Meta’s Andromeda update moved the main performance lever from audience settings to the ads themselves. Meta says this plainly in its own advertiser guidance: creative diversification is now the primary input the delivery system works with. What Meta does not give you is an operating definition of “diverse.” Most accounts I audit think they have creative diversity because they have forty ads live. Then you look, and it is one idea wearing forty outfits.
This is the playbook I wish had existed when the rollout hit: how the system groups your ads, what counts as a genuinely different concept, how many you need, and how to build a production system that keeps up. If you want the background on what Andromeda is and why this shift happened, I covered that in my Meta Andromeda explainer. This article assumes that context and gets tactical.
How Andromeda groups your ads
Andromeda’s retrieval models read your ad the way a person would, then more so. Computer vision on the imagery, language models on the copy, audio analysis on video sound. Out of that it forms a representation of what the ad is, conceptually. Ads whose representations sit close together get treated as the same idea when the system builds its short list of candidates for a given user.
Practitioners call these clusters “entity IDs.” You will not find that term in Meta’s documentation, and I would not build a religion on the exact mechanics, but the observable behavior is well established across enough accounts to treat as a working rule: near-duplicate ads share one retrieval slot. Twenty variations of the same concept give you roughly the coverage of one ad. Twenty distinct concepts give you twenty entry points into different retrieval contexts.
What does not create a new concept in the system’s eyes: a different headline color, a reordered opening sentence, a new crop of the same photo, different background music, a swapped CTA button. These are variants. Variants have a real job, which I will get to, but they do not expand your reach into new pockets of the audience.
Concepts versus variants: the distinction that decides everything
A concept is a distinct answer to the question “why should this specific kind of person care?” Change the persona, the pain point, the message architecture, or the format family, and you have a new concept. A variant is the same answer delivered with a cosmetic difference.
Take a sleep supplement. These are four different concepts:
- A UGC-style video of a shift nurse describing falling asleep at 8 a.m. after work
- A static comparison chart against melatonin, aimed at people who already tried melatonin and hated the grogginess
- A founder-to-camera video explaining why they formulated without melatonin at all
- A carousel walking through the ingredient research, aimed at the skeptical label-reader
And these are four variants of one concept: the nurse video with a different hook line, the nurse video cropped to square, the nurse video with captions in a new style, the nurse video trimmed to 15 seconds. Useful assets. One retrieval slot.
The five dimensions of real diversity
When I plan a creative batch, every concept has to differ from the others on at least two of these five dimensions. This is an extension of the four-dimension framework AdsUploader published, which is the most operational treatment of this I have seen. I add message architecture as a fifth, because two ads to the same persona in the same format can still be completely different ideas.
| Dimension | What it means | Examples |
|---|---|---|
| Format | The structural type of the asset | Static, short video, UGC video, carousel, text-heavy, meme-style |
| Persona | Who the ad is visibly for | New parent, shift worker, athlete, retiree, small-business owner |
| Angle | The benefit or pain point carried | Save time, stop a specific pain, social belonging, price, status |
| Setting | The world the ad happens in | Home, gym, office, commute, outdoors, on the job |
| Message architecture | The persuasion structure | Testimonial, us-versus-old-way, founder story, demonstration, statistic-led, objection handling |
A batch built this way cannot collapse into one cluster, because the ads are different objects all the way down, and each one lands in a different retrieval context. That is the entire trick.
How many concepts you actually need
More is not the answer. This deserves emphasis because “Andromeda rewards volume” became a popular misreading of the update, and it cost people money. When creative supply outruns the budget’s ability to generate conversion data per concept, learning fragments. Monica Shukla made this case in AdExchanger in January 2026, and the account-level evidence backs her: twenty-plus concepts on a small budget mostly buys you twenty ads stuck in learning.
The working numbers I use, consistent with what Atria and other creative analytics platforms recommend:
- Five to ten live concepts per campaign for most accounts. Start at five if your budget is under about 50 dollars a day, closer to ten as budget grows.
- Two to four variants per proven concept, added after a concept shows it can win, to extend its lifespan and cover placements.
- A refresh rhythm of two to three weeks for active prospecting campaigns, because fatigue arrives faster post-Andromeda. Watch the signals rather than the calendar; I keep a separate guide on diagnosing creative fatigue with the specific thresholds, publishing later this week.
The ratio worth internalizing: concepts buy reach into new retrieval contexts, variants buy longevity inside contexts you have already won. Accounts that only make variants plateau. Accounts that only make concepts burn winners without milking them.
The fake-diversity audit
Run this on your account this week. It takes twenty minutes.
Pull every active ad in your main prospecting campaign into a spreadsheet. For each one, write down its value on the five dimensions above. Then group ads that match on four or more dimensions. Each group is, functionally, one concept.
Most accounts that “have 30 ads live” discover they have three to five concepts. One D2C account I reviewed recently had 24 active ads that reduced to two concepts: a product-on-white static in eight variants, and one testimonial video in sixteen trims and crops. On paper, a diverse account. In retrieval terms, two tickets.
The gap between your ad count and your concept count is your to-do list.
Building the production system
The uncomfortable arithmetic: eight live concepts refreshed every three weeks means producing roughly ten to twelve genuinely new concepts a month, before variants. Most in-house setups were sized for a fraction of that. A few things that keep the pipeline affordable:
Mine your five dimensions as a grid, then fill cheap cells first. A founder selfie video costs an afternoon. A text-only static costs an hour. Not every concept needs production value; the dimension grid matters more than polish, and raw formats routinely beat produced ones on this platform.
Write the persona into everything. Vocabulary, casting, setting, on-screen text. Andromeda matches ads to people based on what the ad contains. A generic ad targets no one, which post-Andromeda means it reaches no one in particular.
Recycle angles across formats deliberately. A winning testimonial video’s core claim can become a statistic-led static, a comparison carousel, and an objection-handling founder video. Same insight, four concepts, because format and message architecture both changed.
Log everything. Concept, dimensions, launch date, outcome. Without a log you will re-test dead angles within a quarter. With one, your hit rate compounds, because you learn which dimensions move the needle for your product. My creative testing framework, coming later in this series, covers how to structure those reads without fooling yourself.
How to brief creators and designers for concept diversity
Production partners default to variants because variants are what most clients ask for and what existing assets make easy. If you want concepts, the brief has to force them. What changed in my briefs after the rollout:
Brief the dimension grid, one row per concept. Each row specifies persona, angle, format, setting, and message architecture. The creator’s freedom lives inside the row; the row itself is fixed. This single change ends the “here are ten versions of the thing you sent us” problem, because the deliverable is defined as ten different rows.
Give each concept its own reference ad, pulled from your own history or from the ad library, that shows the message architecture you want. “Make it like this structure, for this persona, about this pain” gets you a distinct concept. “Make some variations” gets you crops.
Ask for the hook in writing before production. If two concepts read the same in their first line, they will cluster together no matter how different the footage looks. Kill the collision at the script stage, where it costs nothing.
Accept lower polish on more rows. Given a fixed budget, eight rough concepts beat two premium ones in this system, because each concept is a separate lottery ticket into retrieval, and polish rarely decides which ticket wins. Spend the premium treatment on proven winners at the variant stage instead.
Measuring whether diversity is working
Two numbers in Ads Manager tell you whether your diversification is real or cosmetic, and both got more useful post-Andromeda:
- Creative similarity flags. Meta increasingly surfaces similarity warnings and fatigue states in the delivery column. If freshly launched ads flag as similar, the system just told you they share a cluster, whatever your spreadsheet says. Rebuild along a missing dimension.
- First-time impression ratio. A healthy diverse batch keeps finding new people, so the ratio stays high across the batch even as individual ads age. A batch that collapses into one cluster exhausts the same pocket of people together, and you see the ratio sink account-wide at the same rate.
Beyond the platform metrics, run the spend-share check weekly: what percentage of prospecting spend sits on the single biggest concept? Above roughly 60% and you are one fatigue event away from an account-level dip, which is a diversity failure even if the concept count looks fine on paper. The bench exists to keep that number honest.
Catalog ads count as diversity too
The Confect study of the Andromeda rollout, 115.7 billion impressions across 3,014 ecommerce advertisers, produced one finding worth stealing even if you distrust vendor research: advertisers with 60 to 100% of spend on catalog ads came through the transition dramatically better, showing 44% higher ROAS and 68% lower CPA than advertisers under 30% catalog share. Their best performers also layered design rules and five-plus design variants per product onto those catalogs, and video catalog formats over static ones.
The mechanism fits everything above. A catalog with a thousand products, each renderable in multiple designed templates, is a massive supply of machine-legible, genuinely distinct ad units, which is exactly what a retrieval engine wants to work with. If you sell physical products and your catalog ads are an afterthought with raw product feeds and default templates, that is probably the biggest diversity gap in your account, ahead of anything your creative team ships this month.
Where broad targeting fits
Diversified creative and broad targeting are one strategy, not two. The creative carries the targeting information, so the audience settings need to stay out of the way. Remove the interest stacks and lookalike restrictions on prospecting, keep only real constraints like geography, language, and regulated-category rules, and let each concept find its pocket. Tara Zirker’s recommendation, quoted by Social Media Examiner, goes as far as whole-country targeting with no age or gender selection, and in most accounts I have seen that direction hold up. Narrowing the audience while diversifying the creative is paying for reach twice and using it once.
A worked example of a launch batch
Product: a meal-prep subscription. Budget: 100 dollars a day. A first Andromeda-era batch might look like this, eight concepts, every pair differing on at least two dimensions:
- UGC video, busy parent, “weeknights back,” kitchen setting, testimonial structure
- Static, gym-goer, macro targets, gym setting, statistic-led
- Founder video, general, “why we started,” workshop setting, founder story
- Carousel, remote worker, lunch decision fatigue, home office, demonstration
- Text-only static, budget-conscious, cost-per-meal math, no setting, us-versus-takeout comparison
- Short video, new mover, “no kitchen setup yet,” empty apartment, objection handling
- Meme-style static, younger audience, cooking-fail humor, chaotic kitchen, relatable
- UGC video, older couple, portion control without waste, dining room, testimonial
Every one of those enters retrieval as a separate idea with a legible audience signal baked in. That is what creative diversity means now. Not more ads. More ideas, each wearing its audience on its sleeve.
Frequently asked questions
How many ad concepts should I run on Meta in 2026?
Five to ten genuinely distinct concepts per prospecting campaign, scaled to budget. Under about 50 dollars a day, start at five. Add variants only behind concepts that have already proven themselves.
What counts as a new concept versus a variation?
Change the persona, the pain point, the message structure, or the format family and you have a new concept. Change the headline, crop, color, music, or CTA and you have a variant of the same concept, which shares its retrieval slot.
Does creative diversity replace audience targeting completely?
Functionally, on prospecting, mostly yes. Keep geography, language, and any regulated-category constraints. The persona work moves inside the ads, where the delivery system can read it.
Do I need high production value for every concept?
No, and over-polishing is a common budget leak. Raw formats like founder selfies and text-heavy statics regularly win, and cheap formats let you cover more of the dimension grid. Save premium production for scaling proven winners.
How do I know if my ads are clustering together?
Watch for similarity and fatigue flags in the delivery column, spend piling onto one ad while siblings starve, and near-identical performance decay curves across supposedly different ads. Then run the fake-diversity audit above; a four-of-five dimension match between two ads means one cluster.
Sources: Meta for Business guidance on creative diversification with Andromeda, Meta Engineering (December 2024), AdsUploader on entity clustering, AdExchanger (January 2026), Atria creative strategy guide, Social Media Examiner.
