Every media buyer running Meta’s ads algorithm right now is watching the same pattern play out. CPMs keep climbing, conversion rates from Meta traffic are sitting near record lows according to Northbeam’s latest data, and performance inside individual accounts swings harder week to week than it used to. Marianne, Senior Paid Performance Specialist at Smart Marketer, runs some of our agency’s highest-spending client accounts, and her read on the volatility isn’t what most media buyers want to hear. People keep blaming the algorithm. The real issue is how tightly they’re still trying to control it.
Key Takeaways
- CPMs are climbing and conversion rates from Meta traffic are near record lows, but the accounts still scaling share one trait: media buyers who let Meta’s AI systems explore instead of shutting campaigns down at the first dip.
- Meta’s decision-making now happens at the platform level, across systems like Andromeda, Lattice, and GEM, rather than inside individual ad accounts, so heavy manual intervention interrupts learning instead of improving it.
- A recurring giveaway funnel optimized for purchases, not leads, is one of the most consistent offer structures working across ecommerce accounts right now.
- Building 10 to 20 genuinely different creative concepts before a launch, instead of iterating on one winning ad, is what keeps campaigns resilient heading into Q4.
Why are Meta ad accounts performing so inconsistently right now?
Meta’s AI systems, including Andromeda, Lattice, GEM, and sequence learning, are running way more experimentation on ad accounts than they used to. That experimentation is what produces the swings that feel like instability. An account that looks strong on Monday and weak by Friday is usually mid-experiment, not actually broken.
Marianne’s read on this is that the industry handed a lot of decision-making power to AI faster than most media buyers got comfortable with. The instinct when performance dips is to intervene: turn off an adset, rewrite a headline, kill the campaign. That instinct made sense under the old rules, back when a human was making most of the targeting decisions by hand. Under the current system, that same intervention interrupts a learning process that was already headed somewhere good.
Where does Meta’s AI actually make its decisions?
Meta’s learning happens at the platform level, not inside a single ad account, adset, or pixel. Once a creative goes live, Meta evaluates it as part of the broader ecosystem rather than in isolation, which is why micromanaging at the account level carries a lot less weight than it used to.
That changes what control actually means for a media buyer. You still decide how much budget an asset gets and which campaign structure it launches in, but the fine-grained optimization that used to live in manual bid adjustments and audience exclusions now happens upstream, across the platform. Marianne’s own preference is one adset per campaign with multiple ads cultivated inside it, though she’s quick to point out there’s no universally correct structure. The right structure is whichever one you understand well enough to run consistently.
What actually separates winning Meta ad accounts from struggling ones?
Creative and offer strength matter more than targeting precision right now, because the levers that used to compensate for a weak offer or flat creative, like narrow interest targeting and audience exclusions, carry a lot less weight in an AI-driven system. A weak offer has nowhere left to hide.
Accounts that used to survive on smart audience segmentation alone are the ones struggling most in this environment. The businesses still scaling are the ones treating creative strategy and offer design as the main lever, not a backup plan. This is also where Marianne spends most of her own time: digging into Reddit threads and customer reviews for language and objections, then building creative concepts from what she actually finds there instead of guessing.
What offer structures are working best on Meta right now?
A recurring giveaway funnel is one of the most consistently effective offer types across ecommerce accounts. It runs as an always-on funnel with a monthly winner, built around a high-value bundle of the brand’s own products instead of a generic prize.
The mechanics matter here. Someone who enters the giveaway starts imagining themselves owning the product before they’ve paid for it, which creates real ownership regardless of whether they win. The entry page leads to a thank-you page with a genuine purchase offer, not just a discount code for signing up. The messaging needs to directly address the objection of waiting to see if they win, reassuring entrants that a purchase now gets refunded if they end up winning the giveaway. The campaign gets optimized for purchase events, not lead volume, and it works across both cold and retargeted audiences without needing heavy exclusions.
How should you structure catalog ads for consistent performance?
There are two catalog ad, or DPA, structures worth testing, and most accounts respond well to one but not both at the same time. Instead of running them side by side as a traditional split test, pick one, give it real time to perform, and only move to the other if it underperforms.
| Approach | How It’s Structured | Best Fit |
| Evergreen single DPA | One campaign, one adset, one catalog ad left running continuously. Creative gets refreshed with seasonal frames instead of being replaced. | Accounts that want a stable, low-maintenance baseline that never fully turns off. |
| Refreshed DPA campaign | One campaign holding a catalog ad alongside a rotating set of regular ads, refreshed at least weekly. | Accounts with a steady stream of new creative that benefit from constant testing next to the catalog. |
For accounts that respond well to AI creative enhancements, particularly background generation on catalog ads, letting Meta’s automated enhancements run tends to outperform a rigid, manually controlled setup.
How should you approach onboarding or auditing a new Meta ad account?
Start with three to four products at different price points to create multiple entry points into the brand, then review up to a year of ad history before building anything new. Plenty of accounts have previously successful ads that were simply forgotten, not tested to exhaustion.
Older creative that performed well and quietly got shelved is often more useful than anything currently running. From there, build each campaign around a single adset and feed proven ads into it rather than spreading tests thin across multiple adsets. Where the account allows it, segment creative by customer avatar so Meta gets a clearer signal to work with.
The instinct to keep iterating on one winning ad, sometimes called a unicorn ad, is understandable but limiting. A stronger move is building additional ads designed to convert the people who saw the unicorn ad and didn’t buy on the first pass. Most accounts need patience here. A new campaign structure doesn’t perform on day one. It needs to be fed consistently before it starts producing a reliable signal.
How should brands prepare their Meta ads strategy for Q4?
Performance on Meta typically picks up after summer, and the accounts that handle Q4 well are the ones that build out their creative library and learn the algorithm’s behavior well before Black Friday hits, not during it.
That means building 10 to 20 genuinely different creative concepts before launch, not 10 variations of the same headline or hook, and using the slower summer months to study how Andromeda, Lattice, and GEM behave inside your own account while spend and stakes are still low. Offers and creative need to be ready early enough to leave room for testing before the real volume shows up. The hardest part is resisting the urge to tighten control as performance gets more volatile heading into the holidays. That instinct tends to make results worse, not better.
Understanding how the system behaves before you need it to perform under pressure is the difference between reacting to Q4 and actually being ready for it.
Where does AI actually fit into a media buyer’s day-to-day work?
AI tools like OpenAI’s ChatGPT and Anthropic’s Claude are most useful as research and strategy partners fed with an account’s full history, not as a generic source of tactics. Media buyers who consult AI without feeding it accurate, current context risk getting confident answers built on outdated assumptions about how Meta’s algorithm actually works.
Marianne keeps a dedicated project for each client loaded with that account’s full research history, then uses it to dig into audience language, run account analysis, and shape creative strategy. The same caution applies to clients who bring AI-generated suggestions to our team. A chatbot without access to an account’s real data and recent performance will often produce advice that sounds specific but is built on stale information. The agencies and media buyers getting the most value from AI right now are the ones treating it as a research accelerant layered on top of real account history and experience, not a replacement for either.
