Most operators I speak to describe the same symptom. Nothing obvious broke. The offer is the same, the landing page is the same, the creative is the same standard it was last year. But performance drifted, costs crept, and the tactics that reliably fixed things in 2023 now do nothing. Duplicating the winning ad set does not scale it any more. Adding another lookalike does not open a new pocket of demand. Pausing the underperformer does not lift the rest.
The reason is structural. Between late 2024 and 2026 Meta rebuilt the layer that decides which ads even get considered for a user, and that change quietly rewrote the rules for how a DTC ad account should be built. This is what actually changed, why your old structure now works against you, and the sequence to fix it without torching the account in the process.
What Andromeda Actually Is
There are two stages to serving an ad. First, retrieval: out of the millions of eligible ads, which few thousand are worth considering for this specific person right now. Second, ranking: of those, which one wins the impression. Andromeda is the rebuilt retrieval stage. Lattice is the ranking stage.
Retrieval used to be relatively crude, which is why your targeting inputs carried so much weight. You told Meta who to show the ad to, and the system largely obeyed. Andromeda inverted that. It now does the heavy lifting of matching an ad to a person, and it does that matching primarily on the content of the creative itself, not on the audience label you attached to the ad set.
The consequence that matters commercially: each genuinely distinct creative is treated as its own entity in retrieval. Twenty distinct creatives give the system twenty different ways to find a buyer for you. Twenty variations of the same concept, different colour grade, different caption, same hook and same visual structure, give it roughly one.
Your creative library is no longer a set of assets you test against each other. It is the map the system uses to find your customers. A narrow library means a narrow map, and no amount of budget widens it.
Why broad targeting quietly started winning
This is also why broad beats lookalikes in most accounts now, and why that shift felt like it happened without an announcement. It did not change because broad got smarter in isolation. It changed because retrieval got good enough that narrowing the audience mostly removes options from a system that is better at finding those people than you are. When you stack a 1% lookalike on top of Andromeda, you are pruning the candidate pool before the part of the machine that is actually good at selection gets to do its job.
The Four Fixes, In Order
Order matters here. Most brands attempt these in reverse, start with the creative sprint, see no lift because the account structure is still fragmenting the signal, and conclude that the creative was the problem. Work through them in this sequence.
Fix Signal Quality First
None of the rest works on bad data. Before you touch structure or creative, confirm the Pixel and the Conversions API are both firing, that they are deduplicating properly rather than double counting, and that you have picked one meaningful conversion event to optimise toward rather than spreading optimisation across four.
The most common version of this problem in DTC accounts is a Shopify store firing purchase events from three sources with inconsistent event IDs. The platform sees inflated, noisy conversion data, optimises against it, and the account underperforms for reasons that look like a creative problem and are not. This is unglamorous work that takes a day and quietly determines whether everything downstream functions.
Consolidate The Account Structure
Move from ten or more ad sets per campaign down to two or three. One broad prospecting ad set. One retargeting ad set. Optionally one testing ad set if you genuinely have a testing cadence to feed it. That is the shape.
Fragmented structures were a rational response to a system where you had to manually carve out audiences. Now they split your conversion signal into pieces too small for any of them to exit learning cleanly, and they create audience overlap where your own ad sets bid against each other. Consolidation concentrates the signal where the system can use it.
Do this gradually. Building the consolidated campaign alongside the existing one, moving 20 to 30% of budget into it, and migrating the rest once it matches or beats the old structure on contribution margin will cost you two weeks. Doing it overnight resets learning across the whole account and typically costs you a month.
Build For Diversity, Not Volume
The headline advice circulating is to produce more creative. That is half right and expensive to get wrong. The requirement is diversity, not volume. Thirty near-identical assets cost thirty times as much to produce and buy you almost nothing in retrieval.
Practically, this means the creative brief changes. Instead of specifying a visual treatment, specify four things per concept: the pain point being addressed, the persona it speaks to, the awareness level of that person, and the format. Change any one of those and you have a genuinely distinct creative. Change none of them and you have a variant.
The working target is 20 to 30 distinct creatives per ad set across mixed formats: user generated content, studio product, testimonial, demonstration, and catalogue. Test similarity by asking whether two ads could plausibly be shown to two completely different people for two completely different reasons. If not, they are the same ad wearing different clothes.
Split Production Between Drivers And Explorers
This is the budgeting discipline that keeps a diverse library sustainable rather than a one-off sprint you never repeat. Split creative production roughly in half. Drivers are refreshes of visual structures you have already validated, two or three new versions per month, low risk and predictable. Explorers are deliberate swings into new talent, new settings, new narrative pacing, new formats.
Explorers have a lower hit rate and that is the point. The winners open audience pockets you could not have reached with a refresh of an existing winner, and under a retrieval driven system those new pockets are where incremental growth actually comes from. Brands that only produce drivers see performance decay slowly and predictably, because they are refreshing the map without ever extending it.
The brands winning on Meta right now are not the ones with the best single ad. They are the ones with the widest set of genuinely different reasons a stranger might buy, each one built into its own creative.
The Reporting Trap Nobody Warns You About
There is a measurement problem baked into this shift that catches out careful operators. When you consolidate to two ad sets and run thirty creatives inside them, your per-ad reporting becomes far less useful for decision making. Individual creatives get uneven delivery by design, because the system is deliberately routing different creatives to different pockets. Killing the ad with the worst surface level ROAS often kills the one thing reaching a segment nothing else reaches.
The correction is to judge at the campaign level on contribution margin and blended efficiency, and to judge creative at the concept level rather than the asset level. Group your creatives by the four brief attributes, pain point, persona, awareness level and format, and evaluate which concepts are earning delivery over a two to four week window. That tells you what to produce more of. Individual asset ROAS over three days tells you almost nothing.
This is where a lot of accounts get quietly sabotaged by good intentions. Daily optimisation habits that made sense under the old system now actively prune the diversity the new system depends on.
What This Looks Like in Practice
A wellness brand I work with came to me with fourteen ad sets across three campaigns and a creative library of roughly forty assets that, on inspection, represented four concepts. Same founder to camera hook, same before and after structure, same ingredient callout, same discount close, all recut. Performance had drifted for five months and the team had responded by adding more ad sets, which made it worse.
We fixed deduplication on their Conversions API setup first, which alone cleaned up a meaningful chunk of phantom conversion data. Then we built a single consolidated prospecting campaign alongside the old structure and moved a quarter of budget into it. In parallel we rewrote the creative brief around persona and awareness level rather than visual treatment, and produced eleven genuinely distinct concepts over three weeks.
The first two weeks looked flat, which is normal and is the point at which most teams abandon the change. By week six the consolidated campaign was carrying the majority of spend at a materially better contribution margin than the old structure had managed in half a year. The lift did not come from a breakthrough ad. It came from giving the system more ways to find a buyer.
Inside the system
How we build this for brands
The bottleneck in adapting to Andromeda is almost never strategy. It is production throughput. Knowing you need eleven distinct concepts is easy. Producing them at a cadence you can sustain every month, while still running the business, is where most brands stall. We solve that with a voice of customer engine that mines reviews and support conversations into the raw material for creative, so every concept starts from a pain point a real customer articulated rather than a guess in a brief, and a creative agent that turns those angles into image and video assets for Meta and TikTok without a two week studio cycle.
Sitting underneath that, a profit dashboard built from live Shopify and ad data, with a reporting agent that surfaces contribution margin weekly rather than surface level ROAS daily, so the account is judged on the metric that actually reflects the change. Part of this runs live for portfolio brands today; the full system is what we deploy when we take a brand on.
Meta Ads Audit
Find Out Why Your Meta Account Stopped Scaling
I will go through your account structure, your signal setup, and your creative library, tell you how much of the auction you are actually reaching, and give you the restructure sequence in the order that will not crash your learning. Just the diagnosis and the plan.
Book Your AuditFrequently asked questions
What is Meta Andromeda?
Andromeda is the retrieval layer Meta rebuilt between late 2024 and 2026. It decides which ads from the enormous pool of eligible ads get pulled forward and considered for a given user, before the auction itself runs. Meta Lattice then handles ranking, deciding which of those retrieved ads wins the impression. The practical effect is that creative diversity now determines how much of the auction you can reach, because each genuinely distinct creative is treated as its own entity for retrieval.
How many creatives should I run per ad set under Andromeda?
Aim for 20 to 30 genuinely different creatives per ad set, mixing formats such as UGC, studio product shots, testimonials, demonstrations, and catalogue ads. The number matters far less than the diversity. Twenty five creatives that are variations of the same concept behave like one creative in retrieval. Eight meaningfully different concepts, each with a distinct pain point, persona, and format, will outperform thirty near-identical variants.
How many ad sets should a DTC brand run per campaign in 2026?
Two to three per campaign is the working standard: one broad prospecting, one retargeting, and optionally one testing. Legacy structures with ten or more ad sets fragment your conversion signal, create audience overlap where your own ad sets compete, and slow the learning phase. Consolidating concentrates signal where the system can actually use it.
What is a creative similarity score and why does it matter?
Creative similarity measures how alike your ads are across hook, visual structure, message, and format. Keeping similarity below roughly 40% across a campaign means your creatives are distinct enough to be retrieved for different audience pockets. High similarity means you are competing against yourself for the same slice of the auction rather than expanding into new segments.
Will restructuring my Meta ad account crash performance?
It can if you do it all at once. Consolidating ten ad sets into two overnight resets learning across the account and typically produces a two to three week dip. Build the new consolidated structure alongside the existing one, shift 20 to 30% of budget into it, let it exit learning, then migrate the rest once it is matching or beating the old structure on contribution margin.
Does broad targeting really beat lookalike audiences now?
In most DTC accounts, yes. Because retrieval leans on creative signals to find the right pockets of users, manually narrowing the audience mostly removes options from a system that is better at finding those people than you are. Lookalikes still have a role in retargeting and in very small or niche catalogues, but broad plus diverse creative has become the default prospecting structure across ecommerce accounts.
How long does it take to see results after restructuring?
Expect two weeks of flat or slightly down performance while the consolidated structure exits learning, then a clearer read from weeks three to six. Judge the change on contribution margin at campaign level over that window, not on daily ROAS. Most teams that abandon the restructure do so in week two, which is precisely when the data is least informative.
About the author
Caner Veli built Liquiproof to global distribution across 3,000+ retailers, then exited. He now runs Purposeful Profits using a combination of operator strategy and AI-powered systems he has built and uses daily, having 10x'd monthly revenue in his own business in the last 90 days.