The End of Micro-Testing? Meta's New Andromeda Is Upending Everything I Thought I Knew About Digital Marketing
At the beginning of May, I saw an ad with a guy soy-facing next to text that read, "How everyone reacted when they saw their April Meta ads results."
Well damn.
Great callout, because I had been wondering the same thing and possibly making the same face.
Up until April of this year, the Meta ads I had been running for my clients were hovering consistently in the 15 to 25 ROAS range. Not to toot my own horn, but I have done a lot of studying on this. Not just the marketing skillset, but the statistical skillsets involved in data analysis: how to perform mathematical A/B tests, how to identify small changes that accumulate into large qualitative shifts in performance, and how to trim the fat from an ad campaign until the strongest version is left standing.
That method, small incremental changes and "trimming of fat," has kind of been my thing for the entire time I have worked in marketing.
Until April 2026.
I had heard whispers for a few months that Meta was rebuilding the way its ads algorithm worked, but I did not think too much of it. These platforms change all the time. Usually, it's easier to roll with the rollouts, so to speak, than to obsess over every technical release and try to get everything right on day one.
But what April did to my ROAS rates spun me into a light crisis of faith.
A system that had been working beautifully up to that point now looked like it had been run through a woodchipper. A ROAS of 5 or 6 would be perfectly respectable for a lot of companies. But again, at the risk of horn-tooting, that was far below what I had become accustomed to.
So I set out to learn about Andromeda, Meta's updated ad delivery architecture.
Oh buddy. The times, they are a-changin'.
And what I think we are seeing is Meta trying to bring the simplicity of Instagram's front-end "boost post" experience into the backend Ads Manager environment. Maybe not explicitly a 1:1 redesign, it is still too early for that, but little by little, it seems they are trying to wean the number nerds like me off of our need to have our hands on the wheel.
As of right now, Ads Manager still looks like it was designed by a fellow spreadsheet-pervert, but on the backend, changes must have been made, corners child-proofed, to make my usual inputs result in wildly different outputs.
In this week's article, I want to walk through what seems to be changing, why so many advertisers saw strange performance shifts this past April, and what I think the new best practices are for keeping ads profitable in this environment.
Long story short, as I am fond of saying, if you can develop a bedrock of strong creative skills, you will probably do just fine.
If you rely purely on the numbers, you may find yourself swimming against the current, making little progress, and spending a lot of cash while doing it.
A Warm Welcome to Our New Machine Overlords
Ad testing is not dead. I do not want to give that impression.
Rather, the polite suggestions Meta has been giving us for the past few years, the ones that say, "Oh, don't go through all that trouble. Why don't you let us tweak that setting for you? Just click this button, and we'll apply this change to every single ad you have running! No need to look first. Trust us. Think of the potential +0.5% in CTR!" are now less like suggestions and more like the axioms, the load-bearing pillars that make up the platform itself.
I have discussed machine learning quite a bit in this blog before, so I won't make this article a mini-bootcamp. But I will say this: if you do not understand machine learning at least a little bit, you may find yourself attributing magic to a very machinic process.
To use a slightly technical analogy, Meta's new system seems akin to a jump from "supervised" to "unsupervised" machine learning. The system appears to be moving from advertiser-constrained optimization toward algorithmic interpretation and delivery.
In English, dammit!
Meta would like very much if you gave their algorithm some initial data points to get started, and trusted their ability to find your customers better than you. And after a couple of months of testing, I admit they might have a point.
In the old paradigm, the advertiser gave Meta the map: people to target, instructions to follow. If I wanted to run an ad, I gave my exact customers through a Pixel, I chose the headlines based on previous ad performance and psychographic persona research, and for the creative, if I wanted a 1:1, 4:5, 16:9, or 9:16 image or video, I had to supply all of those formats individually. I trusted Meta's Advantage+ Audience suggestions, so I would generally turn that on, but I would also create lookalike audiences using the data collected from a Pixel, and use various ad sets to set different audiences. Only on rare occasions would I take Meta's "enhancement" suggestions to my ad headlines or images.
In the newer model, those levers still exist, and you can pull them to your heart's content, but you may be working harder than Meta would like you to.
Your desired audience, even a data-backed lookalike audience built from your Pixel or Conversions API, and the assets you upload, increasingly function more like suggestions than commands. Meta still takes your inputs seriously, but it also seems to trust its own interpretation of the ad's intended purpose, intended user, and intended moment of delivery more than it trusts your cute little audience theory.
Meta is playing matchmaker. They think they can find a "soulmate" for your ads, and they would rather you not get in their way.
Meta's new approach, I've heard described as "hyperspecific" targeting. They have crossed a qualitative threshold from using demographic signals like location, bounce rates, or engagement rates based on ads with certain keywords to matching a specific person with a specific creative asset at a specific moment when they've determined that person is most likely to respond.
Meta is not just looking at "Women, 35 to 50, interested in home decor. Ah, there's one now, deliver her the ad with the highest bid for that demo." Rather, now the algorithm is looking at tens of thousands of micro-signals that no human media buyer could reasonably track by hand.
- What pixel patterns are present in the content this person watches?
- What scroll movements or velocities correlate with various conversion goals?
- Which videos do they finish, and are there qualitative commonalities between them?
- Which colors, faces, environments, gestures, and words tend to make them pause?
That level of testing is not feasible for a human to perform. A person cannot reasonably test the performance implications of every frame, every facial expression, every line break, every caption structure, every background prop, every tone of voice, and every tiny visual cue across millions of users.
Technically, neither can a machine, but machines can make the attempt more easily.
Meta is opting for this route, not because their machine now understands "creative." Machines still do not "understand" it. Rather, Meta is taking this route now because their AI model has had a serious upgrade in its computer vision abilities and is sophisticated enough to process a huge number of signals and look for statistical relationships between tiny bits of creative material and user behavior.
I began to suspect a change like this after a recent Adam Mosseri announcement regarding how Instagram would treat unoriginal content going forward. He said the platform would begin reducing recommendations for accounts that regularly post unoriginal photos and carousels, unless that content was materially changed enough to count as something new. On face value, in the organic context, this is about rewarding original creators. But I realized something deeper in the paid-media context: Meta can increasingly read visual content at scale, judge similarity, and infer when two assets are symbolically or structurally close to each other.
That means ad creative is no longer just a singular data object you upload. It is a data class that the system can interpret and file objects under.
I do like the idea of things becoming more "idiot proof," but I cannot shake the feeling that the knobs and dials in Ads Manager are becoming increasingly decorative.
Standard targeting options are still there. You are still given placements. You are still given creative variation tools. You can still pretend that every decision is under your full control.
But I cannot, for the life of me, shake the familiar feeling like when my older brother would give me an unplugged controller while he and I played video games as children. I mashed buttons. He told me I was doing a great job. I felt pride. He got to carry on unbothered by my interests and goals.
There are pros and cons to this strange new world. On one hand, the system may be able to find buyers you never would have found manually. On the other hand, the advertiser's role changes in a way that can feel profoundly annoying if your whole personal brand is built on careful experimentation.
Machine learning models run on data. The larger the pool they can draw from, the better they tend to work. In the past, your campaigns may have benefited from sectioning off different tests into carefully controlled containers. In the new system, that same fragmentation can indirectly punish you because it reduces the amount of data available to each part of the system.
Rather than a series of neat experiments you monitor like a scientist, the future of Meta ads looks more like a witch's cauldron.
You throw in a bunch of parts and ingredients, stir, and wait for the dark magic to do its work on the other side of the bubbling fog.
The trick is making sure the feet of toad and eyes of newt are of good stock.
Known Unknown Entities
A term you are likely to hear more of is "Entity ID."
Historically, advertising platforms have categorized ads in terms of Creative IDs. Every ad you would upload to Meta would be assigned its own unique identifier, like a physical product being given a serial number. In the old model, if you duplicated an ad, you had a brand new ad with a brand new creative ID. Maybe the same image with a different headline. Maybe the same video with a slightly different opening scene. Maybe the same creator in the same room, but with three versions of the first sentence. To the old Meta system, the algorithm was none the wiser. They were all completely different ad creatives.
If I wanted to pay to run the same ad with very slight variations, I was able to do so. To the statistically minded advertiser, those unique identifiers allowed for various gradients of A/B testing.
To the new machine vision system, all those tests with slight variations just look like many copies of the same ad concept wearing the same long trench coat.
Hence the witch's-cauldron analogy.
Now, when you upload creative to Meta, it does something akin to boiling all the images down and comparing the viscosity of the resulting goo.
Your ad creative gets saved not only as the file that you upload, but also more deeply, as the relations between the pixels, their arrangements, and their movements. The same kind of "tokening" method that allows large language models to create or understand brand-new words also allows Meta AI to break ad creative down into microscopic datapoints and weight the relations between them.
Here's the first of a few kickers.
If Meta can interpret the underlying qualitative structure of an ad, then it does not have to treat every upload as a unique entity. It can see ads as particular embodiments of a general ad concept. It can recognize when several ads share the same visual elements, the same actor, the same opening scene, the same product arrangement, the same emotional posturing, or the same conceptual hook.
In other words, the platform may not care that you made five "different" ads if those five ads are, to the machine, basically the same goo.
Imagine I wanted to optimize an ad's creative asset in the old system. Let's say I was trying to sell a Bahamas vacation package using a photo of a model on a beach. In the old framework, I might have run several image variations from the same photoshoot:
- A model wearing a blue dress
- A model wearing a red dress
- An overjoyed expression
- A tranquil expression
If I kept everything else about the ad the same, such as the headline, the call to action, and the target audience, the ad's results might reveal something about how shirt color or facial expression affects click-through rate or purchase behavior in this audience.
That kind of test is being forced into irrelevance in the machine learning age.
Why? Because if, in two ads, the model is wearing two different colors of dress but standing in the same place, posed the same way, framed the same way, smiling the same way, with the same headline over them, then Meta is going to treat those two ads as conceptually the same thing.
The system will likely collapse them into the same conceptual bucket and assign them all the same Entity ID number. The machine does not necessarily see "three clean experiments." It may now see "same person, same product, same setting, same general message, minor shirt variation."
If you're the kind of person to say, "I don't care, man, I like the way I was doing things. I'm going to keep doing it the old way anyway!" first off, you sound fun, and I'd like to buy you a drink and pick your brain sometime. Secondly, though, Meta is going to punish your ads for this similarity. Rather than being a series of small, parallel tests, your ads will now be treated as the same Entity. Like five identical baby birds fighting for the same worm, they're all going to go hungry.
A trick to winning in this system is knowing where Meta's AI is putting the majority of its focus when it's determining whether to assign a creative a new or existing Entity ID. As all good content creators know, the first five seconds are the most important. If you upload ten videos with different scripts, but all ten open with the same creator standing in the same environment, Meta AI is going to disregard the vast majority of the differences, and there's a very good chance you're going to end up with bad results because your ads all got classified under the same Entity ID.
The Poisoned Chalice of Lazy Volume
Generative "AI" technologies have made it easier than ever to churn out a thousand little versions of the same ad with very slight tweaks to them, hoping that by uploading them all at the same time, you stumble on one with the "secret sauce" and then redirect all funds toward that winner. From a statistical perspective, that's just betting on noise. There is no science behind that tactic. It's what has lately come to be termed "lazy volume."
On the one hand, I think many of Meta's new changes are meant to combat "lazy volume" by forcing ad creators to come to the table with fresh and truly unique ideas. On the other hand, it does seem to really be punishing ad analysts since genuine A/B testing and "lazy volume" can sometimes seem like the same thing from a distance.
Here's the part that gets me. Meta is not taking a moral stand against using generative AI in advertising. They would just prefer you let their generative AI do the work of generating the lazy volume for you. Rather than you going through the trouble of putting your headline on your creative in the way that you want, Meta would prefer you give their platform all the disassembled pieces and let it do the mixing and matching for you.
To keep with the same Bahamas image-optimization example, you would not be uploading multiple shots from the same shoot where your model wore different colored dresses. You would upload the best image from each scene. The images you upload should be as visually diverse as possible: different locations, models, angles, and so on.
Then, Meta would like it very much if you, at that point, allowed them to add "enhancements" to your ad creative. Doing so allows Meta to generate new versions of your creative with your model in different colored dresses, different clothing entirely, a different pose, or a different location if it deems that would lead to more conversions. That's mostly harmless, but with GenAI, there is always a non-zero chance of getting an image you would never have approved, or copy for a service or feature that you do not actually sell.
That's still one of my biggest issues with generative AI in marketing: the technology is non-deterministic by design. You might be able to predict the output 99% of the time, but that 1% deviation might cost you thousands. That is a serious issue for brands that take their branding seriously or do not have piles of excess cash to throw at PR firms to fix bad ads.
But I digress.
Once again, this does not mean testing is dead. It just means you need to think a bit bigger when putting your ads together.
Do not bother with testing something like shirt color that can be easily changed by the AI. You, as the human in the situation, have a deeper grasp of what makes your audience tick. Use that knowledge to come up with drastically different ad approaches and styles.
The more I sit with the new algorithm changes, the more I think that fortune has never favored the bold more.
Test if a founder story beats a product demo. A testimonial versus motion graphics. Throw in a direct-response video sales letter for good measure, and a lifestyle image too, while you're at it. A raw phone-shot creator clip. A polished editorial visual. A selfie video on the fear of wasting money. An AI-generated song about aspiring to a better life.
Brother, there are no wrong answers. It's all goo in the end.
Look, I am all for anything that incentivizes pushing new artistic depths. I am just a little bitter for now because micro-testing was kind of my thing.
After some time to wrap my head around the best possible tactic, I have arrived at what I think will be the most helpful mindset going forward for people who want to split the difference between rigor and envelope-pushing.
Fun With Funnels
In the past, I would develop personas based on data signals and research, create assets based on unique selling points and pain point alleviation, then run small tests on controlled variations. One aspect at a time. One variable at a time. Watch CTR. Watch CPC. Watch ROAS. Watch landing page views. Watch near-conversion signals. Trim the fat. Scale the winners.
The approach still has some merit, but the new Meta system begs a looser relationship between strategy and execution.
Rather than separating every concept into its own little box, the platform seems to prefer that you build a large, singular creative ecosystem and let the algorithm handle the matching.
I believe the single best mindset a person can take to increase their chances of success while making their lives a little easier in Meta's new ad paradigm is to return to basics and focus on their customer funnel.
Let's briefly zoom out.
All of the marketing efforts you put into the world can be thought of as a giant net, or a funnel. The goal is to sift through the general public, find people who may be interested in your product or service, and then shepherd them along to the next stage of the journey.
That journey typically ends in a purchase, a phone call, a quote request, a booked consultation, or a visit to your website, and it can also loop back in on itself by retargeting past customers and reentering them into the funnel.
Each step in the funnel that a lead takes on the path towards becoming a customer represents a different stage of awareness.
Top-of-funnel leads are drivers on the highway before they see your billboard. They may have a problem, but they might barely be aware of it. Your billboard gets them to recognize the problem and suggests that a solution exists.
Middle-of-funnel leads are the ones who visit your website later. They are now aware of the problem and actively searching for solutions. They may be comparing you to competitors. They are asking whether you are real, whether you are trustworthy, and whether your offer fits their situation.
Bottom-of-funnel leads are on the precipice of buying. They are the ones calling you, messaging you, adding to cart, booking a meeting, checking the price, reading the guarantee, or looking for one final reason to act.
Post-purchase customers are the people who have already bought from you. You do not just let them drift away. Ideally, you continue nurturing them. You give them reasons to come back, upgrade, refer, review, or deepen the relationship.
People at each stage need different kinds of information to push them to the next.
Top-of-funnel leads need a reason to stop and care. They need something educational, surprising, funny, beautiful, weird, useful, or emotionally charged enough to interrupt the scroll. They need to be shown a dreamy outcome. They need the problem to become visible.
Middle-of-funnel leads need proof. They need testimonials, reviews, demonstrations, comparisons, process explanations, founder credibility, and evidence that you understand the world they live in. Evidence that you are who you say you are, do what you say you do, and why they should choose you over someone else who offers the same things.
Bottom-of-funnel leads need clarity and friction reduction. They need pricing, promo codes, limited-time offers, guarantees, FAQs, easy checkout, obvious next steps, or a strong reason not to wait. They are on the edge. All they need is that last little push to convert.
Post-purchase customers need reinforcement. They need check-ins, new offers, useful updates, community, and a reason to remember that buying from you was a good decision.
Any marketer worth their salt has likely already theorized these "stages" for their company or client. But this is often a part of the strategy phase, not the tactical, hands-on, "okay now let's actually make all the ads we said we should" phase.
If the new system rewards creative diversity in your ad ideas, then keeping your funnel stages in mind seems to be the easiest way to quickly and efficiently create many ads that would appeal to various people in various parts of your customer journey.
Keeping your funnel in mind will help you readjust your campaign structure as well. Whereas you may have previously used different ad sets to target different audience segments and target different creatives within that set, now it's probably better to have a single ad set using Pixel data to aid Meta's automatic Advantage+ targeting, populated with many individual ads aimed at various funnel stages.
Sticking with the Bahamas example, what are you trying to do with the eye-catching model and eye-catching-er offer? You are trying to engage top-of-funnel leads with big, fast value or a dream outcome that forces them to stop what they are doing and realize they have the problem you are talking about.
Rather than creating an ad set targeting a lookalike audience based on people who have not visited your website and using the ads within it to test various creatives, you would create diverse assets such as images without text, headlines, and primary texts that differ in approach, but align conceptually with appealing to top-of-funnel leads.
For each funnel stage, determine what information would appeal to them most and ideate a bunch of creative ways to convey that information. Each unique idea is a separate ad. Next, collect the assets you would use in each of those ads and upload them as separate assets that Meta can mix and match for you.
How (I Think) You Should Structure Your Ads
Having all your funnel stages covered by a few ad options is better than having a million different ads targeting the same stage. Meta can fudge the specifics. It needs help with spread.
To make a new ad campaign, start by brainstorming new ad concepts that would appeal to leads in all stages of your customer funnel.
Dream big. Remember, Meta is going to melt the things down into goo anyway.
Then, critically, put all of those ads into the same ad set and let Meta sort out the delivery.
This was the most unintuitive part for me. Take internal notes on which ads are aimed at which funnel stages. You can even put indicators like "ToF" or "BoF" in the ad title to help you remember which is which, but all of the ads, for all of the stages, need to go into the same ad set targeting an Advantage+ audience, possibly informed by a lookalike audience from your Pixel.
My new criteria for when to create new campaigns versus sets versus ads looks like this:
- Create new campaigns when ad objectives differ. If all your ads are aimed at online purchases, they can probably live in the same campaign. If some aim for website visitors, some aim for leads, and some aim for purchases, they need separate campaigns. The same goes if the conversion event changes, the geography or language changes, the budget source needs to be separated, or special ad category rules differ.
- Create new ad sets when the delivery logic actually needs to be different. If the placement strategy differs for a proven reason, create a new ad set. If you need to protect spend for a specific creative concept family, use a new ad set. If you have a large enough retargeting pool to justify isolation, use a new ad set. But do not create a new ad set just because you have a slightly different Advantage+ Audience suggestion. In that case, you may not be separating funnel stages. You may just be splitting the same purchase signal into smaller puddles.
- Create new ads for distinct creative ideas, not small superficial tests. The hook should change meaningfully. The format should change. The awareness stage, objection, proof mechanism, offer, CTA, use case, emotional trigger, or visual language should change. If you want to test blue dress versus red dress, that is fine. But understand that Meta may see those as two versions of the same basic idea, not two truly different ads, and you will split your potential audience between the two.
"Different" also needs to really mean different. For each funnel stage, try a different concept, a different creative medium, a different use case, and a different reason to care.
Of course, there will be exceptions to these "rules," so I urge you to do your own research. For example, I will still sometimes use two campaigns even if the objective is the same because one I use with a large budget for proven "winning" ads, and the other I use a fraction of the budget for evaluating "experimental" ad approaches. I run a separate ad set for retargeting because I have a custom audience of past purchasers identified by my Meta Pixel. That is a solved audience pool. There is no need for Meta's grubby hands to touch my targeting data.
You can take the testing out of the platform, but you cannot take the testing out of the tester.
Remember, your job is to fill the cauldron with the right ingredients. Meta's job is to decide which spoonful of which potion gets served to whom.
Thinking through your customer funnels will help you more easily create sufficiently diverse ads so as not to step on each other's toes.
Creative Is the New Targeting
A raw phone video of a tired founder talking about a problem in their garage does not send the same signal as a polished studio product demo. A testimonial from a calm older customer does not send the same signal as a fast-cut UGC clip from a young creator. A static lifestyle image does not send the same signal as a text-heavy motion graphic. A meme does not send the same signal as a video sales letter.
Each unique creative gives the algorithm a different set of clues about who might care.
If Meta is ingesting and interpreting the creative asset itself, then the creative itself becomes the input for the targeting mechanism rather than an explicit audience setting.
Of course, this assumes you have the technical foundation in place: a working Meta Pixel, clean conversion tracking, a decent landing page, and enough first-party signal for the system to learn from. Beyond that, you are encouraged to throw things at the wall to see what sticks.
New Meta ad success requires an active Pixel to give the targeting algorithm an initial signal to go in, and a wide variety of unique and diverse creatives.
Real creative diversity means changing the format, the environment, the actor, the emotional appeal, the funnel stage, the level of polish, the narrative structure, and the psychological trigger. If that sounds daunting, think through your customer funnel and ask yourself where the people in each stage are, what mindset they are in, what info they need, and how best to deliver that info to them.
One ad might teach. One might entertain. One might answer an objection. One might show the product in use. One might tell a customer story. One might explain the cost of doing nothing. One might give a direct offer or promo code.
For each of those ads, you might try a few different concepts. Do them all and add them all to the same ad set.
To me, for all my huffing and puffing, this does represent a way of giving power back to the true creatives by taking power from the statistics-brains.
A person who knows how to shoot, edit, frame, script, caption, light, score, and structure a video has an advantage. A person who understands why the first three seconds matter has an advantage. A person who understands the difference between information and persuasion has an advantage. A person who can translate a buyer persona into five different visual languages has an advantage.
A person who can tell you what a p-value is, for maybe the first time in recent memory, now has less of an advantage.
The New Testing Philosophy
So where does that leave testing? Alive, but maimed.
The test-and-learn mindset still works. The best idea should still win. One should still document what worked, what failed, and what to do next, and one should still avoid making business decisions based on vibes and the highest-paid person's opinion.
But the hypothesis has to get bigger and more, dare I say, more holistic.
What was once:
"I think headline A will outperform headline B, and dress A will outperform dress B. Let's run all 4 combinations and see what gets the best results."
Now becomes:
"I think this MoF persona responds better to practical demonstrations than polished brand claims. We should do this as an influencer selfie video and as an image carousel with trending audio."
A 2025 arXiv paper on Meta advertising experiments argued that Meta A/B tests can suffer from what the authors called "divergent delivery." In plain English, the system may not show each version of an ad to the same kind of audience. The paper analyzed thousands of Meta experiments and found that A/B tests showed clear audience imbalance, while lift tests did not show meaningful imbalance.
That means even the results of a Meta A/B test can reflect both the creative itself and the audience Meta decided to route that creative toward.
So no, testing is not dead, but if it was not scientific to start with, why bother arguing over whether or not to pull the plug?
The new Meta ad analyst should not ask, "Which ad is better?" They should instead ask, "What kind of person does this creative speak to?"
That is messier, but holistic.
And probably closer to reality.
And if Meta decides to break everything again a couple of months from now, assuming I still have hair to pull out by then, I will see you right back here for a new analysis.