About Patrick: LinkedIn | Substack

Connect with Sociaaal : Webpage | LinkedIn | Substack

About Rocketship HQ : Website | LinkedIn | Newsletter | Youtube | Podcast Website

FULL TRANSCRIPT BELOW

Shamanth: Quick note, some of you heard me shut down the Mobile UA show rather publicly a few months ago in order to launch Intelligent Artifice as something new, as a new show, a new podcast. Turns out those should not have been two different things. I was trying to pivot and I should have been trying to evolve, so I’m back.

This is the Mobile UA Show. It’s reopened. Intelligent Artifice is our deep dive series on AI and mobile. Same me, same depth, one single feed.

So welcome and welcome back to the Mobile UA Show. Today’s episode is a part of our series Intelligent Artifice, a deep dive on how AI is transforming performance marketing. Since 2018 and over 250+ episodes, we’ve been deconstructing how the best performance marketers actually win. Every week I sit down with the operator shaping it.

Or we tear about  the ad systems behind the highest performing advertisers in the world. I’m Shamanth Rao, founder of RocketShip HQ. Let’s get into it.

Our guest today is Patrick Stuart-Constant the CEO of Sociaaal, which is an AI operator for apps,

Sociaaal buys apps, and through frontier data and generative AI helps make them into great businesses. Under his leadership, Sociaaal has grown to a $13M run rate, achieved profitability and scaled multiple apps through a lean model powered by frontier technology. I’m excited to speak with Patrick because I’ve known him for a very long time and he and his team at Sociaaal have been at the forefront of adopting AI way before it was cool, they’ve tested massive volumes, iterated aggressively they’ve tested massive volumes and adopted frontier models and technology well before they became mainstream. In today’s interview, I’m excited to dive into his perspectives on how generative AI has completely transformed growth marketing  and apps.

I’m excited to welcome Patrick Stuart-Constant to Intelligent Artifice. Patrick, welcome to the show.

Patrick: Lovely to be here.

Shamanth: Yep. Excited to have you, Patrick. I’ve admired a lot of your work for a very long time. Among other reasons for this is the fact that you guys were really adopting AI and automation way before it was cool, way before it was easy even, and we’ll talk about a lot of your journey and your learnings as we go ahead, but let’s start perhaps at the beginning, what was your creative operation like, when you guys really started doubling down on AI and automation, and what was the inspiration for starting to integrate AI so early, also given that, AI wasn’t nearly as easy or as advanced as it is today.

Patrick: Yeah, definitely. Thanks for the question. So I’d say that well, the company is three years old. For the first year we were working with an  agency and they were producing creatives the traditional way. And we were also doing a lot of like fixed CPI deals.

And then we started to internalize user acquisition and so also internalized creative production. And so that was about two years ago and maybe, yeah, actually maybe even a bit more, two and a half years ago now. And we went straight for AI creatives. I guess the first reason was to reinvent how an app operator functions in the world of generative AI, and that was part of the premise from the very start. I was personally like a relatively early adopter of Gen AI. I was using LLMs like models from OpenAI like DaVinci before ChatGPT and before it became really mainstream and I was following very closely all of the work.

…different people at Google were doing and what you could find out about it from different whistleblowers and things like that. So I was fairly fascinated with LLMs and generative AI for many years now. And yeah, the natural thing we did is we went  straight for Gen AI creatives. The bet has quite a lot of things built upon generative AI, is that if you can build the right systems around these foundational models to extract maximum value and really lean into the infinite potential of generative AI, then as models progress, a tide that lifts all boats. So if you’ve got, the right systems, skills, teams, processes, even ways of tracking data models, etc, in place.

…then as models progress, then like your whole user acquisition is lifted and it accelerates faster and faster. And so that’s what we saw. Initially when we were, when we started with generative AI ads, it was the Will Smith spaghetti era basically of generative AI.

I don’t know if you know that meme, but our bet was if we can produce something today, which is just about as good as using live creators, then, in a few years time  we’ll be really outperforming live creators and actors.

And so yeah that’s how we started with Gen AI creatives.

Shamanth: Yeah. And that’s an interesting pattern. I’ve noticed when I’ve talked to folks who are really pushing the envelope of AI today, that they were comfortable starting when the output quality was not good. Yeah. And that’s something I’ve noticed with a lot of the folks on the leading edge, that they stuck with it because I think there was a conviction that it would get better, and when things do get better, you guys would be at an advantage.

Sounds like that was the case for you as well. Switching gears to today, everyone talks about audiences getting tired of AI-generated creative. So talk to me about the ways in which AI fatigue is manifesting and also what you think is the way to counteract that as a marketer or an entrepreneur.

Patrick: Yeah.

Interesting question because I think, a lot of marketers are worried about this sort of AI fatigue and don’t want to lean into generic ads for this reason. I think there are two main dimensions to  Gen AI fatigue and AI fatigue in ads. The first, it’s like visual and you get an ad and the ad looks a tiny bit AI, and then you get those comments, oh, stop using AI in your ads, etc, etc. That dimension was like very true. Six months ago now we hardly get these comments under our videos because they’ve become nearly impossible to distinguish from reality. And within another six to 12 months, I don’t think you’d really be able to tell, like visually if an ad has been like filmed with a human actor or if it’s Gen AI.

So I think that dimension, it’s less of a problem, it’s less of a problem than it was six months ago and it’s not going to be a problem for very long. I think the other one, which is more fundamental, it’s people they might not even know it’s because of AI. It’s people are tired of AI slop and, AI slop is just basically meh content and I think that’s where you really have to inject a lot of creativity into the ads. You use Gen AI to build. You don’t just rely on LLMs to find the  concepts and find, these sort of ideas and marketing angles. But you really work on that with highly creative people to produce content, which is, interesting to people and which is original.

And your world of problems is that. Most people use similar prompts, like you generate a script for this app, etc, and like it converges towards this sort of quite mediocre content, which doesn’t stand out, which, is the opposite of what marketing should be doing, which is to stand out from the crowd.

And yeah, you have for that, you have to be careful how to use how you use Gen AI and to maintain true creativity in your marketing.

Shamanth: Yeah. Yeah. So from what you’re saying, it’s, the problem isn’t so much AI itself, the problem is lazy marketing. The people who just take the AI output without double checking.

And that is what is the root cause of AI slop. Yeah. Now if you actually put in a lot of thought and discernment into this, it wouldn’t be AI slop. Yeah. And that’s a great point and it’s certainly something I’m noticing as  well because I think if there’s enough thought and attention going into an ad, I think the AI can produce very good quality ads.

I think the slop happens when there isn’t that kind of thought and attention.

Patrick: Yeah. Exactly. I think, the problem in AI slop isn’t the AI part of it, it’s the slop. Yes.

Shamanth: Yeah.

For sure. Yeah. So how, on your team, how do you ensure process-wise that there’s less slop because I would imagine, if a marketer or a video editor is tasked with coming up with 10 ads, the natural temptation is to come up with the ones that are easiest to make, which would just be either copy-paste from a competitor or copy-paste what’s already working, or just take whatever the AI gives you.

What are some of the ways you’ve thought about doing this?

Patrick: Yeah.

You have to find the right balance between automating things and having, keeping humans in the loop and keeping the right humans in the loop. And so it’s you want  to have, and that’s the whole challenge is to have a high volume of high quality creatives. It’s pretty easy to not as, it isn’t actually, but it has been feasible for, many years to make a few a few dozen good ads, like high quality ads. Now, it’s also very easy to produce hundreds, thousands of ads. The challenge is producing thousands of high quality ads.

Yeah, you need to have you need to create the right system. I would say and system is both, you know, humans, tooling, like, you know, workflows, processes. In our case, we’re very sort of data-centric in our way of operating even in like for the creatives.

And so it’s also how we’ve trained our own data models to basically allow us to see the signal amongst the noise relatively quickly and relatively cheaply using a Bayesian approach. And so it’s creating this system which allows you to produce this high velocity, high quality, high quantity of it.

Shamanth: Sure. And if I might  drill down a bit on what you said with the system itself, what would you say characterizes the system that you guys have come up with versus a system that would basically result in AI slop. What do you think the big characteristics are?

Patrick: So I think it’s keeping humans to inject creativity. So to find new marketing angles, new ad concepts, new video types. I think that’s one of the very important things is keeping like, one of the departments in the company where we use AI the most is our creative department, and it’s also the department which is growing the most because…

Shamanth: Yeah,

Patrick: I’ve noticed, as we scaled up from a 100 to 500 to 1,000…

2000, now we’re close to 3000, we’re heading to 10,000 within the next few months. You have to have quite a lot of humans to make sure you keep this creativity, keep your ads refreshing and you keep innovating.

Shamanth: Yeah. Yeah. And.  You’re right that because the more ads you have, the more quality checks you need, the more directing and guidance you need. So this definitely makes sense. I was also very intrigued when you said, as you used more AI, you actually needed more humans.

Patrick: Yeah.

Shamanth: Was that surprising to you upfront?

Patrick: Yeah, it was. Initially, when we started with AI ads before I really understood that you needed to maintain a high level of creativity. I was like, this is amazing, like with two people, I can produce 2000 ads a month. That’s, traditional app studio might have eight people producing 50 ads a month. I was like, this is huge and even two people producing 2000 ads a month especially back then. Because that’s what we did originally. We scaled quite quickly with only a two-person team to 2,000 ads a month. And it was good. And, we were performing, we were beating like benchmarks, like market benchmarks and like we were scaling and it was doing okay.

But I had this feeling that we could do better and to do better, we basically needed to expand the team. And then we multiplied the number of  people in the creative team by three. So we went from two to six and we stayed at 2,000 ads a month for quite a few months. We saw performance, basically double. And so yeah, I guess it goes back to that point where it’s having a high quantity of ads and the high, velocity is very important, but, it’s having that also a high quality of ads and that human touch and creativity, which takes your sort of user acquisition to the next level.

Shamanth: That is so interesting.

I would love for you to take me back to the point of time when you had two creative team members, you’re producing 2000 ads. What were you noticing in the setup at the time that made you say, okay, I don’t want to have more ads per designer now actually want to add more humans. What was breaking in that point of time that led you to say, okay, we actually need more humans.

Patrick: So it was that multiplying variations. We’ve got quite a systematic process where we test like each variable by the ad and we test like dozens of hooks, dozens of actors, etc., for the  same winning ad and we notice that performs and it does increase performance, but those videos also have a much shorter lifetime.

And that the ones which really outperformed and. Which, nearly got well, or got to what we call internally unicorn status, which is, ads that we can run at incredible, incredibly low CPI/CPAs. The performance was just great on these ads at really high volumes of spend. Those ones were coming from basically brand new ads we were building using AI tools and so, you know, the idea was: what if we can build a high volume of brand new, completely different ads, then that’s where we will be able to really outperform. And I guess, basically it’s what, it’s one of the operating principles of the company since. Follow the data and whatever is working, double down on it.

And that’s what we track the data very closely. As soon as we see something performing, outperforming, we just double the resources on it, triple the resources.

Shamanth: Okay.

Patrick:  And that’s how we’ve been growing the company.

Shamanth: Okay. Yeah. So interesting. So obviously from what you’re saying, the end goal isn’t to produce 2000 ads, right?

The end game is to come up with a small number of unicorn ads. And the 2000 ads is in service of that. So if I understand correctly, you started noticing that the unicorn ads weren’t made with AI-driven iterations.

Patrick: So they were AI ads, but they were like nearly like side projects, like brand new ads. Oh, let’s do this. Let’s just try this. They weren’t a variation of an existing marketing angle.

Shamanth: Is there an example that comes to mind of a unicorn, an ad like this that was almost like a of being had that,

Patrick: so this goes back to the point of, you can’t rely too much on the LLMs to come up with the ad concepts themselves. And have that creativity and you, that’s not what’s also, and they’re like, it’s artificial intelligence. It’s not artificial creativity, artificial imagination. It’s, you know what, it’s not what these models are  optimized for, and it’s not really what you should expect of them and it’s not what they’re particularly good at. Ask an LLM to tell you a joke and the joke might be okay, but it will sound like very familiar and just be like a variation of some very classic joke and not usually that funny.

Like every time, that’s one of the tests I do. Every time there’s a new model, I ask it to. I’ve got a few questions, but one of them is I’ve got different prompts to ask, like to generate jokes to see if they’re getting any funnier or not. And they’re getting a tiny bit funnier, but it’s not yet hilarious.

But yeah, to get back to your question an example, it’s on one of our apps, we were getting like, not a lot, but a few reviews, maybe like 1% of reviews or something. People who couldn’t understand how to use the app. And if we’d recommendations and come up with script ideas, it would’ve probably told us to do a tutorial.

To better explain, or, just to improve onboarding and UX of the app. What we decided to do is take all of these reviews of people who didn’t understand how the app worked and. Basically done in a sort of very, troll-like, light-humor manner  insult the people who didn’t understand like, how can’t they understand this?

It’s the simplest thing, etc., etc. And yeah. And roasting the people who, who couldn’t understand how to use this fairly simple app. Yeah. And that ad really outperformed, like it really performed well. And yeah. And that’s not the sort of ad I’ve ever seen an LLM come up with. Yeah, it’s just because of probably guardrails, etc.

If you ask Claude to, to generate something like that, it won’t ever. And I think sometimes it’s by producing like in an. Like world of basically, one pretty soon, nearly infinite at least unlimited content being produced by LLMs, the way to stand out in your marketing, which is one of the main objectives of marketing.

The way to stand out is to produce things that would never produce.

Shamanth: Yeah. And. You are so right that this is not the sort of thing that an AI would produce unless you see the idea first.

Patrick: Yeah, if, and get true. But the idea, and that’s and you, that’s why you have to have a human loop in the right place.

Everywhere. Exactly. But yeah, you do need to see that, that  idea and yeah. If you see that idea and then you can, generate the scripts and things for it and that can do quite well sometimes. Sure. Depending which LLM and how you prompt it, but yeah. Exactly.

Shamanth: Yeah. Very interesting.

Very interesting.

Patrick: I want to encourage all brands out there to go out insulting their users. For one app where the brand works quite well, it’s quite a Gen Z

Shamanth: yeah.

Patrick: app and it worked it worked quite well and under that ad, like all the comments were best ad ever, etc, etc.

People actually, because it was refreshing, I think.

Shamanth: Yeah.

Patrick: And that’s what you need to do is. Even if it looked like it was an AI voice, like doing the whole thing, etc, there wasn’t a single comment, ah, AI, AI slop, etc.

Shamanth: Yeah.

Patrick: Because

I think it felt refreshing and so it felt human in a way.

Shamanth: Yeah. Yeah. And I can imagine, and obviously if it feels more human, it generates more engagement. If it generates more engagements, the algorithms love it. Algorithms love it. Algorithms give it more love. So that’s a virtuous cycle right there.

Yeah.

Yeah. And to switch gears a bit I know in the prep call for this recording, you mentioned that you are  working on something that could changes how you decide what ads even get tested.

Patrick: Yeah.

The things which, some of the things which we’re working on, which really excite me, it’s originally it was hard to scale efficiently to 5K ads and you know how to build it with some sort of software internally to test manage all those creatives, test them, deploy them on the campaigns refresh. And that was the first one. But that was relatively easy to do. And now we’re scaling up towards 10K ads. But what really excites me is how can we scale up to a 100k ads a month? Because when you’ve got brand new tangents, you can’t just test all those ads on ad networks because of the cost, campaign limits, all of those different things would, I know, and managing that sort of level of creative testing would make it highly complicated and inefficient.

So things we are working on as a first filter, like ultimately I think, humans need to be involved. Humans will be watching it. Like the ultimate test is testing them on the network and seeing if they perform in real life. But one thing we’re working on as a sort of first filter there, it’s using LLMs to  based on audience data, to imitate that audience and to basically pre rank.

Ads before we even test them. It’s in the hundred thousand different ads you’re going to produce a month, it makes the top 10,000 emerge. And those you actually test on the platform and you actually see how they perform with humans. And so I think you can use LLMs and AI to basically.

Like the way I see ads and marketing, it’s that it’s sort of this huge, nearly infinite search space. And so you need to like LLMs help you explore a much wider part of this search space because you have like many more attempts at it, but you have to be looking in the right areas of the search space, and that goes back to the original point.

If you produce 10,000 awful ads a month, basically the totally wrong part of the search space, then you know you won’t perform. If you produce 10 ads a month in the right part of the search space, you might get a few which perform. If you’re producing a hundred  thousand ads a month in the right part of the search space, you know at least 10,000 of them are in the right part of the search space, then you are going to produce a lot of winners.

Shamanth: Yeah. Yeah you’re right. Search space is really infinite. And the way it sounds like you’re looking at using the LLMs is to narrow that down to what has high probability of performing. Yeah. And I think that will also let you explore even more ideas. Yeah. Because if you’re able to narrow down and rule out A, B, C, then you can just.

Patrick: Yeah.

Shamanth: And that’s fascinating. And what’s the actual form that you anticipate that this will take?

Patrick: So there have been a bunch of there have been a few academic papers on this topic. We’re working on a like internal version, but there are a few companies now which have launched, which are focusing on this.

And depending how fast they go we might just use their services. Basically our philosophy is if something doesn’t exist, we build it, but it does exist, and there’s a good company doing it, and we usually go with  that company as long as the pricing’s reasonable.

Shamanth: Yeah. Yeah. Interesting. And I’ve heard of a couple of services that call themselves synthetic focus groups. Haven’t seen the results, so I’d be curious to find out what you find.

Patrick: yeah.

It’s we, where we’re currently working on it, it’s going to be in production quite soon, hopefully.

But some of this sort of academic literature these LLMs would seem better than random. There, it’s not a perfect signal, but any signal better than random.

Shamanth: Yeah.

Certainly. Also to zoom out a bit, you scale your creative production, you scale the number of humans on your team. How do you, and obviously you’ve done this for a number of years, which is a very long time in AI land. So how do you think about moats in an AI world, in an, in a world where ads can be copied almost instantly?

Now, even apps can be copied almost instantly, right? I could drop like a monologue into Claude Code and have something very functional in a couple of hours, sometimes sooner. So what, how, what do you see as a moat? What do you see as what will lead  to sustainable advantage as you grow or as any company?

Patrick: Yeah, so you know, I think generally quite a lot of the classic moats still work in an age of AI. Human networks things like brands and maybe brands move like at one stage more soon, one more than ever. Those things still work. I think for our particular case, we’ve got quite a few apps.

We’re scaling six apps at the moment. Depends if you are category leader and a lot of our good apps are in that particular niche then obviously you can’t copy, like when you are category leader, you can’t copy other there’s no point copying the second or third apps, and that’s where creative velocity is very important because yeah, they can copy your ads, but you’ve run them first. We launch campaigns once or twice a week with new creators and as soon as they start working, they get a lot of spend.

So it would always be us. Because we’re creating this new sort of ad concept for this  particular app category or niche or whatever, we’re going to be the ones, who are showing that ad first to most users, and especially for most relevant users for, that particular app and ads. And you maintain your lead by continuing to innovate.

One thing we do, and I think that’s actually quite interesting for. other person who has an app, which is leader or, amongst the leaders in its category. It’s, we hardly ever look or we, and we never copy what our, direct competitors are doing. But we look at what’s happening in the market and then, completely different app categories and we’ll take inspiration from there sometimes.

And that’s one way where, and you know, it’s you need human taste usually to do that. We’ve tried to automate it and you do that sort of automation. But at the end of the day, human taste has worked best for us. And it’s by yeah, looking at completely different categories or even things which aren’t apps at all.

E-commerce things and taking inspiration from that for your own category.  That’s often worked very well for us.

Shamanth: Interesting. And that’s not easy and that’s not very common. And I’m curious if there is any example you can think of something you’ve looked to in a completely different category for inspiration.

Patrick: An example where we looked at a completely different category one, there are some ads where. We took inspiration from, I dunno if you’ve seen all those, they’re usually just viral videos from restaurants where it’s like a completely unrelated video.

Then there’s a really cool transition and it’s. And it’s like there’s a smooth but very sudden transition.

Shamanth: Okay.

Patrick: And we did that for quite a few of our apps for a while. And those videos performed for a while. And it was like. I dunno a guy, falling, like a guy falling from a roof of a house.

And he lands on our app store page kind of thing.

Shamanth: That’s crazy.

And was that generated with AI?

Patrick: Yes.

Shamanth: Yeah. Okay. That’s amazing.

Patrick: Like one of the things

I’ve really liked about our apps is that we  manage to generate CPIs.

…around 50 cents now, it’s maybe a tiny bit higher towards, more towards 60, 70 for like high quality traffic that we’re bidding on now. But it’s pretty low. So we get a lot of new users, which generate a lot of data, and we have some very data focused with their data centric, and I think that

Proprietary data is one of the things which will continue to yield moats, I would say generates not a moat, but it’s the thing which gives you the opportunity to build a moat is in the era of AI. It’s just velocity. Just moving very fast.

Shamanth: Yeah.

Patrick: Always being on the leading edge of technology. It’s not a moat per se, but it gives you the opportunity to play again and to actually develop.

Shamanth: Yeah. And as you also pointed out, having enough humans in the loop so that it’s not just velocity, it’s velocity in the right direction.

Patrick: Yeah.

Shamanth: Yeah, indeed.

Patrick: Yes. That’s very important.

Yeah. Moving fast in the wrong direction.

Shamanth: Indeed. That would be very good. Yeah.

Patrick: Yeah.

Shamanth: Excellent. Patrick,  this has been incredible. There’s a lot that I’ve learned, and I as I did when I read some of your writings, and I will actually link to those, it’s actually was the, and I would highly recommend that folks check out some of your writings.

Patrick: Thank you very much for having me. It has been a pleasure.

Shamanth: Wonderful. Wonderful.