If your CAC has crept up over the last year and you haven't changed much, you're in good company.
Most of the brands we speak to are still running Meta the way it worked in 2021, with interest audiences, lots of ad sets and a steady stream of ads that all say roughly the same thing.
The platform has moved on from that. Meta now reads your ads, works out who each one is for, and finds those people itself. That changes where your time should go. Less of it goes on audiences and settings, and a lot more goes on research, personas and the concepts you put into the account.
This page goes through the whole thing in order. First how the machine works, then what that means for how you build creative, and every so often, how we do it at DTC Mode.
What changed between 2023 and 2025
Meta used to give everyone labels. If you interacted with a post about dogs, you'd get tagged with an interest in pets, and a pet brand could target that interest. It worked well enough, but an interest in pets doesn't tell you someone owns a pet, and it tells you even less about whether they'd like your brand.
Over three years Meta rolled out three big pieces of AI that replaced most of that system.
One large model that replaced hundreds of smaller ones, so Meta predicts performance across every goal and placement at once.
A new system for picking which ads are even considered for each person, out of tens of millions.
Meta's largest ads model, trained like a large language model, which passes what it learns down to the rest of the system.
Together they mean Meta places people by how they behave instead of by labels. Every time you watch, like, comment or buy, your position shifts a little, and people who behave alike end up close together.
When you launch an ad on broad targeting, Meta spends a little across lots of people, watches who responds and buys, and then moves spend towards the people who look like them. That's what broad targeting does in practice.
The simplest way to sum it up is that your creative now does most of your targeting. The rest of this page explains why, and what to do about it.
What happens before an ad shows
Every time someone opens Instagram or Facebook and there's a slot for an ad, Meta runs the same three steps in a fraction of a second.
1. Retrieval
Out of every live ad, Meta picks a shortlist of a few thousand that could be relevant to this person. This is Andromeda's job.
2. Ranking
Meta then scores that shortlist on how likely each ad is to get this person to click, watch and buy. It uses what they've done on the platform before, with ads and with normal posts. Lattice and the models GEM trains sit here.
3. The auction
The best-ranked ads go into an auction, and the one with the strongest mix of bid, predicted result and quality wins the slot.
Most advice about Meta focuses on the auction, because it's the part with settings in it. The first two steps matter more. If your ad never makes the shortlist, your bid never comes into it.
Andromeda, the part that picks the shortlist
Andromeda is the retrieval step. Meta announced it in December 2024 and reported a 6% improvement in recall (how good the system is at finding the right ads) and an 8% improvement in ad quality on some segments. It runs on NVIDIA Grace Hopper chips and Meta's own AI hardware, which is what lets it look at so many ads so quickly.
The important part for brands is how it decides. Andromeda builds its picture of each ad from the creative: what's said, what's on screen, the text, who's in it. That's why the creative has become the targeting.
A video of a man in his fifties talking about stiff knees will get picked for very different people than a video of a 25 year old runner talking about recovery, even if both sit in the same broad ad set with the same settings.
The trap most accounts fall into
When ads look and sound too alike, Meta treats them as the same ad. They get grouped together and served to the same pool of people. You can have 200 ads live and Meta might count 20.
24 ads in the account
3 ads as far as Meta is concerned
So if your team is making lots of small variations of one idea, most of that work is going to the same people. The fix is variety at the level of the idea, which we'll get to in chapter 09.
Lattice, the part that ranks
Lattice is the oldest of the three. Meta announced it in May 2023. Before it, Meta said "hundreds of models were separately trained, served, and optimized", each one looking after a different placement or goal. Lattice merged them into one large model that predicts how an ad will perform on clicks, video views, conversions and ad quality, across Feed, Stories and Reels, all at once.
Two things about it matter for you.
The first is that it looks at short and long signals together. A click happens in seconds, but a purchase might happen days later, after someone has seen your brand a few times. Lattice was built to learn from both, which is part of why Meta judges an ad on the whole journey it starts and not only on the click.
The second is that what Meta learns in one place carries over to the others. A strong Reels ad teaches the system something it can use in Feed.
You'll see some agencies talk about Lattice as if it's brand new. It has been live since 2023, so if your CAC changed recently, Lattice probably isn't the reason. It's still useful to understand, because it explains how the ranking step got as good as it is.
GEM, the model that teaches the others
GEM stands for Generative Ads Recommendation Model. Meta calls it the largest foundation model for recommendations in the industry, trained at the same kind of scale as a large language model, across thousands of GPUs.
GEM doesn't pick your ads directly. The easiest way to think about it is as a teacher. It learns from a huge amount of behaviour, both with ads and with organic posts, including long histories of what each person has done. Then it passes what it learns down to the hundreds of smaller models that do the live ranking.
Meta reported that once GEM was in place, ad conversions went up 5% on Instagram and 3% on Facebook Feed in the second quarter of 2025.
In plain terms, here's what it means for you:
- Meta learns from what people do with normal posts as well as ads, so ads that look like the content people already watch tend to beat ads that look like adverts.
- The format and the creative itself are part of what the model reads, so the same message in a very different format can reach a different group of people.
- Each update makes Meta better at matching ads to people, which means it needs more creative from you that's different enough to have something good to match.
The auction, and why your bid is your economics
Once ranking is done, the top ads go into the auction. Meta publishes how the winner is chosen. It's the ad with the highest total value:
Two brands showing the same product to the same person can win at very different prices. That leads to two things worth holding on to.
You can't set relevance
Estimated action rate and ad quality are Meta's read on what has already happened. You can't change them in the moment, and you can't make an ad "relevant" in general, because every ad is relevant to different people. What you can do is give Meta more ads that are different from each other, so there's a strong match for more of the people it could reach.
Your bid comes from your margins
The brand that can afford to pay more for a customer wins more auctions. That comes down to gross margin, average order value, and what a customer is worth to you over time. A brand with better unit economics can run weaker ads than you and still outspend you. That's why we always start with the numbers before we look at the ads.
Make sure Meta can see your sales
Signal quality feeds in here too. If your pixel and Conversions API only pass back 70 of every 100 sales, Meta thinks your ads convert worse than they do, and you lose auctions you should be winning. Check the Event Match Quality score in Events Manager and make sure purchases are coming through properly. Past that point, extra tracking tools tend to make a small difference that's hard to measure.
What you still control in targeting
With Advantage+ audience, Meta splits your targeting settings into two groups. Some are firm controls that it sticks to. The rest are treated as suggestions, a starting point Meta will go past if it finds better people.
Controls
Meta won't go outside these.
- ✓Location
- ✓Minimum age
- ✓Exclusions, such as past customers
- ✓Language
Suggestions
A starting point. Meta can reach beyond them.
- ~Age range
- ~Gender
- ~Interests and detailed targeting
- ~Lookalike and custom audiences
In practice that means going broad and putting your effort into exclusions. Meta knows your existing customers are likely to buy, so if you don't exclude them, a good chunk of your prospecting budget can end up on people who were going to buy from you anyway.
Creative is the targeting now
This is the part that changes how you should spend your week, so it's worth slowing down here.
Picture all of Meta's users spread across a huge map, with people who behave alike sitting close together. When you launch an ad, Meta reads it, makes a guess at where on the map the right people are, tests a little spend around there, and then moves towards whoever buys.
A specific ad finds a specific group, which is what you want. The catch is that every group runs out. Once Meta has shown an ad to most of the people in its group, it has two options. It can keep showing it to the same people, so your frequency goes up. Or it can move out to people further away who are less likely to buy, so your CAC goes up. That's what creative fatigue looks like from Meta's side.
The way to grow is to give Meta more groups to go after. Each new idea aimed at a different type of buyer opens up a new part of the map. Extra ads saying the same thing don't do that, because Meta groups them and sends them to the people who have already seen them.
So when you look at your account, count how many different types of people your ads are speaking to. That number tells you far more about your room to scale than the total number of ads.
What a concept is
Because the creative does the targeting, the unit you plan around is the concept. A concept is three things put together, and the format comes after.
The person you're talking to, and who's in the ad.
The problem you talk about, and why your product is the answer.
What you put in front of them: a bundle, a trial, a first-order deal.
Change any one of the three and you have a new concept.
Change the persona, the angle or the offer and you have a new concept that reaches a new group of people. Format comes after. You decide the concept first and then choose the format that gets it across best.
A lot of brands start the other way round. Someone says "we need more UGC", and the brand ends up with ten new videos that all make the same point to the same person.
- Persona
- A keen walker in her late fifties whose knees ache on the way down hills, and who's started cutting her weekend walks short.
- Angle
- Stiff knees don't have to mean shorter walks. Show her back on the route she gave up.
- Offer
- 30 day starter pack with free delivery.
- Formats
- Talking-head testimonial from a real walker, a before-and-after route static, and a short founder explainer on the ingredients.
The order to test in
Most brands test in the wrong order. They spend months on formats, which is the smallest lever, and get to the angle last. We go the other way.
- AngleWhere weak ads usually fail. The argument for buying isn't strong enough.
- OfferBig swings in results, though offers tire quickly.
- PersonaOpens up new groups of buyers once the angle is working.
- FormatHelps with reach. Rarely the reason an ad wins.
If an ad isn't working, putting the same weak argument in front of a different persona won't fix it. Fix the angle first.
Inside a concept that's working, the hook is the cheapest thing to test. A new hook keeps the same concept but counts as a new ad, so it can reach new people. Record five or six hooks every time you film, and you've got tests ready to go for weeks.
How to do the research
The best creative strategists spend most of their time on research. It's also the part most brands skip, which is exactly why it gives you an edge. This is where we look, in the order we look.
Your own ad account
Before making anything new, find what's already proven. Which concepts have held the most spend at a CAC you're happy with? Which winners have gone tired? New work should usually start close to something that has already worked.
Ads Manager, broken down by conceptYour reviews
Read your reviews, support emails and post-purchase survey answers, and copy out the exact words people use. Tag each one by the problem it describes. Phrases that come up again and again make your best hooks.
Reviews · support inbox · surveysReddit and forums
Reviews tell you why people bought. Forums tell you how people talk about the problem before they know your product exists. Note what they complain about, how often it comes up and how strongly they feel about it.
Subreddits · Facebook groups · niche forumsCompetitors, and brands outside your category
Use the Meta Ad Library to see what others are running and which ads have been live longest, since a long-running ad is usually a profitable one. Read competitor reviews for complaints you can answer. Then look at other categories for formats and angles, because if you only copy direct competitors, your ads end up looking like theirs.
Meta Ad Library · ForeplayWhat makes your product different
List what your product does that the alternatives don't, and match each point to a complaint people have about those alternatives. That's your reason to be chosen, and you'll want proof for every claim.
Product specs · competitor complaints
By the end you should have one document full of real customer language, sorted by problem. That's what the personas get built from.
How to build proper personas
A persona is a specific person with a specific problem in a specific situation. It sounds obvious, but most persona documents we're sent are a list of demographics.
Women, 25 to 45, interested in fashion.
Every ad written for this person ends up vague, because there's nothing specific to say to her.
A mum in her early thirties who needs one dress for a client meeting and a Friday night out, for under £80.
You can already picture the ad: the setting, the problem and the line she'd stop scrolling for.
The same product can serve lots of personas, because each problem is a different way in. Acne at 20 is a different problem from hormonal acne at 32, and that's different again at 42. Each one is its own persona, its own concept and, on Meta, its own group of buyers.
For every persona, we write down the same set of things:
Back at work after her second child, juggling client meetings with nursery pick-ups.
"Nothing in my wardrobe works for both bits of my life anymore."
A few pieces she can wear to work and straight out after.
A work event or a night out in the diary for next week.
She's been burnt by fit before and hates sending things back.
Real women her size wearing it, and an easy returns line.
That last line matters more than people expect. Someone who already knows the kind of solution they want needs to hear why yours is better. Someone who's only aware of the problem needs to be taught first. Those are different ads, and often different landing pages. We usually start a new brand at solution aware because it converts faster, then move down to problem aware for scale.
Three to five strong personas is plenty to start with. In our experience, brands get into trouble by keeping personas too broad far more often than by making them too narrow.
How to set concepts up in the account
Account structure exists for a reason. Campaigns and ad sets tell Meta which ads should learn together and which should learn separately. Data doesn't flow freely between ad sets, so each one builds its own picture of who to go after.
That gives you two rules.
One concept per ad set
Put three concepts in one ad set, get a 3x return, and you have no idea which concept did it. With one concept per ad set, a winner is obvious, and you know exactly what to make more of.
Keep everything else together
Lots of campaigns and ad sets, each with a small budget, split your conversion data so thinly that none of them learn. Fewer, bigger buckets learn faster.
Statics first, video to scale
Statics are quick and cheap to make, so we use them to find out which angle, offer and persona work, often within a few days. When a concept wins, that's when we put money into video for it, because video is usually where you scale.
Testing and iterating
Once concepts are live, the job turns into a weekly loop.
Look at results by concept, at the ad set level. Ad-level ROAS is unreliable, so avoid big calls from it.
Cut or scale using CAC against what you can afford to pay for a customer.
New hooks, creators and formats on the winners. Tired winners often come back with fresh hooks.
Launch new concepts every round, so Meta always has new groups to go after.
Data makes the decisions. It tells Meta who your buyer is so it can find more of them, and it tells you when to kill an ad and when to put more behind it.
The brands that get dangerously good at Meta run this loop every week, and they start each new round of creative from what the last one taught them.
Common myths we still hear
A lot of Meta advice was true in 2020 and hasn't been updated. These are the ones that come up most on our calls.
