← Video Library · AI Tools

How to Scale Your Meta Ads with AI Like a PRO

Published January 1, 2025 · 279 views on YouTube

Key takeaways

How does Meta’s AI actually decide which ads to show?

Meta uses its own machine-learning model to predict how well an ad will perform before it ever fully runs, scoring it in the ad auction using signals, model training, and your account history. Ads that score higher get subsidized visibility, better placement, and cheaper cost-per-acquisition, so understanding what feeds that score is the foundation for using AI to improve ad performance.

According to the video, the formula Meta uses internally is: maximizes estimated [defined goals] subject to constraints. That single sentence explains why “signals” (your creative, copy, images, and video) and “constraints” (budget, targeting, interest restrictions) are the two levers advertisers actually control.

What are “signals” and why do they matter more than manual targeting?

Signals are the information advertisers give Meta’s AI through ad creative, copy, images, and video, rather than through manual interest or demographic targeting, and they let advertisers tell the algorithm who to find instead of guessing at interest categories. Because Meta has reduced access to many interest-based targeting options, sending strong creative signals has become the more reliable way to reach a specific audience.

To generate signals for an ideal audience, the presenter’s process is:

  1. Describe the ideal audience’s characteristics in detail (e.g., innovative, willing to experiment, open to networking).
  2. Feed that description into ChatGPT (or Claude) along with the offer and landing page content to draft ad copy.
  3. Ask the AI to rewrite the copy in different copywriting frameworks and “voices” so the tests are genuinely distinct from each other, not minor variations.
  4. Pick 2-3 of those variations at random to test rather than trying to test everything at once.
  5. Ask the AI to make the copy compliant with Facebook’s ad policies before publishing.

Why do “constraints” work against Meta’s AI?

Constraints are anything that limits the AI’s ability to explore, including budget caps, demographic targeting, and interest targeting, and the video’s guidance is to use as few constraints as possible so the algorithm can learn and improve. The comparison drawn in the video is to ChatGPT itself: the AI needs enough context and setup (in Meta’s case, correct optimization objectives and Advantage Plus settings) to have a real chance at finding good results, and heavy constraints operate like tying the AI’s hands behind its back.

What AI tools are used to test ad creative and copy?

The video walks through a specific creative-testing workflow that pairs ChatGPT or Claude for copywriting with Canva for quick static designs and AI avatar platforms for video ads. The tools named are:

Purpose Tool(s) mentioned
Ad copy and headline generation ChatGPT (primary), Claude, Bing/Copilot (free option)
Quick static image ad design Canva
AI avatar / URL-to-video ads Creatify, HeyGen, Arcads (described as a personal favorite for raw avatar footage)
Captions on video ads Captions app (phone)

The stated workflow is to test creatives (images/video) first, and once a winning creative is found, layer in copy testing next, especially the first three lines of the ad. For video, the process is to first validate a message works in copy form via ChatGPT, then generate an AI avatar script from that proven message, testing multiple avatar personas and backgrounds to see which resonates.

What do you do when a working ad campaign suddenly stops converting?

When a previously converting campaign abruptly stalls for several days, the recommended fix is to deliberately create and run a creative that looks the visual opposite of what’s currently running, in order to send Meta’s AI a different signal and jolt it into a new audience pocket. In the video, this is done by asking ChatGPT for a list of distinct visual design styles (e.g., “colourpop,” “cosmic”), building creatives around a style very different from the current ads, and running that new creative even if it doesn’t convert immediately, with the expectation that the shift can restart the original ads’ performance within about a week.

Full video transcript

Meta is actually using AI to predict every time any of us puts up an ad. It tries to predict how well that ad will perform, and it uses machine learning, it uses signals, and it uses model training. So why do we care about this?

Thanks for tuning in to this session of our AI Hack Demo with Lyanna Ling, who’s going to be demonstrating her experience running Facebook ads and how to use AI so that you can run more effective Facebook ads. Lyanna, thanks so much for joining us. Thanks for having me, I’m so excited to be here and to share some really actionable stuff. So do you mind, Darby, if I just jump right in? Go ahead.

All right, I wanted to show you basically three things. The first one is an AI that probably everybody’s been using, but they don’t really realize that they’ve been using it if you’ve been running Facebook ads, and I find that a lot of people don’t talk about it. I’ll show you some fun stuff, some kind of sexy stuff here that’s going to just help you create some amazing ads really quickly, in the blink of an eye. However, I did want to make sure that we covered this part because it’s so crucial to the foundation: it’s actually Meta’s own AI. I don’t know if you know this, but Meta has actually been using AI for a while with how to optimize your ads and figure out what posts to put out there, and in the last couple of years they’ve actually released a new model that they’ve trained up. The key thing here is understanding how that works so that you can then use AI to leverage that and make it truly your superpower. It’s not about taking AI and churning out thousands of different ads; I prefer to do things a little bit more intentionally.

So let me share one thing that I learned. This is a new model that they’re training, which means everything we know about ads, you have to throw out the window and start fresh. Let me share my screen. I went to an invite-only event at Meta headquarters in Toronto, and they had one of their data scientists there. I think I was the only one at the break who literally ran over to the corner to ask him questions, but he was talking about how they value ads in the auction, and I promise I’ll show you how this is relevant to you.

This has been my go-to for the past year, in which I’ve been able to get results over and over again in almost any industry. I know it looks a bit simple, but this is what they shared: this is how they determine the value of your ad in the Meta ads auction. Why should you care about this? If your ad is valued highly, Meta will actually subsidize you so that you can show up higher in the auction, and the better people engage with it and the more conversions it gets, again Meta will put a higher value on that, and again you’ll get a higher position in the auction, which means cheaper CPA and a higher-quality audience.

So essentially, this is what it is: maximizes estimated defined goals subject to constraints. Let me break this down to show you how this fits into what I’m going to have you do, because you can literally take what I’m about to show you and start creating better ads. The first part here is really key: it says “maximizes estimated.” Meta is actually using AI to predict, every time any of us puts up an ad, how well that ad will perform, and it uses machine learning, it uses signals, and it uses model training. So why do we care about this? This was the keyword that I focused on: signals.

I learned that we can actually give signals to Meta’s AI in terms of what we want, and how we do that is through the creative, through the copy, and through the images and the videos. People have been complaining and saying they don’t have these interests anymore, they can’t choose certain demographics, and this is actually better because we have much more control over it. Before, we didn’t. I don’t even know, for example, Russell Brunson at one point or Tony Robbins, I have no idea how they became an interest. This way I can actually signal to Meta, instead of guessing what’s in their interest, the type of people that I want to attract.

Darby, as an example here, since I know you were going to sit back and let me teach, but I like to be interactive, so I’m going to pull you in for a minute: give me some characteristics of the type of people in your ideal audience. Yeah, absolutely, our ideal audience characteristic-wise, from a solopreneur, entrepreneur, or content-creator standpoint, they’re usually innovative, they’re trying out new tools and technologies, they’re willing to experiment and get outside their comfort zone a little bit more, they’re open to networking, connecting, and putting themselves in different environments to grow and adapt. Okay, perfect, hold that thought, I’m going to use that as our example for creating the ads we’re going to create next.

So now we can take that and create ads which will then send signals to Meta that this is the type of person we’re looking for. That’s the first thing, and why I think this is really a turning point for advertisers, because we now have control over this. The rest of it: “define goal” means the optimization objective that you choose on your ads, and “subject to constraints” means literally anything, like the budget. They know nobody has an unlimited budget, but when we say spend $100 a day, that’s a constraint. When we add in demographic targeting, when we add in interest targeting, any of that is a constraint. What they want is to have as few constraints as possible to allow their AI to explore, to learn, and to evolve, which should in theory give us better results.

I’m starting to see that the reason some people may not be seeing results, and this is just a guess, is that they don’t truly understand how it works and they haven’t made sure they’ve set up everything so the AI is set up for success. It’s similar to how, with ChatGPT, you have to give it some information, you have to give it some context. Same thing here with Meta’s newest model, a lot of which is under what they call Advantage Plus. You need to make sure you have certain settings set up, your audiences loaded, give it the best chance you can for it to work.

I don’t have time to go through all of these, but just know that these are important, so you should take some time to learn and think about which optimization objective you’re going to choose, because that will affect all of this, and think about the types of constraints you’re putting on it, because the more you put on, the more the AI will basically be operating as if its hands are tied behind its back, and it’s not going to give you the best results.

I wanted to focus on signals because I’ve found that this is the biggest needle-mover, and I want you to experience that as well. I’m going to show you a couple of different tools I use, and I know some of you may have heard of these tools, but instead of saying “yeah, I’ve heard of it before,” I’d like you to commit while watching this video to actually using it. How much have you actually used it? Because I still use these tools to this day to manage millions of dollars a year, and it still works even on the smaller campaigns, where we’re running for a nonprofit, for example, at $10 a day.

So the first thing we’re going to do is take a look at the copy, and I usually get ChatGPT to help me. I’m going to share a couple of little tricks that I use. I have the paid version, and I’ve actually sent people to the free version. If you’re watching this, you can go to Bing.com, I think it’s called Copilot now, and try using it there as well to see what results you get. You can also use Claude, however you pronounce it, that’s also pretty good at writing, but I tend to stick with ChatGPT for my first round because I also find Claude is a little more polite and has a bit more limitations on it, so it might tell me off for saying I can’t write like this person. I find ChatGPT lets me be a little freer.

So what I’ll do is just ask it, and I don’t do really fancy prompts, I know I should probably get better at prompting, but I’m usually a little impatient and I just want results. I’ll do something like: “You’re a direct-response marketer and copywriter, please help me write a Facebook ad,” and I’ll tell it exactly what I’m looking for. I’ll say my audience is people who like to network, they’re on the cutting edge or want to be, business owners, influencers, and so on, and I’ll put the offer, and then I’ll go to the landing page and copy everything in there. For this example, I’ll put the offer as an amazing free opt-in by Darby Rollins that will help me use AI never before.

So I’ll do something basic here, because I want to show you how I take this to the next level. I just ignore all the emojis, I personally don’t use emojis a lot in ads. What I do is a lot of testing, so the key thing is getting you started and then using it to test a ton of different things, because that’s actually how I win: I just get something up and then I start testing. By the way, you can also say “please rewrite so that it is in compliance with Facebook ads policy,” and it still works even if I’m not spelling right. It’s not bad with this, but you should still review everything to confirm, because it sometimes doesn’t pick up nuances very well, but it’ll get you pretty far.

I’ll create a version zero of this, and then I’ll start to change the voice. I have a Google doc, which I’m happy to share with everybody, where I systematically go through different frameworks and different approaches and different styles until I find one that works. When I’m testing Meta ads, I’ll actually do creatives first, and I’ll show you how in a moment, but I wanted to hit copy first because it’s a little more involved. I’ll use all the same copy and test different videos and images, and once I find a winning one, I’ll start to test different copy, especially the first three lines of the Facebook ad, using a lot of the frameworks and voices I have here.

I have 17 frameworks and a bunch of different voices, I just asked it for different types of copywriters and marketers it already knew about. It depends on your budget: some budgets are big enough to test everything, but if not, I’ll just pick three things at random to test. Even if you don’t have access to this document, you can just ask it to pick three copywriting frameworks it knows and rewrite the copy with them, and a lot of times one of them will just take off and I won’t even have to go further, and that’s the beauty of AI, it’s sped everything up for me.

Even these different voices, when I’m testing ads, I find that what moves the needle more is creating tests that are very different from each other, not too similar. Try to find voices that sound very different, like a Robert Collier voice will sound very different from a Harry Marshall voice, which will sound very different from a Jay Abraham voice. I’ll ask it to rewrite in various voices to get something very different to test, because I’m using this to send signals to Meta about the type of person I’m looking for, and if it’s taking too long to find somebody, you have to come up with different angles to nudge it and say you’re not looking in the right place, so look in different places.

I also changed up the hooks, the first three lines, based on emotion, and tested literally every single one of those emotions, and I’m often surprised by which ones win. I think one was gratitude for a SaaS product, and I thought there’s no way this was silly, why should I put this in here, but I tested it anyway and it ended up working, even though it wasn’t close to Thanksgiving. So sometimes you never know what’s actually going to work. I also have tone and voice, there’s so much you could test, that’s why I said just choose three things for your copy to test.

By the way, you can do the same thing with your emails, and test the same thing on your landing page. If my ads are working really well, a little extra tip is I’ll go in and start making changes to the headline at the top of the landing page, that tends to move the needle for me too. I’ll go through the same process, choosing a different framework, voice, or hook, and start testing that on the landing page as well.

Let me go to creatives, because creatives are more fun and it’s usually the first thing I test. I don’t always necessarily use AI for this, AI is absolutely my superpower to make things go faster, but a lot of times when starting out with a new Facebook ad and a new funnel, I’ll go into Canva, create an Instagram post, a square, and type in a search, in this case something like “technology” since Darby is in technology and AI. What I like about this is I can quickly scroll through and see what sticks out, similar to how people encounter your ad on Facebook or Instagram: they’re scrolling and not paying much attention, then they see your ad and you want something that stops their scroll so they’ll look further.

A lot of these designs are really busy. I’ll pick something simpler, maybe just a headline. I’ll ask ChatGPT to write five persuasive and compelling headlines, something like “unlock AI strategies to grow your influence,” and put that in the headline. It’s obviously big at first but we can make it smaller. I’ll use the same creative and test that, picking one randomly, and just look for something really simple to start with. This isn’t using some big fancy AI SaaS, but I find it quite effective, and I personally prefer to choose manually because I find AI still can’t quite replicate that instinct yet, though I’m sure there’s a formula someone will figure out eventually.

So that’s the first thing: find some type of design inside Canva, then ask ChatGPT or Claude to create a couple of really compelling, persuasive headlines, and test that. Right now, at the time we’re recording this, images are actually working really well, better than video. I just launched a lead-gen campaign two days ago and I’m already under $5, and it’s just the same design with different people and different hooks in the copy, and it found the one that’s working, so we’re very happy with that cost.

In terms of using AI a little more, I also use it to create AI avatars. I use tools like Creatify, which makes it easy to create ads, they have a URL-to-video-ad feature where you paste your link, whether you’re selling something from your store or you have a summit or a free offer, and it writes the scripts for you and you choose whichever AI avatar you want. What I like about this is, once I know a message works because we tested it with ChatGPT, I’ll create a script and have an AI avatar deliver it, choosing different personas: this one looks more casual in a kitchen, this one looks more professional, this one looks more like a tech person, this one looks more like an influencer with purple lights behind him. I’ll create a whole bunch of these to see which resonates, always on the lookout for better and more diverse AI avatars.

I also use HeyGen, they have a lot of pretty good avatars with good backgrounds, including one I used until it stopped working where she was actually walking and people were amazed, even if the lip sync was a bit off, most people don’t really care, they’re just focused on the message. HeyGen also lets you paste in a URL and it’ll write and produce the video for you. My absolute favorite is Arcads, which has the best types of avatars, the ones that work best for me. What I do there is create just the raw footage, I’m not using any of their templates for anything fancy, I just want the footage so I can give it to a video editor to add their own b-roll, or I’ll put it into Descript or use a captions app on my phone to add captions, since I want it to look and feel like my other social media videos, similar to how a lot of people create viral videos on TikTok, Instagram, Facebook, and YouTube.

I’m only using these avatar tools to get the AI actor to say the script, then I do whatever I want after that. Sometimes I’ll just throw it up as-is to see what happens, to show you the variety of actors available. After choosing one, you also have options to change the voice and the speed; decreasing stability makes it sound less flat and more expressive. A little tip: before you render and use your credits, play the preview so you can hear how it sounds, and adjust speed or stability as needed before clicking generate, because I’ve wasted way too many credits clicking generate impatiently.

I write the script with ChatGPT, edit it a little for compliance and clarity, then preview it before generating. We used to spend so much time and money hiring human actors from Fiverr to do this, so now I test a whole bunch of AI avatar variations instead. What I sometimes do is hire a human once I know what the winning creative is, or if you have access to a group of human influencers, you can give them the winning message and let them be creative in delivering it their own way, which has also produced some really good creative, though not everyone has access to that, which is why I always point people to AI avatars as the alternative.

Once you understand how Meta’s ad AI works, you can use that to your advantage: use AI to create variations, create different hooks, and send those signals to Meta using tools like Canva, HeyGen, and Arcads, along with others out there that do the same thing.

One last thing I wanted to mention, going back to the formula: if you’ve run ads for any length of time, whether you’re a beginner or more advanced, you’ll find you get caught in a pocket where everything is going well and then suddenly stops for several days and you can’t figure out why. Going back to this formula, all we need to do is send a different signal to Meta, and remember, we do that through the creative. One of the best tips I’ve gotten is that, especially if I’m running a purchase campaign that’s just not converting and all the activity falls off a cliff for several days, I’ll go into Canva and use whatever AI tool I want to create creatives that look the opposite of what I’m already using.

I went into ChatGPT and asked it for a list of all the different varieties of design styles out there that are very different from each other, and put them together inside a Canva document, naming them, like “colourpop” and “cosmic.” I don’t just stick with the tool’s suggestion, I go back to the design; the keyword just gives me a direction. If a style looks the opposite of what I would normally run, I’ll choose that and edit it a little to make it relevant to my offer, then run it. It may not get conversions right away, but the point is that I’m sending a signal to Meta that I want to get out of this pocket and go look at another audience, do something different. It jolts the campaign out and pushes it into a different pocket because it’s reading your signal, and that’s been one of the best tips I’ve gotten, because I’ll run it for maybe a week and then the other ads start working again because I’ve forced it out of that other pocket.

That was the other tip I had for you and your audience, Darby, which I hope people will be able to take and put into action. Absolutely, and I appreciate you sharing those tips with us, Lyanna, super helpful stuff for getting started and getting out there, testing creatives, working on those signals, and shaking things up. Don’t be afraid to go opposite of what you might think is working to reset the algorithm, and tap into AI for that extra spark of creativity, and don’t forget to test and try things out. Any last words or tips you’d want to leave with everyone? Just take action, pick one thing, and take action. Love it. All right, Lyanna, thank you so much. For those who want to connect with you outside the summit, where can they find you? You can find me on social media at “the lead gen queen,” and if you want to dive deeper into Meta and Facebook ads, you can find me inside the Meta Insiders community. All right, sounds good, thanks for tuning in, we’ll see you all in the next video.

Watch on YouTube

Related tutorials