta actually replace its ad ranking engine entirely, or is this incremental?
Neither framing is quite right, and it's worth being precise here. Andromeda is one stage — retrieval — inside what Meta describes as a multi-stage ads recommendation system; ranking and auction are still separate stages downstream of it. Calling it a full "replacement" of the ranking engine overstates what Andromeda itself does, even though the practical effect on advertisers has been large.
Meta has also built a downstream model, often referred to as GEM (Generative Ads Model, not "Generative Embeddings Model" — a mislabel that circulates in agency content), which functions as a kind of teacher model that transfers what it learns to more specific ranking models through knowledge distillation. In its Q4 2025 earnings call, Meta stated it had doubled the GPU allocation dedicated to training GEM — a real, disclosed data point, though we can't verify the more dramatic framing ("three additional AI systems," "five interlocking systems," an "eighteen-month buildup") that shows up in some agency recaps of this story; that level of detail isn't in Meta's own public statements as far as we can find.
What's genuinely changed for account structure: Meta increasingly treats your creative set as a sequence rather than independent, interchangeable options. If you pause a top-of-funnel ad because its in-platform ROAS looks weak, you may be disrupting a sequence that a lower-funnel ad depends on — you likely won't see that dependency in a 7-day attribution window. We don't have a reliable published figure for how much ROAS typically drops when advertisers make this mistake, so we're not going to quote one — but the mechanism itself is real and worth understanding before you cut a top-of-funnel ad on its face-value ROAS alone.
What does the benchmark data say about Advantage+ Shopping vs. manual campaigns?
Two different benchmark sources give two different confidence levels here, and it's worth distinguishing them rather than blending them into one number.
One source (Skale Strategy) puts Advantage+ Shopping Campaigns (ASC) at roughly 4.5x ROAS versus about 3.7x for manually configured campaigns — a lift in the 15–25% range, with a claimed ~32% lower cost per acquisition. That source discloses no sample size or methodology behind those figures, so treat them as an agency's directional claim rather than a verified benchmark.
A more transparent source (MHI Growth Engine, disclosing a sample of 1,247 active Meta ad accounts and $87M in spend over 2025) states Advantage+ Shopping now represents 62% of ecommerce Meta ad spend (up from 34% in 2024) and delivers an average 17% lower CPA than manual campaigns — for brands meeting readiness criteria: 30+ SKUs, 15+ creatives, and sufficient conversion data. Below those thresholds, the algorithm doesn't have enough material to work with. We couldn't verify a specific spend-threshold claim (a figure like "£10,000/month" circulates in some agency content) against any source we could actually locate, so we're not repeating it here — the qualitative point (more budget and data volume generally helps ASC perform better) is reasonable, but we don't have a hard number to back a specific cutoff.
A claim that broad targeting now beats detailed interest targeting by a specific margin for supplement brands shows up in one benchmark source but isn't corroborated by that same publisher's own more detailed writing on the supplements category — we're flagging that inconsistency rather than repeating the number. The directional point (broad + Advantage+ increasingly outperforms manual interest stacking for many DTC accounts) is consistent with what we see in accounts we manage, even without a precise, verified percentage attached to it.
If you're unsure whether your account structure and creative pipeline are actually set up to benefit from these shifts, that's exactly the kind of diagnostic work we do through SciGrowth's Meta Ads consulting — reviewing signal quality, campaign architecture, and creative volume before recommending any structural changes.
What are the levers that actually matter under this system?
1. Creative diversity — not volume for its own sake. Common Thread Collective's actual published creative research (a large-scale analysis of ad spend across hundreds of stores) found that only a small fraction of ads become high-spending "whales" that drive the majority of results — well under 1% of ads in their dataset. A separate, larger study (Motion, ~550,000 ads) found similarly concentrated results. The exact numbers vary between these two real studies, but the pattern is consistent: most ads don't work, a small number carry most of the spend, and your job is to produce enough genuinely varied creative — different hooks, formats, angles — that the algorithm has real options to test. Minor copy edits to the same hero image aren't creative variation. Vertical (9:16) format is broadly dominant across placements at this point, consistent with Reels/Stories taking a growing share of inventory industry-wide.
2. Simplified structure. Simpler campaign structures with broader targeting and Advantage+ placements are consistently reported as outperforming the old multi-ad-set, interest-stacked approach. In practice, I run 3–5 campaigns max per account. Every extra ad set you add tends to fragment learning signal that the algorithm would otherwise consolidate.
3. Signal quality — CAPI is table stakes. Pixel-only tracking misses a meaningful share of conversions in 2026's browser-restricted environment — estimates for exactly how much vary by source, so we're not quoting a specific percentage as settled fact, but the direction is well-documented and CAPI running alongside your pixel materially closes the gap. Run Conversions API server-side in parallel with your pixel. Check your Event Match Quality score in Events Manager — treat a low score as a real signal problem, not a vanity metric. On value-based optimization, Meta's Advantage+ Sales campaigns include a predicted lifetime value (pLTV) layer that needs a meaningful volume of high-value purchase events with value data attached to stabilize — we don't have a verified specific weekly threshold to quote, so test what your account needs rather than targeting a borrowed number.
What should you actually stop doing right now?
- Interest stacking and lookalike layers as primary targeting. Most advertisers don't have a targeting problem anymore — they have a creative specificity problem. Interest-based targeting was rarely the variable that separated high-performing accounts from low-performing ones; creative quality was. The current system makes that true at the algorithm level too.
- Placement restrictions. Restricting to "Feed only" or "Reels only" limits the algorithm's ability to find efficient delivery across the full surface area. Let Advantage+ Placements run unless you have a hard brand-safety reason not to.
- Pausing top-of-funnel ads solely because in-platform ROAS looks weak. Under a sequence-oriented delivery model, those ads may be doing structural work you won't see in a short attribution window.
- Manual bid caps on prospecting campaigns. Automated bidding generally has more room to work with real-time auction dynamics than a manual cap allows — we won't quote a precise CPA-improvement percentage for this since we couldn't verify one, but the mechanism (a cap constrains the algorithm's ability to bid opportunistically) is straightforward.
What's the honest trade-off — is there anything Advantage+ still can't do?
Yes. The recommended approach for most advertisers is running both automated and manual in parallel — use Advantage+ Creative on core performance campaigns while keeping manual control on brand and compliance work. If you have strict geographic exclusions, audience suppression requirements (e.g., excluding existing subscribers from prospecting), or brand guidelines that Advantage+ Creative's auto-enhancements would violate, manual campaigns still serve a real purpose.
Advantage+ Audience treats your inputs as suggestions rather than hard rules in most cases — the actual hard controls it respects are narrower than "everything you set," and include things like location, minimum age, language settings, audience exclusions, and Special Ad Category restrictions where applicable. Don't assume every targeting input you set is being strictly honored — check which ones are actually enforced versus treated as a signal.
The CPM environment is also tougher industry-wide. Costs have generally increased faster than performance improvements for average-performing accounts, compressing margins — while accounts with strong creative and testing velocity have tended to maintain or improve ROAS through the same period. The system rewards advertisers who give it the best inputs, which is actually good news if you're willing to invest in creative infrastructure rather than a reason to give up on the platform.
What's the one-paragraph summary for a founder who has 90 seconds?
Meta's ad delivery system increasingly uses your creative — not just your targeting settings — to decide who sees your ads. It's not a brand-new 2026 system; the core retrieval technology was detailed publicly in December 2024, and its effects became more visible to advertisers through 2025 as it extended across more of the platform. Advantage+ Shopping campaigns are reported to outperform manual setups on average, though the exact size of that gap depends on which source you trust and generally requires real scale (SKU count, creative volume, conversion data) to materialize. Collapse your campaign structure, kill unnecessary interest stacking, build genuine creative variation, and don't pause top-of-funnel ads on face-value ROAS alone.
If you want a second set of eyes on your account structure, signal setup, and creative pipeline before you restructure, SciGrowth's Meta Ads consulting is built for exactly this kind of audit — hands-on, practitioner-led, specific to your numbers. No decks, no generic recommendations.
Frequently Asked Questions
- Is Project Andromeda the same thing as Advantage+ Shopping Campaigns?
- No. Andromeda is the underlying ad retrieval system that runs across all Meta campaign types. Advantage+ Shopping is a campaign type that leverages it most fully. Andromeda affects every campaign you run — manual or automated. ASC is the campaign structure that gives the algorithm the most latitude to use what retrieval surfaces.
- Do I need to migrate everything to Advantage+ Shopping immediately?
- Not necessarily. Start by migrating your highest-spend conversion campaigns first and run a parallel test for several weeks — same creative, same budget — before making structural decisions. ASC isn't the right fit for every brand or objective, but the reported performance gap is consistent enough across the sources we could check to take seriously. Brands with strict audience exclusion or geo requirements may still need manual ad sets for specific use cases.
- My Event Match Quality score is low — how much does that actually cost me?
- A low EMQ score degrades the retrieval system's ability to match your ads to high-intent users, because it's working from incomplete or unreliable signal about who actually converted. Combined with pixel-only tracking (which misses a real, if imprecisely quantified, share of conversions in the current browser environment), the algorithm ends up optimizing toward a meaningfully incomplete picture of your actual buyers. Implement CAPI server-side and push server events that match on email, phone, and click ID to close the gap.
- Why does creative fatigue seem to be accelerating?
- Meta's retrieval system evaluates and cycles through creative variants far faster than older ranking approaches did, which tends to exhaust a static creative set more quickly — it reaches the highest-probability audience segments sooner. We're not going to quote a specific comparison to other AI systems here since we couldn't verify one; the practical takeaway is the same regardless: you need continuous creative input, not quarterly refreshes, to keep giving the system new patterns to work with.
- Should I be worried about the rising CPM environment making Meta unworkable for DTC?
- Rising average CPMs are real, but they're not evenly distributed. In its Q4 2025 earnings call, Meta stated that AI-driven ad redistribution had a revenue impact roughly four times larger than increasing ad load — directionally suggesting the system leans on relevance rather than pure inventory expansion. Advertisers with strong creative quality and signal infrastructure tend to face better effective CPMs because their ads win auctions at higher relevance scores. The gap between average and top-quartile accounts appears to be widening, which is worth tracking in your own account more than any industry-wide CPM figure.
- What budget or scale do I need before Advantage+ Shopping actually works properly?
- We don't have a verified specific spend threshold to give you — treat any precise number you see quoted for this (including in earlier versions of this article) with skepticism. What's better supported: readiness criteria around catalog size (30+ SKUs), creative volume (15+ active creatives), and having enough conversion data flowing through CAPI for the algorithm to learn from. Below that kind of scale, you can still run ASC, but the headline performance gains reported in benchmark studies were measured on accounts with meaningful data volume, so calibrate your expectations accordingly.
Sources:
Jetfuel Agency — Meta Algorithm Changes 2026: Andromeda Update Explained
GregHal — How Meta's Ads Algorithm Works in 2026: Lattice, UTIS & Andromeda
Skale Strategy — Meta Advantage+ Shopping Campaigns in 2026 (no disclosed methodology for ROAS/CPA figures)
MHI Growth Engine — Meta Ads Benchmarks for Ecommerce 2026 (1,247 accounts, $87M spend, 2025)
Andromeda launch details, GEM naming, and Q4 2025 GPU/redistribution figures: Meta's own engineering blog (December 2024) and Q4 2025 earnings call.