Stop Making 5,000 Variations: What 'Reasonable Creative Diversity' Actually Means After Andromeda

Picture two ad accounts on the same Monday morning.

The first one ships 80 new ads. Same hook, same offer, same person on camera, just the background swapped, the headline reworded, the text nudged 4px to the left. Brute force. The bet is that volume alone will shake a winner loose.

The second ships 12 ads. But they're built across four different angles, aimed at three different buyers, in a mix of formats. Fewer files, more actual variety.

A year ago I'd have told you the first account had the edge. More shots on goal, more chances to hit. After Andromeda, I think it's the second one that wins, and it isn't close. Here's my take on why, and what "reasonable diversity" actually means now that the machine is doing work the spend used to do.

Andromeda is a retrieval problem, not a volume contest

The clearest explanation I've heard of Andromeda came from someone on the Meta side describing what the system actually does. Before any ranking happens, before the auction even kicks in, Meta has to retrieve which ads are even worth considering for the person who's just landed on a feed. That retrieval step is what Andromeda rebuilt.

And the reason they rebuilt it is us. Generative tools let everyone spin up a huge number of creative variations almost overnight. Backgrounds, copy, mixing and matching. One person on the platform called it a "Cambrian explosion" of creatives. Suddenly the system had far more ads to sift through for every single impression, and the old retrieval engine couldn't keep up.

So here's the part that should change how you brief creative. If you make 25 versions of one ad, only a handful of those are genuinely relevant to any given person. The colours, the context, the background it shows up against, the type of content that person actually consumes. The rest just aren't a match for them, even if they're technically "different" files.

Andromeda's whole job is to pull the ones most likely to land for that specific person. Which means stacking up 25 near-identical variations doesn't give the system more to work with. It gives it 25 versions of roughly the same signal.

Why three of twenty getting spend is the system working

This is the bit that trips people up, so I want to sit on it.

You launch 20 ads. Three of them get real delivery. The other 17 barely spend. The instinct is to read that as failure, like you wasted the effort on 17 duds.

I'd read it the opposite way. That's the machine doing its job.

The way it was put to me: Meta attempts delivery across your variations, works out which ones are actually performing, and moves the dollars onto those. The other 17 weren't necessarily bad. They were just lower-performing for the people it tested them on, and you genuinely don't want your budget sitting on lower-performing ads. A lot of the sorting that used to cost you spend to figure out is now done up front, in those server rooms, before your money's properly in play.

To put that in perspective: in the old world, finding your three winners out of twenty meant paying to test all twenty at some manual budget. You ate that cost. Now the system shortcuts a big chunk of it. So the goal isn't "get all 20 to spend." The goal is to give it a strong, varied set of 20 so the three it picks are genuinely good, not the least-bad of a samey bunch.

What "reasonable diversity" actually means

Same source again, and this line stuck with me: creating an extremely large number of ads probably won't help you more than having a reasonable diversity of them.

"Reasonable diversity" is doing a lot of work in that sentence, so let me define it the way I'd brief it.

It does not mean random. It's not 5,000 wild swings where half the ads are built for an avatar that has nothing to do with your brand. That's not diversity, that's noise, and the system still has to wade through it.

What it means is deliberate variety across the things that actually change who an ad resonates with:

• Angle. The reason someone buys. A skincare buyer might want fewer steps in their routine, or fewer breakouts, or to look less tired in photos. Three different angles, three different pockets of the market.
• Avatar. Who you're talking to. The 25-year-old and the 55-year-old don't see themselves in the same person on camera, even with an identical script.
• Format. Static, UGC, founder piece to camera, demo, a copy-led image. Different formats reach different parts of the funnel.
• Awareness level. Someone who already knows your category needs a different message from someone who doesn't even know the problem has a fix yet.

Think of your whole market as one pie. Each genuinely different concept reaches a different slice. Brute-force volume just keeps re-serving the same slice from slightly different angles.

Group by avatar AND concept, not one or the other

Here's the practical shift I've made, and it's the single biggest one.

It's not enough to group your testing by concept alone. You group by avatar and concept together. The "fewer steps" angle aimed at the time-poor parent is a different unit from the "fewer steps" angle aimed at the minimalist twenty-something, even if the words overlap. Built and grouped separately, the system can match each to the right people. Mashed together, your targeting goes slightly mushy and neither does as well as it should.

I started paying real attention to this around the middle of 2025, right as the Andromeda changes were landing, and the timing wasn't a coincidence. Once you group properly, the diversity stops being cosmetic. Each cluster is a clean, distinct signal pointed at a clean, distinct audience.

It also reframes what you're producing. You're not making "80 ads this week." You're making four or five concept-avatar clusters, each with a few variations so Meta has room to pick a winner inside the cluster. That's a far smaller number of files doing a far better job.

Where the AI supply chain helps, and where it doesn't

I'm not anti-volume tooling. The cheap, fast generation of creative is real, and you should use it. But you need to be honest about which half of the job it's good at.

The way one of the Meta folks framed the long game stuck with me. Machines are never going to be out-optimised by a human again. They move budget faster, react to the auction quicker, sort performance better than any of us sitting in Ads Manager refreshing a dashboard. That fight is over, and we lost it, and that's fine.

But machines are not good at the part that actually matters most. Coming up with a novel message. Looking at the performance data and reasoning about why this angle landed with this audience and that one didn't. That's human work. Taste, intuition, a real read on the customer.

So here's how I split it. Use AI to produce the variations once you know what you're saying. Re-skinning a proven concept across formats, generating backgrounds, churning out the supporting versions inside a cluster. That's the supply chain, and the machine is brilliant at it.

But the angle itself, the avatar you're talking to, the read on why something worked so you can do more of it? That's where a human earns their keep. No model is going to read a batch of reviews and feel the thing your customer is too polite to say out loud.

If you flip that, if you let AI invent the strategy and you spend your day manually shoving budget around, you've got it exactly backwards. The machine should be moving the money. You should be deciding what's worth saying.

What I'd actually do this week

If your creative process right now is "make as many variations as possible and pray," I'd stop and rebuild it around clusters.

Pick three or four real angles, grounded in things your customers actually say. Map each to the avatar it speaks to. Then, and only then, use whatever AI tooling you like to build a handful of variations per cluster. Group by avatar and concept. Let Meta pick the winners inside each one, and read the results to learn which concepts are working, not just which files.

You'll ship fewer ads and learn more from them. That's the whole trade after Andromeda.

If you've been pumping out volume for months and you're not sure whether your "diversity" is real diversity or 50 versions of the same signal, that's exactly the kind of thing a fresh look at your creative history will surface fast. A Signal/Noise Audit walks back through what you've actually been testing and shows you where the genuine gaps are, which avatars and angles you've never properly briefed. Worth doing before you generate another hundred ads into the same slice of the pie.