For years, the Meta Ads playbook was: define your audience well (age, interests, lookalikes) and then put a good creative behind it. That playbook no longer describes how the platform actually works today. With Meta's current ranking system, publicly known as Andromeda, manual segmentation has lost weight against a different logic: the algorithm explores who to show each ad to based on who genuinely reacts, and it's the creative itself that provides that signal. In practice, the ad stopped being just the message, it became the targeting tool.
How Meta's segmentation logic has changed
The old model worked in blocks: you defined an audience with certain parameters, and that audience saw the ad. The current model works differently: Meta evaluates a much larger number of possible person ad combinations, identifying multidimensional behavioral patterns, not a single declared interest, but the combination of signals that best predicts who will react to a specific message. We can call these groups, hard to define by hand because they combine too many variables at once, affinity "bubbles": sets of people a human would never manually assemble with traditional segmentation tools, but that the system can identify based on how they react to content.
The practical consequence: the narrower you define the audience by hand, the less room you leave the algorithm to find those bubbles on its own. That's why Meta has been pushing toward broad audiences and Advantage+ style campaigns for a while now, giving the system more freedom to explore, in exchange for asking advertisers for more creative variety, not less.
The creative as a targeting signal, not just a message
Here's the most important mindset shift: every different creative you upload isn't just another way of saying the same thing, it's a different "hook" the algorithm can use to find a different pocket of audience. A video that attacks the "I don't have time" pain point and a carousel that attacks the "I don't trust the quality" pain point aren't competing for the same slice of audience: each one lets the system discover who cares more about which thing, something no manual interest based segmentation could anticipate with that level of precision.
In the old model, the sequence was: segment the audience → show that audience the right creative. In the current model, the sequence is reversed: upload diverse creatives → the system finds, for each one, the audience that responds best. Restricting manual segmentation in this context isn't caution, it's taking away the information the algorithm needs to do its job.
What "strategic creative diversity" means in this context
It's not enough to add variety for variety's sake, diversity that actually works as segmentation has to give the algorithm genuinely different angles to explore:
| Variation axis | Why it works as a targeting signal |
|---|---|
| Format (video, static, carousel, UGC) | Each format generates a different interaction pattern, and the system learns from that interaction who responds to which |
| Pain point / angle (price, time, trust, status) | Each pain point speaks to a different segment of your potential audience, even if they all buy the same product |
| Funnel stage (see the consumer journey) | An awareness creative and a decision stage creative draw reactions from people at different moments, widening the range of "bubbles" the system can map |
| Voice (founder, real customer, team, see real faces and voices) | Different voices generate different levels of trust depending on who's watching, another variable the algorithm can learn to associate with a segment |
The core idea: instead of chasing a closed segmentation and a single "perfect" creative for it, the goal is to cast several different hooks, different formats, different pain points, different funnel stages, and let the system explore which ones generate real affinity, then scale the ones that work best.
From exploration to scale
- Launch broad variety: several formats, several pain point angles, without over restricting the audience, let the system have something to explore with.
- Give it enough time and volume: the algorithm needs data to learn which hook works with whom, the same minimum learning volume principle developed in how much to invest in digital advertising.
- Identify which combinations generate the best interaction and conversion: that's where the system is indirectly revealing which segment has the highest affinity for each message.
- Scale the winning creatives: not necessarily by narrowing the audience by hand, but by letting the system itself concentrate delivery where it has already shown results (see how to lower your CAC).
So is manual segmentation useless now?
It still has a role, just a different one than before. Exclusions (removing people from the audience who are clearly not customers) still add value, because they save the system from exploring where it's not worth it, the same criteria developed in lever #1 of how to lower your CAC. What's lost weight is ultra specific inclusion based segmentation (stacking interests, narrow age ranges, hand picked narrow audiences): that kind of restriction now competes against the algorithm itself, instead of helping it.
Common mistakes
Still building ultra segmented ad sets with just one or two creatives each. This fragments both the audience and the algorithm's learning, consolidating structure, as explained in how to lower your CAC, usually performs better.
Diversifying only the visual side, not the angle or pain point. Three color variations of the same message don't give the system genuinely different hooks to explore.
Cutting "underperforming" creatives too quickly. The system needs time and data to find each piece's audience pocket, judging it within the first few hours doesn't give it that chance.
Restricting the audience by hand "just to be safe." In the current model, this usually has the opposite effect, it takes away the room the algorithm needs to find real affinity.
Frequently Asked Questions
What is Andromeda, Meta's algorithm?
It's the ad ranking and delivery system currently used by Meta, which evaluates a much larger number of person ad combinations than previous models, allowing for finer personalization without depending as heavily on the segmentation parameters set by the advertiser.
Does interest based segmentation still make sense on Meta Ads?
Ultra specific inclusion segmentation has lost relevance compared to broader audiences combined with diverse creatives. Exclusions (removing people who aren't customers) are still useful.
How many different creatives do I need for the algorithm to explore well?
There's no fixed number, it depends on budget and available data volume. The criteria matters more than the number: real variety in format and angle, not several cosmetic versions of the same idea.
Does this only apply to Meta, or also to Google Ads and TikTok Ads?
Andromeda is specific to Meta, but the underlying logic, systems that explore with broad audiences and rely increasingly on the creative as a signal, is also present in Google's Performance Max and TikTok Ads' delivery system.
Conclusion
Creative diversity stopped being just a way to avoid ad fatigue, today it's, literally, how segmentation is done on Meta. The more different hooks you give the algorithm to explore, the faster it will find the audience pockets with the highest real affinity, and the better it will be able to scale what works.
Still segmenting by hand instead of letting your creative do the work? Book a free audit with KLIV and let's review together how your account is structured. Book a call →