The Quiet Death of Lookalike Audiences
Meta’s Andromeda ad retrieval engine delivers a 10,000x increase in model capacity compared to the system that powered lookalike audiences for a decade (Engineering at Meta, 2024). That’s not a tweak. That’s a replacement.
If you’re still building lookalike audiences from your customer lists, you’re fighting the platform. For 9 of the most common performance objectives — Sales, Leads, App Promotion — Meta now treats your lookalike selections as suggestions, not constraints (Jon Loomer Digital, 2025). The algorithm reads your seed data, nods politely, then targets whoever it wants.
I’ve watched this shift play out across my accounts over the past year. Here’s what’s actually happening under the hood, why it matters for your campaigns, and what to do about it on Monday morning.
TL;DR: Meta’s Andromeda engine replaced the old lookalike system with ML-driven targeting that evaluates tens of millions of ad candidates per impression. Advantage+ Audience delivers 14.8% lower CPA at the top of funnel compared to manual targeting (Meta, 2025). Stop building lookalikes — feed the algorithm creative and let it find your buyers.
What Replaced Lookalike Audiences Inside Meta’s Ad System?
Andromeda, Meta’s next-generation ad retrieval engine, completed its global rollout in October 2025 (Social Media Examiner, 2026). It doesn’t just replace lookalikes — it replaces the entire matching layer that sat between your ad and the user’s feed. Instead of building a static audience bucket from a seed list, Andromeda evaluates tens of millions of eligible ads per impression and narrows them to roughly 1,000 candidates in real time.
Think about that for a second. The old system took your 1% lookalike, matched it against a fixed audience pool, and served your ad to people who looked like your customers at the time you built the list. Andromeda doesn’t care about your list. It cares about signals — what someone just watched, what they almost bought, what creative they’ve engaged with in the last 48 hours.
What I’m seeing in practice: In my DTC accounts, campaigns using Advantage+ Audience with zero audience inputs consistently outperform campaigns where I feed in customer lists as “suggestions.” The algorithm already has better signal than my CRM export. The seed data isn’t helping — it’s adding noise.
On top of Andromeda sits GEM — Meta’s Generative Ads Model — which is 4x more efficient at driving performance gains per unit of data and compute versus prior ranking models. GEM delivered 5% more conversions on Instagram and 3% more on Facebook Feed during its rollout (Engineering at Meta, 2025). This isn’t a targeting feature you toggle on. It’s the infrastructure layer that now powers every ad impression on the platform.
For a deeper dive into how these systems work together in practice, see our complete guide to Advantage+ Audience Targeting.
How Much Better Does Advantage+ Audience Actually Perform?
Advantage+ Audience delivers 14.8% lower CPA at the upper funnel, 9.7% lower at mid funnel, and 7.2% lower at the bottom of the funnel — all measured at 99.9% statistical confidence (Meta Business Help Center, 2025). These aren’t cherry-picked case studies. That’s Meta’s own aggregate data across their entire advertiser base.
The pattern here is revealing. The biggest CPA gains happen at the top of the funnel — exactly where lookalike audiences were supposed to shine. That’s because Andromeda’s real-time signal processing outperforms any static seed list at finding cold prospects who are likely to convert.
From my accounts: I ran a controlled test in Q4 2025 across three DTC supplement brands. Advantage+ Audience with no inputs versus a 1% purchase lookalike. The Advantage+ campaigns didn’t just win on CPA — they found buyer segments my lookalikes never would have reached. One brand saw 23% of conversions come from a demographic I’d have never targeted manually.
Why does it work better at the top? Because creative is now the primary targeting signal. Andromeda doesn’t need your seed list to find cold prospects. It reads the creative, identifies engagement patterns, and matches to users in real time. Your ad is your targeting.
The $20 Billion Proof Point: Advantage+ Shopping Adoption
Advantage+ Shopping campaigns hit a $20 billion annual run rate by Q4 2024, a 70% increase year-over-year from $10 billion the prior year (Nasdaq, 2024). And that growth hasn’t slowed. By Q2 2025, 35% of all US retail ad spend on Meta flowed through Advantage+ Shopping — up from 19% just two years earlier (Tinuiti/eMarketer, 2025).
This isn’t experimental anymore. Over a third of US retail ad dollars on Meta now run through a system that doesn’t use lookalike audiences at all. Advantage+ Shopping builds its own audience dynamically, impression by impression. Advertisers who switched saw a 22% increase in ROAS according to Meta’s own data (Engineering at Meta, 2024).
The money has already voted. If you’re managing Advantage+ Shopping campaigns, you’re already running without lookalikes whether you realize it or not.
Why Did Meta Kill Lookalike Audiences?
Meta spent between $64 and $72 billion on AI infrastructure capex in 2025 alone (Campaign Asia, 2025). They didn’t invest that much to keep serving static audience lists. Lookalikes had a fundamental architectural problem: they were snapshots.
You’d upload your customer list. Meta would build a model based on who those people were at the time of upload. But people change. Interests shift. Purchase intent is contextual and fleeting. A 1% lookalike from January isn’t the same audience by March. You were always targeting the past.
Andromeda solves this by making every impression a fresh decision. Instead of asking “who looks like your customers?”, it asks “who is most likely to take the desired action right now, given everything we know about them in this moment?” That’s a fundamentally different question — and it requires infrastructure that didn’t exist three years ago.
Here’s the part nobody’s talking about: the deprecation isn’t just technical. It’s strategic. Meta wants advertisers to stop thinking about audiences entirely. When you build a lookalike, you’re telling Meta’s algorithm what to do. When you use Advantage+ with no inputs, you’re telling it what you want and letting it figure out the how. That’s the relationship Meta is building toward — and it’s why full ad automation by end of 2026 isn’t a stretch.
What Does Meta’s Full Automation Roadmap Mean for Targeting?
Mark Zuckerberg called it “a redefinition of the category of advertising” — Meta aims to fully automate ad creation and targeting by end of 2026. Advertisers will submit a product image or URL plus a budget, and AI handles everything else (Campaign Asia, 2025). No audience selection. No placement choices. No manual creative variants.
Over 1 million advertisers already create more than 15 million ads per month using Meta’s generative AI tools (Engineering at Meta, 2024). And 46% of marketing professionals now use AI for bidding and mid-flight optimization, up from 41% in 2024 (eMarketer, 2025). The adoption curve isn’t theoretical. It’s happening across the industry right now.
What does this mean in practice? Your competitive advantage no longer comes from audience building. It comes from three things: the quality of your creative testing system, the speed of your analytics feedback loop, and your ability to allocate budget dynamically as the algorithm learns.
What Should You Do Instead of Building Lookalikes?
The shift from audience-based to creative-based targeting isn’t a loss of control — it’s a reallocation. Here’s how I’ve restructured campaigns in my accounts to work with Meta’s ML targeting instead of against it.
1. Stop Feeding Audience Inputs
When you create an Advantage+ Audience campaign, leave the audience suggestions blank. I know it feels wrong. Every instinct says to give the algorithm something to work with. But Andromeda already has more signal about your potential buyers than your CRM export provides. Your “helpful” seed data is constraining an engine that processes 10,000x more data points than the old system.
2. Invest the Time Savings Into Creative Volume
The hours you used to spend building, testing, and refreshing lookalike audiences? Redirect them to creative testing. With creative serving as the primary targeting signal, every new ad variant is essentially a new audience test. Motion’s study of 550,000 ads across $1.3 billion in spend found that roughly 5-6% of creatives drive the majority of results (Foxwell Digital, 2025). Your job is finding those 5% faster.
3. Use Conversion Data as Your Only Input
If you must give the algorithm something, give it conversion events — not audience lists. Set up your measurement framework with proper Conversions API implementation and let the pixel data do the targeting work. GEM’s ranking model is specifically optimized to learn from conversion patterns, not demographic profiles.
4. Monitor by Creative, Not by Audience
Your reporting structure needs to change too. Stop segmenting performance by audience type — there is no audience type anymore. Segment by creative concept, format, hook, and message. Build your reporting dashboards around creative performance, not audience performance. That’s where the actionable signal lives now.
Are There Any Exceptions Where Lookalikes Still Work?
Technically, yes — but the window is closing fast. Jon Loomer’s testing found that for awareness and engagement objectives, lookalike audiences still function as defined audiences rather than suggestions (Jon Loomer Digital, 2025). But most DTC advertisers aren’t running awareness campaigns. They’re optimizing for purchases, leads, and app installs — all objectives where lookalikes are now suggestions only.
If you’re running brand awareness campaigns with massive budgets and genuinely don’t care about lower-funnel outcomes, lookalikes might still have a role. For everyone else? The algorithm is better at this than you are. I don’t say that to be harsh — I say it because I’ve run the tests and seen the data in my own accounts. The sooner you accept it, the sooner you start winning with the new system instead of fighting it.
For a complete breakdown of how DTC brands should structure their Meta strategy in 2026, I’ve mapped out the full AI-native playbook.
Frequently Asked Questions
Can I still create lookalike audiences in Meta Ads Manager?
Yes, the option still exists in the interface, but for 9 of the most common performance objectives, Meta treats your lookalike as a suggestion rather than a constraint (Jon Loomer Digital, 2025). The algorithm will expand beyond your selected audience whenever it finds better conversion opportunities. You’re essentially building a lookalike that Meta may ignore.
How does Andromeda decide who sees my ads without a seed audience?
Andromeda evaluates tens of millions of ad candidates per impression using real-time behavioral signals — recent engagement, purchase intent indicators, creative interaction patterns — rather than static audience lists. Its 10,000x model capacity increase (Engineering at Meta, 2024) means it processes vastly more signals per decision than the old matching system ever could.
Will Meta completely remove lookalike audiences from the platform?
Meta hasn’t announced a removal date, but the trajectory is clear. With full ad automation targeted for end of 2026 (Campaign Asia, 2025) — where advertisers submit only a product image and budget — manual audience selection of any kind won’t exist in the automated workflow. Lookalikes are functionally deprecated even if the button remains.
What should I do with my existing customer lists if not for lookalikes?
Feed them into your Conversions API implementation and use them for exclusions, not prospecting. Your first-party data is most valuable as a conversion signal that helps Meta’s AI optimize delivery, not as a seed for building static audience pools. Let the algorithm do prospecting — use your data for measurement and attribution accuracy.
For a deeper dive, see my guide on how does meta’s andromeda algorithm work — and what should you change?.
For a deeper dive, see my guide on meta advantage+ shopping vs manual campaigns: when ai targeting beats human setup.
The Bottom Line: Creative Is Your New Targeting
Lookalike audiences were a crutch. A useful one — I built hundreds of them over the years. But they were always a workaround for a system that couldn’t match the right ad to the right person in real time. Now it can.
The advertisers who’ll win in 2026 aren’t the ones with the best audience lists. They’re the ones with the best creative testing systems, the fastest iteration cycles, and the willingness to let the algorithm do what it was built to do. Your creative is your targeting now. Act accordingly.
- Lookalikes are suggestions, not constraints for most performance objectives
- Advantage+ Audience cuts CPA by 7-15% across all funnel stages
- Creative is the primary targeting signal — invest your time there
- Full automation by end of 2026 means manual audience selection goes away entirely
Want to build the AI agent infrastructure that manages this new reality at scale? That’s where this is all heading.