Meta rebuilt its ads retrieval layer in December 2024, and most accounts I audit are still structured for the auction of 2019. Andromeda, the retrieval engine Meta announced then, increased model capacity at the retrieval stage by 10,000x and, by Meta’s own measurement, delivered a 6% recall improvement and an 8% ads quality improvement on selected segments.
Retrieval sounds like plumbing. It is not. Retrieval decides which handful of ads, out of millions of candidates, even enters the ranking auction for a given person. When that stage gets four orders of magnitude more capacity, the system no longer needs your ad set boundaries to figure out who should see what. It reads the creative and finds the buyer.
What Andromeda actually changed
The engineering details matter less than the design assumption. Meta built Andromeda on new accelerator hardware explicitly so retrieval could scale to a large volume of ad creatives per advertiser, evaluated per person. The intended input is many distinguishable creatives; the model does the matching. The same post noted that more than a million advertisers had used Meta’s generative AI tools to produce over 15 million ads in a month, and that advertisers using image generation saw a 7% increase in conversions.
Follow that logic to its end and creative volume becomes a targeting input. Every distinct concept you feed the system is a probe into an audience pocket you would never have built as an interest stack. Uploading one hero ad into a hyper-segmented account starves exactly the machinery Meta spent the last two years scaling.
Advantage+ swallowed the account structure
“Adoption of Advantage+ shopping campaigns continues to scale with revenues surpassing a $20 billion annual run rate and growing 70% year over year in Q4.” Susan Li, Meta CFO, January 2025.
That was the quarter automation stopped being a campaign type and became the default. In February 2025, Meta began rolling out a streamlined Advantage+ setup: no more choosing between a manual campaign and an Advantage+ shopping campaign, with sales, app and leads flows starting AI-on by default. Meta’s early testing put leads campaigns with Advantage+ turned on at a 10% lower cost per qualified lead.
The models underneath kept turning over. GEM, the ads foundation model Meta trained on thousands of GPUs, delivered a 5% increase in ad conversions on Instagram and 3% on Facebook Feed in Q2 2025. By Q1 2026, more than 8 million advertisers were using at least one of Meta’s AI creative tools, double the count from the end of 2024, and ad revenue grew 33% year over year on 19% more impressions and a 12% higher average price per ad.
Read those last three numbers together. The machine is showing more ads and charging more per ad at the same time, which only works if conversion prediction keeps improving. My read: the gains land disproportionately on advertisers who feed the system what it now runs on, which is consolidated signal and diverse creative.
How I structure accounts now
Testing changed first. I used to A/B audiences: interest stack against lookalike against broad, same creative in each cell. That test is now mostly meaningless, because the delivery system overrides your audience hypothesis within days. What I test instead is concepts against concepts in broad delivery, and I read results at the concept level, not the ad level, so a fatigued variant does not get misread as a dead idea.
I spent 2025 consolidating client accounts, and the playbook has settled:
- Fewer campaigns, pooled learning. Fragmented ad sets split conversion signal that the retrieval and ranking models would rather see in one pool. Structure now exists for budget governance and measurement, not for targeting.
- Concepts are the segmentation. I brief creative as distinct concepts against distinct buying motivations (price anchor, problem agitation, social proof, demonstration) and let delivery find the audience for each. The audience tab of my media plans got shorter; the creative brief got longer.
- Signal quality outranks structure cleverness. Server-side events through the Conversions API, deduplicated, carrying value data. A consolidated account with weak signal is just a black box with fewer buttons.
- Keep the controls that still matter: exclusions, geo, budget caps, and measurement outside the platform. Automation optimizes toward whatever you report to it, so what you report to it is the last real lever.
The one place I still add structure is measurement. Geo splits and holdout campaigns exist so I can estimate incrementality without trusting the platform’s own scorecard, and they are worth their overhead precisely because everything else got consolidated. Structure for proof, not for targeting.
A fragmented e-commerce account came to me in 2025 running dozens of ad sets, each an artisanal interest stack with its own starving budget. Consolidating into a handful of campaigns with a wider concept slate stabilized both delivery and costs. Directional, not a case study with percentages. The counterintuitive part was psychological: performance improved as my visible control decreased. What I lost was the illusion of steering. What the account gained was concentrated signal.
The trade on the table
You do not have to like this direction, and parts of it deserve suspicion. Advantage+ grades its own homework, and platform-reported lift is not incrementality. Hold that skepticism and consolidate anyway, because the alternative, fighting a 10,000x retrieval model with 2019 account architecture, loses on math.
Structure less. Brief more concepts. Guard your signal and your measurement. On Meta, that is now most of the job.