Yes, Meta Sequences Your Ads. No, That Doesn't Mean You Should Stop Optimizing.

Every few months a screenshot from a Meta engineering paper makes the rounds on Twitter and LinkedIn. The takeaway, almost always, is the same.

"See? The algorithm knows. Stop killing ads with bad ROAS. They're doing more than you think."

I've now had this conversation with founders spending $40k a month and CMOs running nine-figure brands. The technology they're pointing at is real. The conclusion they're drawing from it is wrong.

The Sequencing Story Most Operators Have Half-Right

Meta has been public about three pieces of infrastructure that genuinely changed how the auction works.

Andromeda is the retrieval layer. It groups creatives that look semantically similar into clusters, and treats those clusters, not individual ad IDs, as the unit of competition. GEM is the generative ranking model that predicts which next-best ad to show a given user, based on the sequence of ads they've already seen. Sequence Learning preserves the order and timing of every interaction a user has with an account.

The headline numbers Meta has published are not nothing. Single-digit conversion lifts on Instagram and Facebook from sequencing alone. At Impremis, where we manage $250M+ a year in ad spend, even a two-percent shift in conversion rate translates into millions in margin. So no, I'm not here to argue the technology doesn't work.

The technology works exactly as advertised. The interpretation has gone sideways.

The interpretation goes like this. "Some of your ads with bad ROAS are doing the heavy lifting upstream. The algorithm is using them to warm cold users. If you cut them, you'll break the funnel." Therefore, do less. Trust more. Stop pruning.

What "Sequencing" Actually Means in Practice

When the model decides to show an ad, it's predicting which creative, in this user's specific path, will produce the best outcome for Meta's optimization target. That target is a conversion event you defined, weighted by Meta's revenue-per-impression goals.

That is not the same thing as: "this ad with 0.4x ROAS is secretly responsible for the 4x ROAS ad downstream."

For that claim to hold, the underperforming ad would have to be doing measurable work that simply isn't showing up in any attribution window. No 7-day click. No 1-day view. Nothing.

Think about what that requires. An ad important enough to influence the entire account's performance, but not important enough to generate a single click across any reporting window the platform offers. Those two things don't sit comfortably together.

If an ad's impact is invisible to every signal Meta itself captures, the more parsimonious explanation is that it isn't doing what people are claiming it does.

The Three Things the Algorithm Cannot Do for You

I've watched this play out across hundreds of accounts. Here's where the "trust the algorithm" worldview breaks.

The optimization target isn't your P&L

Meta optimizes for events. You optimize for first-time customer margin. Those are not the same thing. We've covered this in detail in the first-time customer P&L vs blended MER, but the short version is that the auction will happily spend you to death on returning customers and call it a win.

The algorithm cannot create what doesn't exist

GEM sequences from the creative library you've given it. If your library is exhausted, sequencing turns into rotating between fatigued assets. The fix isn't more trust. The fix is feeding the system new creative on a known cadence, which is the entire premise of scaling through systems, not winners.

The algorithm cannot tell a $40 LTV customer from a $400 one

It sees an event. It does not see that this cohort never reorders, while that cohort subscribes for 14 months. Without your hand on the wheel, it will optimize toward the cheaper conversion every time, which is rarely the more valuable one.

The Real Spectrum Is Touch, Not Trust

The right question is not "should I let the algorithm cook?" The right question is how much human oversight does this account need to produce the outcome I actually care about?

That depends on a handful of things.

| Factor | Lighter touch | Heavier touch | |---|---|---| | Spend volume | $10-30k / mo | $200k+ / mo | | Product complexity | Single SKU, single offer | Multi-SKU, multi-LTV cohorts | | Audience clarity | Mass appeal, broad demo | Niche or segment-driven | | Reporting gap (MTA vs. platform) | Small | Large | | Margin sensitivity | Strong margins, room to absorb error | Thin margins, every dollar matters |

A founder running $15k a month on a single hero SKU should be in the platform once a week, killing obvious losers, launching new concepts, and trusting Meta to do the rest. A team running $400k a day across multiple LTV bands should be looking at first-time customer economics every morning and intervening with intent.

What I Actually Do at $250M+ in Annual Spend

The portfolio I'm responsible for at Impremis isn't homogeneous. Some accounts spend $60k a month. Some spend $60k a day. Across all of them, the same three checks run weekly.

1. Is new creative getting fair test data? Anything launched in the last 14 days needs at least 3-5x its target CPA in spend before I make a judgment.
2. Are clear losers being cut decisively? A clear loser is one with statistically meaningful spend, no top-funnel signal, and no path to fix. Not a gut feeling.
3. Are scalers actually being scaled? It's amazing how often the answer is no. A great ad in a $10/day budget cap might as well not exist.

That's the rhythm. That's not a contradiction of sequencing. It's the work that has to happen on top of sequencing for the sequencing to matter.

A Production Rule I'd Defend Anywhere

For every $10k of monthly spend, you should be producing roughly one new concept per month, with three meaningful variations of each. That's the floor.

A $50k/mo brand should be launching ~5 concepts and ~15 variations every 30 days. A $500k/mo brand needs roughly 10x that. AdFuse, the ad ops platform we built, exists in part because the calendaring and tagging behind that volume is more than a spreadsheet can handle once you cross a few hundred ads a month.

This is also the only way to stay out of the creative graveyard problem. When you optimize aggressively on last-click ROAS without a steady inflow of new awareness-level creative, you slowly contract your addressable audience. Every cycle, the algorithm gets better at finding the cheapest converters from a narrower and narrower pool. The account looks healthy on the dashboard until growth flatlines and you cannot figure out why.

The answer is almost always that the top of the funnel has been quietly starved.

What I Would Push Back on Hard

A few claims I see almost every week that don't survive contact with reality.

• "Don't kill any ads. Let Meta sort it out." No. Meta sorts the ads it has. You decide what it has.
• "Bad ROAS ads are first-touch winners." Maybe. Show me the view-through data. Show me the brand search lift. If both are flat, you're rationalizing.
• "Andromeda makes creative testing obsolete." The opposite. Because creatives are now competing as clusters, the diversity of what you produce matters more, not less.
• "Just trust the algorithm." The algorithm is one of three or four parties at the table. It is not the operator.

The Operator's Take

Meta's sequencing is real. It is also one component in a system you are responsible for. The algorithm decides which of your ads to show next. You decide what ads exist, what they cost you, what audience they speak to, and what business outcome you're optimizing toward.

The brands I see lose are not the ones who pruned too aggressively. They're the ones who confused passivity with patience. They stopped producing. They stopped killing. They stopped paying attention to the gap between platform reporting and their actual P&L. And by the time the dashboard told them there was a problem, the addressable audience had already collapsed.

Trust is not a strategy. Structure is. The algorithm is doing its job. Yours hasn't changed.

FAQ

Should I really kill ads with bad ROAS, even though sequencing is real?

Kill the ones that have meaningful spend, no measurable signal in any attribution window, and no plausible fix. Don't kill new ads that haven't had a fair shot. Sequencing doesn't change either rule.

What's a fair amount of spend to judge an ad on?

A reasonable rule of thumb is 3-5x your target CPA. So if you're acquiring at a $40 CPA target, give an ad $120-200 of spend before you call it.

How many ads should I be launching per month?

Rough rule: one new concept per $10k of monthly spend, with three variations per concept. A $100k/mo brand should be launching ~10 concepts and ~30 variations every month.

Is broad targeting always the right move now?

For most accounts under $200k/mo, yes. For larger accounts with distinct LTV cohorts or multiple offers, you almost certainly need exclusion logic and segment-level audiences. The platform default is rarely the optimum at scale.

What does "creative graveyard" actually look like?

Flat new-customer rate, slowly rising blended ROAS, declining brand search volume, and a sense that the account is "working" until growth stalls. By the time it's obvious, you've spent six months mining a narrowing audience.

How does this change for very small accounts?

If you're under $20k/mo, the cost of over-managing is real. Light touch, broad targeting, kill obvious losers, focus most of your time on producing more creative.

Where does first-party data fit?

Your exclusion lists, customer match audiences, and CRM-side LTV signals are how you make Meta's optimizer optimize for your business, not theirs. Without them, sequencing optimizes toward the cheapest conversion. With them, it optimizes toward the conversions you actually want.

Has any of this changed how I think about ad ops at scale?

It's the reason AdFuse exists. Once you're past a few hundred ads a month across multiple brands, sequencing-aware reporting and creative tagging stop being nice-to-have and start being the difference between knowing what works and guessing.

Keep reading

Pieces I've written on related topics that pair well with this one:

• Catalog Ads Are Underrated: The Facebook Dynamic Product Ad Setup Most Brands Get Wrong — The default Facebook catalog setup is wrong for most DTC brands.
• How to Actually Get New Customers from Facebook (Not Just Old Ones) — Meta will happily sell your ads to your existing customers and call it growth. Here's how we force the platform to find net-new buyers at Impremis.
• Long-Form Ads Are Working on Meta. Volume Is Still a Trap — Why 5-minute and 14-minute ads are outperforming on Meta, and why producing 100 ads a month is the wrong response to it.
• Meta Doesn't Care About Your Margin: Take Back Control — Meta optimizes for what it can see. Your business runs on what it can't. Here's the three-lever system I use across $250M+ in spend.
• The Power Law of Ad Creative: Why 5% of Ads Do 95% of the Work — Meta isn't a normal distribution. It's a power law. Stop spreading budget evenly and start treating outliers like the only thing that matters.