Andromeda Meta Ads: How to Brief Truly Different Creative

Everybody and their grandma wants "creative diversity" now… more variations, more hooks, more offers, more, more MORE!

But if you've tried the "shotgun" approach, and the numbers haven't gone up like you want, then you probably made one of these mistakes.

You rewrote the same promise 10 different ways, or you put the same pitch in a new outfit, or you came up with a great new idea and stuffed it into your old template where nobody could tell it was new.

All 3 end the same tragic way, because you're still giving people the same reason to care, however many ads you upload.

There are 3 real factors to consider here: what Meta says, what marketers do and our opinion here at Coppica.

Let's dive in.

IWhat Meta says Andromeda does, and why "upload more ads" is the wrong takeaway

Okay, so Andromeda is the big fancy retrieval stage of Meta's new ad system, which is an elaborate way of saying it's the first big filter.

In an October 2026 engineering post, Meta says retrieval "narrows a vast pool of eligible ads down to a small set of strong candidates in milliseconds", and then a slower, more expensive ranking stage takes a close look at whatever survived.

The numbers get a little ridiculous from there. Meta says the current system "scores every eligible ad" and surfaces the top candidates at over 1 billion ads per second, which it puts at 100 times the throughput of the 2024 system. What comes out the other end is what Meta calls a "user-ad relevance score".

It's basically a credit score for your ads.

Now, Meta does report a 2.4% increase in conversions on Facebook from Andromeda since 2025, and it tells advertisers to expect "better performance automatically". (Yeah, we've heard that one before.)

Meta's April 2025 business post says generative tools and automation let advertisers upload more creative, so the system has more to process. It also defines creative diversification as "creating a wide range of ad creatives with different themes, messages, and visuals to cater to diverse audience segments".

Plenty of marketers read all that and decided the answer was just simply more ads.

And while that can work...

There's only one way it actually will work, guaranteed.

And you kinda gotta think about it in the opposite way most marketers do.

First, getting looked at isn't the hard part anymore. If every eligible ad gets its credit score, your 7th rewrite isn't getting any consideration your 1st didn't already get.

And because the system can only match a person to what you gave it, even a system that checks every ad can't give a repeated idea a second reason to work. In other words, once Meta has decided your idea sucks, your idea sucks. Period.

Now, Meta's guidance around this is to use "volume and diversity", and most marketers only really remember volume.

To understand Meta's thinking a little more, let's look at the customer Meta chose to quote. Ben Paster of Dribbleup says the team "maintained steady performance while ramping from 3-4 new creatives a week to almost 50 on average".

The word there is steady, and he credits "new, diversified concepts", so even Meta's happy customer isn't saying that just more ads made more money.

Finally, Meta's percentages are also measuring other stuff. That 2.4% is a result Meta credits to its own system on Facebook, and a gain that shows up automatically tells you nothing about your creative.

In April, Meta reported a 9% lower cost per action for the streamlined Advantage+ sales campaign setup, and an 11% higher click-through rate and 7.6% higher conversion rate for campaigns using its image generation feature. I don't know about you, but most of the people I've talked to haven't seen those gains.

In fact, most marketers right now are just trying to scale volume, and they're having a really hard time figuring out exactly how to give Meta the diversity it's asking for. And notice that none of Meta's numbers compares an account full of reworded ads with an account full of genuinely different ideas, which is the one comparison that would actually tell you something, because rewording an ad and having a fundamentally different idea are two very different moves.

IIWhat are Entity IDs and similarity scores, and why do they matter?

Now, most of what you've read about Andromeda punishing similar ads comes from third-party write-ups, and the term "Entity ID" doesn't appear in Meta's 2024 engineering post or its April 2025 business post. So treat this next part as field reports (maybe even myths), and remember the people filing those reports are mostly companies selling ad tools or services.

DataAlly's Colby Flood describes an Entity ID as a group of creatives Meta's system considers visually similar, and says 200 live ads may really amount to 20 distinct ones. Flood also quotes a Meta rep calling it "an internal-only metric". Ads Uploader calls the term "the shorthand the industry uses", more of an explainer than an official Meta term. So marketers can't even agree on whether this thing is real or just a handy metaphor.

The similarity numbers are even softer. TheOptimizer passes along figures from other publishers: admetrics.io says scores above 60% trigger suppression, 303 London says to aim below 40%, and Tara Zirker says to treat these metrics as directional signals. If Creative Similarity has popped up in your Ads Manager, read it as a hint and don't lose sleep over it. Meta's "every eligible ad" line doesn't settle any of this either, because scoring an ad and grouping similar ads are 2 separate questions.

The thing to remember is that most of these write-ups are talking about visual sameness, like the same template, background, lighting or creator. Ads Uploader claims copy and audio count too, and Envisionit reports that 10 versions with slightly different hooks behave as 1. DataAlly says 2 scripts from the same creator in the same setting are likely to group together, which is why a new message in an old frame might not be enough on its own.

IIIWhy does rewording an ad feel like progress but change nothing?

Okay, so let's talk about rewording, because it's the cheapest thing in marketing that still looks like work. You take the winner, write it 6 new ways, slap each one on the winning image and call it a test.

Marketers would tell you that batch looks like 1 ad, and our bigger problem with it is that it also really only says 1 thing, just 6 different ways.

Think about it from the other side of the screen. If somebody didn't care about your 1st version, what new reason did you give them to care about your 7th? (Spoiler: you didn't.)

The words moved and the argument stayed put, so everyone it missed the first time is still missed. Now multiply that by a whole market writing from the same few AI models and you get what we call the Green Desert, where every message passes every check and the entire feed starts to look the same.

Even Ads Uploader's own example does this. It takes 1 benefit and dresses it up as UGC, studio footage and motion graphics to win more "tickets", even though every version communicates the same core benefit.

Now, that might well help delivery, but you've still only got 1 reason to buy.

So ask 2 questions about every new ad: is the answer new, and is the look new? Put those together and you get a simple grid, which also happens to hold all 3 mistakes from the top of this post.

IVHow do you brief a new ad with the nine questions?

Now, Meta's April 2025 post tells you to vary your themes and messages, but it never says how different is different enough. So this part is all us. And to be clear, nothing in any source says Meta's system knows these questions exist, because they're a tool for the humans writing the brief.

In our world, a new message is a new answer to one of the nine questions every reader silently asks. They come from The Persuasion Taxonomy, which is our free catalog of persuasion techniques documented from real ads.

One thing before you go copy those lines: borrow the move and leave the words, because those lines belong to the brands that wrote them. Then put 1 sentence at the top of every brief, which is the question this ad answers and its answer. If 2 briefs match on both, congratulations, you have 1 brief.

And each question has room for plenty of ideas. "Can I trust the claim?" could be a demonstration, an outside validator or a founder's story, and those are 3 very different ads.

You also don't need all 9 running at once. In other words, an ad earns its spot by answering a question your buyers are actually asking.

VSee it work: 1 made-up product, 9 different briefs

Okay, let's make this real and invent a product: a monthly bookkeeping subscription for freelance designers. The usual batch is 10 versions of 1 promise, like "Stop dreading tax season", "Never fear tax time again" and "Make tax season stress free". That's 3 lines and 1 answer, and it only talks to the people already dreading tax season.

Now aim each ad for that same product at a different question and change the execution each time, so every ad lands in the bottom right square of the grid. The lines below are all made up for this hypothetical product.

Look at the spreadsheet ad for a second. It's talking to someone who wants out of their current system and is scared of the move, and the tax season promise barely touches that person.

Of course, all of these ads have to be true of the product. If you can't actually do the migration, or the date is made up, change the brief.

VIHow do you audit your account against the nine questions?

The good news is you can do this in a spreadsheet in an afternoon.

1. Pull your active ads. Export each ad with its copy, a thumbnail and its spend.
2. Write the reason. For each ad, finish the sentence "You should care because...", using only what the ad actually says.
3. Tag the question. Mark the 1 question the ad mainly answers. Get a second person to tag separately, because if 2 people tag an ad differently, the ad is often muddled.
4. Tag the look. Add the format, the setting and who's on screen.
5. Count twice. Tally ads per question, then spend per question. The second tally matters more, because 1 out of 10 ads can eat nearly all the money.
6. Find the pile and the blanks. Note which questions hold most of your ads and spend, and which hold nothing. Ads with the same reason and the same look count as 1 idea.
7. Pick the blanks that matter. Read your sales calls, reviews and refund notes for objections, then pick the empty questions your buyers are really asking.
8. Brief 3 new ads. Give each one a blank question, an answer and a look you're not already running.

And if you want a second opinion on your tagging, paste your copy into our free analyzer, which takes up to 1,500 words and gives you 3 runs a day.

VIIHow do you read your results without fooling yourself?

Now, give each ad time before you judge it. Fair warning, this is marketer advice with no Meta source behind it: Envisionit says at least 5 to 7 days, and TheOptimizer says 7 days minimum, at least 1,000 impressions per creative, and up to 14 days for high-consideration products.

TheOptimizer also calls a 5 to 10% hit rate normal at scale. In other words, expect most of your new ads to lose, and don't panic when they do.

Log the question and the look next to spend and conversions, then read your results by question as well as by ad.

That second read is the important one. One winner could be a fluke of execution, but a question that keeps winning across different executions is telling you something real about your buyers.

Be careful comparing 2 ads head to head, too. Meta says its system tests variations "with different audiences", so your winner may have just been shown to easier buyers. TheOptimizer's setup of 1 creative per ad set with equal budgets gives you a cleaner read, and the price is more manual work.

And look, results vary by account, and a lift doesn't prove the cause. Your offer, the season, your budget, your competitors and Meta's own system can all move at the same time as your new ads.

Finally, if you change the answer and the look together and the numbers go up, you can't say which one did it. When you need to know, run the new answer in both the old look and the new look, then repeat the finding before you build on it.