Marketing Mix Modelling: What It Is, Why It Matters & Where AI Fits In

James Lawrence: Welcome back to the Smarter Marketer Podcast. I’m here today with Henry Innis from Mutinex. Henry, welcome to the pod.

Henry Innis: Nice to be here, James. Thanks for having me.

James Lawrence: Excited for the chat today. I’ll just do a quick introduction for listeners who might not have heard of you or the organisation before.

Henry is the CEO and co-founder of Mutinex, a business designed to disrupt marketing measurement. Mutinex has gone from being a small Australian analytics consultancy to a global SaaS company now valued at over $130 million.

You’ve grown rapidly by productising marketing mix modelling into a platform called Growth OS. Today, you’re working with major brands like Samsung, Asahi, ING, Domino’s, Kia and Dulux to help businesses make better decisions across billions of dollars of marketing spend.

Henry, you moved across to the States in 2024, I believe, to lead the push of Mutinex into the US market, which I guess is the next chapter of the business. Welcome to the pod.

Henry Innis: Thanks, James. I think I already led the push, really. The US business is now doing pretty well by itself.

We work with Hershey’s, Whirlpool and Olaplex in those markets. It is definitely a big and growing part of our business.

James Lawrence: It’s awesome. It’s really cool.

Before jumping into the conversation around Mutinex, where the product is now, use cases and all those types of things, quickly, prior to starting the business, what were you doing? What led to the start of the business back in the day?

Henry Innis: I was in the data arm of Y&R at the time, VML before that. This was before they were all smashed together and merged in the mega-merges that kind of plagued the 2020s.

We were really focused on trying to figure out ROI on creative campaigns and what our best allocation of resources was to maximise campaigns. So, not too dissimilar.

My co-founder and I planned to found our own business inside WPP. That plan got kiboshed with some changes to leadership and things like that, so we felt it was best to go it alone.

We actually started the business in late 2019, I think, not 2017. You could not have picked a worse time to start a business, to be frank with you. We walked straight into COVID, and off the back of COVID, walked straight into high inflation. It has not been a particularly easy few years, I would say.

But the objective of the business when we started it was that we had spoken to so many customers and clients at the time, and every single one of them shared the same pain point, which was, “How do I allocate my capital well and effectively for growth? Do I put it into pricing? Do I put it into media? How do I make those trade-off decisions?”

Their only options at the time were one of two things. Either they made the decision on incomplete information because their data was not well organised and it was very hard to do anything with it, or they made decisions off very untimely information.

Once they had the data organised to run models at scale, this was pre-cloud and stuff like that, it was a very difficult and challenging task. There were not many people who were able to do it. There were a number of consulting firms that ran these models over four, five or six-month build periods.

Of course, if you have ever been in media, if someone tells you, “We have got an insight from six months ago and we would love to tell you how that insight applies to the media market today,” the person planning the media will look at them and go, “That is some good archaeology you have done there, but let us get back to the real job.”

That was at the centre of why we founded the business. We felt that timely insights that encapsulated the total business’s growth were important.

I do not really think we knew the shape of the business and what the business would become back then. We probably thought it would be more of an ROI consultancy. But in 2021, we decided to make it a software as a service business.

There were a few reasons for that. One, COVID gave us space to do it. Two, we felt the business could give better service, better answers and be used more consistently if there was much lower latency for a client to use the actual product.

One of the problems you have with products is if they have to interface through the people in the business to interface with the data. If someone has to pick up the phone to me and say, “Henry,” and then I go and interrogate the data, ultimately I am getting that data through a lens of opacity. There is also a lag to that, which is just very human.

So, we thought removing that would make a lot of sense for what we wanted to do.

James Lawrence: And just for those listeners who might not necessarily be across this space, can you explain marketing mix modelling? What it is, where it has come from and how it works?

Henry Innis: Measurement really came into vogue, let’s call it around 2008, as Meta and Google really started to take hold. They had a pretty novel idea. The idea of advertising ROI was around before this, but it really took hold in the zeitgeist at that moment, which is that when you spend a dollar in advertising, you get some form of sales back, and you can understand that a sale was influenced by a certain piece of advertising or a certain decision.

That really came up through the digital platforms.

For a long time, since about the 1960s, very advanced and sophisticated businesses have done that another way by building quite complex regression models.

What the regression model basically means is that when they see a sales lift in a week of sales, they look for whether any other variables were going up around that sales lift, and whether those variables go up consistently every time they see sales going up, at its simplest form.

Market mix models have become a lot more advanced because those regressions have become a lot more advanced. We know a lot more about marketing today.

For example, we know that TV has a halo effect on search, which then helps lift the number of people searching for the ad, which then lifts the performance of search engine marketing.

At its simplest, the regressions before were quite simple: show me what variables lifted before. Now, as we have gained a larger and larger body of evidence, driven by a lot of institutions and things like that, we are able to start to say, “Actually, we know that marketing works like this. Now tell me what is happening in this business based on these rules that we know.”

That is what you would call a Bayesian market mix model. It is a very fancy, technical word for it. But in effect, what we are doing is trying to look at when something happens, what else did we see happening around it?

It has become in vogue because, in particular, two things have happened. One, a lot of the attribution via pixels, people have realised that is not quite how marketing works. There is a lot of work that happens before that last click to make that click happen.

The second thing is that we used to have a lot of attribution models that tracked people through cookies and things like that. There has been a lot more signal loss.

I think coupled with that, you have a much tougher economic climate. So you have these three forces. One is, my normal tool for understanding what everybody saw has been massively degraded in its effectiveness. Two, I have got a much better understanding of marketing, and that the last click is not where the click came from. Three, I have got a lot more CFOs asking, “What drives our business? What drains our business? What did pricing do? What did media do? How do they work together?”

Because they are confronting more and more tough economic times, that has led to an explosion in software as a service platforms. We are not the only ones. There are other groups like Recast, who are pretty fantastic. The Cassandra guys in the EU are also another really good example of a player that works with much smaller businesses than us. So, there are lots of examples and permutations.

James Lawrence: Yeah, cool. That is a really good overview.

Just a clarification, because you will hear of media mix modelling and market mix modelling. Same thing, different thing? Is it just a semantic difference?

Henry Innis: No, it is quite a big difference.

Media mix modelling refers to, “I just put the media spend in and the sales in.” The problem with that is I can observe a lift in sales, but I may have run a 30 percent discount. In a media mix model, it will only see the media lifting, so it will only attribute that sales lift to the media.

In a market mix, you are bringing in all the variables that would affect the business. Media mix suffers from something called omitted variable bias, which basically means, in English, we have missed a bunch of things. As a result, we are overstating the impact of the thing that is currently in the model.

I think the reason people do media mix is because it is easier. To collect the data on complex pricing and stuff like that, and to understand how those things work, is a much more complex beast to go into market mix than media mix.

For many businesses, they are okay with just getting the directional read off it.

James Lawrence: And the downside being you might be getting more correlation, not causation.

Henry Innis: Yeah, that is right. I think that is correct. Most of the time, you will see correlations happening.

There are two ways to get to causation. The gold standard is experimentation. If I want to minimise the risk of overstating a variable through correlation, once you see all correlations, you have less of a chance of saying this thing spiked and it was just spiking because something else was spiking. You are more likely to see that as a model as a result.

So I think that is why people focus on it.

James Lawrence: Yeah. Can you maybe go into some specifics around what Mutinex actually does? In terms of your conversation before, one of the big problems previously was time, right? The fact that you are looking back six months, a year, and you are looking at old data.

I suspect that part of the solution is getting data coming in closer to real time. I would be so interested in the types of factors that the system is looking at. Is it weather? Is it all kinds of different things? Just to give listeners an idea as to how the platform works.

Henry Innis: So it is weather, competitive spend, market factors like CPI and fuel increases. Australian horror movies are currently an event in many models.

Then there are other factors, like paid media spend, pricing, discounts, offers, trade media and brand equity measures. All of those things generally get brought in. So, they are fairly large beasts.

I think what we have done quite successfully is make it relatively easy for someone to put unstructured data in.

If you think about how I would do a data project six years ago, I would have formatted everything super neatly and then sent it off to the modellers. We do not really need to do that anymore.

We can start to take a lot more unstructured data. The CSV report that you are looking at on a week-to-week basis can be hoovered up by the systems. That has been a big shift, and we invested a lot in that in 2023. Our business did not really hit an inflection point until that point.

James Lawrence: Yeah. Okay. Where are the use cases particularly strong?

You mentioned the European competitor being better with smaller businesses. I presume that there is some requirement here for big enough data sets, big enough spends, enough online and offline data to make it worthwhile?

Henry Innis: It is very hard to estimate a brand new business, for example, with a model. I think that is an incredibly challenging thing to do.

You need some depth in data. You need some understanding of what works and what does not. For us, the juice has to be worth the squeeze, right?

I always say to people, if you are running a $10,000 media budget, it is probably not worth investing in a media mix model or a market mix model. You are much better off investing in good incrementality testing, or just using first-click attribution or something like that.

There are different ways to approach the problem based on scale. As with all things, it is about how complex your environment is. More complexity generally requires more modelling.

And it is about your return on effort, because a lot of these things are really effort based. The SaaS platforms bring down the raw cost, but there is still some degree of effort to them.

James Lawrence: Yeah. And any difference between B2B and B2C?

Henry Innis: Not in my experience, no. Although, having said that, B2B enterprise is a very hard thing to model.

You have got to model, say, enterprise leads as opposed to enterprise sales, as a good example. Generally, getting close to the marketing action is what you need to do.

James Lawrence: Yeah, okay. And what impact has AI had? I presume there has been a lot of work going on with AI in the last couple of years. What is the product plan there?

Henry Innis: I think you have got to break AI down into what it is trying to do.

Pre-artificial intelligence, a lot of quite high-value tasks were priced very highly. What AI is allowing us to do is take very high-value tasks and diminish the value of those tasks immensely, while allowing you to scale the repetition of those tasks immensely.

A really good example of that is production assets. For me to do a photo shoot five years ago was quite a lot of work. To my mind, it is much easier now for me to generate 500 shoots today. That has brought the value of that task down considerably.

So I think our supply is much, much greater in terms of intellectual capital. That is bringing the cost down dramatically, as you would expect. But I think the explosion on the other side is that there is a lot more demand.

My suspicion is that demand for services, production, content and all of these different things that everyone says are going to be destroyed by AI is going to massively explode.

Now, how does that apply to our business? Market mix modelling used to be a very costly exercise, and therefore we would not go hyper-granular on the models because there was a pretty large cost to that.

For me to run 500 models for a major US brand would probably cost me, even if I had a fair degree of automation, I would still have probably five or so people working at that scale to maintain and run that.

Now I can do that with a fraction of the people, and that allows the average business to scale up into a lot more MMMs, with a lot more granularity than they otherwise would.

The second thing is that I would have to structure a lot of the data. So, intuitively, I would have a requirement to have a lot of data engineering or data maturity for these sorts of businesses to make this work. Again, AI has made that very high-value task very cheap.

So I can now have lots of brands engineered ready for these models without necessarily needing a lot of money behind them.

The third aspect is the insight layer. Usually, I would have to have a bunch of very smart people to interpret these models and apply them in clever ways. Now AI has made it very easy to read these models, get value out of them, and ask questions through a translation layer.

All of those things contribute to turning the sector into something very different. I think the sector is really going to reinvent itself, and I think most sectors are like this. There is a wave of services-based sectors that are effectively flipping from being high-value, low-volume to high-volume, low-value businesses.

Structurally, that is how I think about AI.

Practically, we are constantly launching AI agents and agentic workflows into our product to begin to take on more and more of the outcomes and the outcomes work that our clients do.

I foresee, in the future, us being able to provision growth as a service back to the market. But I think we are some way off that in the product roadmap. That is where we want to be in the medium term.

James Lawrence: If you look back at the product two years ago to now, is it completely different?

Henry Innis: Completely different.

James Lawrence: Is it just transformative in terms of the actual results you are driving for clients?

Henry Innis: Yeah. Let us just take an average query through our AI interface in our agentic workforce.

I reckon the average query would have taken our marketing science people, let’s call it, 10 hours to answer. Last week we did 1,000 queries. Some weeks we will do 3,000. It is all quite variable, but call it 10,000 hours.

For me to be able to service that previously, I would probably need about 260 people. So the volume and velocity at which people can get value out of these things is greatly improved.

I do think AI is raising the intellectual floor. We talk about AI hallucinations a lot and things like that, but it is much harder to be really stupid in the world of AI.

There is something to be said about the intellectual floor significantly raising. I never see a really bad deck or a really badly architected email. That averaging out of really mediocre communication and work has been quite dramatic in the past six months. We have seen that in our own product and how people interact with our product.

James Lawrence: Yeah, I have not thought about that. Obviously, humans make mistakes, so the whole hallucination thing is analogous to that. But the concept of lifting up the bottom end, I have not thought of it from that viewpoint. Interesting.

Henry Innis: When was the last time you saw a really poorly structured email that read exceptionally poorly? It just does not really happen anymore.

Most people who are really bad at this stuff, or really bad at communication, are running it through ChatGPT now. So the intellectual floor has been significantly raised.

I have not seen a bad deck in months because people are just getting really good at it.

James Lawrence: How are you finding the business culture in America? Particularly from a marketing viewpoint, what are the similarities in the outlook there?

But probably more interesting for Australian marketers, what are the differences? How does marketing work operationally in America?

Henry Innis: Let’s start with the similarities. The similarities are the language. That is about it.

The Americans have a very high risk-taking culture. They probably do not bias towards research and being thoughtful about things. They bias towards action and testing things. They have a much higher acceptance of failure, correspondingly. They do not mind failing at something as long as they do it quickly.

I think the Americans are impressive there. They have been impressive for some time there.

I think where the Americans are going, and how they also operate, is that they operate far less on relationships than Australians do. An Australian will rely on relationship first, which allows you, when you are going through a bad time, to have a solid way to work through things. The Australian instinct is to try to work through things together.

I think the American instinct is just to move on. It has its strengths and weaknesses. You do not build as long-term or meaningful partnerships there as you do in Australia. But on the flip side, you do get a lot more signal from the market if you are not doing things well there.

I know that because someone fired me quite willingly in my first year there, and they did it quite quickly and probably much more aggressively than I would have thought.

James Lawrence: Interesting observations. I found that interesting.

How is the outlook over there from a media and marketing viewpoint?

Henry Innis: You have two camps. You have the international businesses, and the international businesses with international supply chains are probably a little bit more spooked at the moment.

Then you have the domestic businesses. I think it is very important to understand, in America, that they are a net exporter of oil. So the Strait of Hormuz impacts their domestic situation a bit, but it is less existential than, say, Australia. It is less existential than, say, Europe. That is an important piece of geopolitical context.

Most marketers there are always planning to grow and are always looking for ways to grow. That is their culture.

They do not have many businesses that sit at the 50 or 60 percent market shares that Australia has. Their landscape tends to be more fragmented, and that makes the hyper-competition for growth a lot more important to them.

So, they have some differences and changes like that.

James Lawrence: Interesting.

You guys published a really good white paper this year on whether your organisation is measurement ready in 2026. We will include a link to that in the show notes.

I was reading through it while preparing for the conversation today. There is a really nice phrase, which is, “Marketing has changed faster than measurement.” I think that is a nice framing. It would be good if you could elaborate on that.

Henry Innis: Yeah. Look at the volume of variables and the disciplines that we have to deal with as marketers.

Marketing has gone from a profession where we ran one campaign on six channels once a year to a continuous smorgasbord of campaigns run across a plethora of channels, with a plethora of external forces also buffeting us, because everyone is changing price every week as well.

The ability for somebody to change and respond to the market has dramatically increased. It has never been easier to buy ads, change creative, change price, run discounts and run offers. That means the market is hyper-reactive. People are moving and reacting to what they see in the market and what signals they get back from the market.

I think most measurement systems just have not kept pace with that.

If I look at most organisations doing market mix modelling, they do it once a year. They might do incrementality testing once or twice a year. That means I am getting signal at a much slower pace.

It would be a little bit like me driving with the Yellow Pages when I could buy a GPS. On one, I am getting course corrections and I know which way I am turning on a street-to-street basis. On the other, I am having to pull it out once every month or so.

James Lawrence: Good. And how does marketing mix modelling deal with brand versus performance?

Henry Innis: I think they are separate.

I will myth-bust one thing. The idea that a brand ad does not sell anything today but magically sells stuff in two years is bullshit. I do not think anybody sees that.

I think the effect of a brand ad can be realised over two years.

James Lawrence: Yeah.

Henry Innis: You get the effect, and then it bleeds and has a very long tail, effectively. But this idea that you should never see a sales response from a brand ad, I think, is not correct.

My general feeling is that we see this show up in the models in two ways. One is that performance media tends to be demand capture, so it tends to be capturing demand. It tends to have high synergy with the upper-funnel channels.

The second thing is we tend to see that brand media has a longer ad stock or memory effect. Ad stock means that I saw that ad three weeks ago and it influenced me to walk into the car dealership, that type of thing.

That is where you see the differences.

Generally, a brand versus performance channel would be segmented. If I am buying Meta on a reach and frequency basis, I would separate that from Meta on a performance marketing basis, for example.

James Lawrence: Interesting. I think Ritson says something along the lines of, good long-form marketing can have a good short-term impact, but rarely does short-term marketing have a long-term impact.

Henry Innis: Yeah. I think that is right, and I think that is a good articulation of it.

I think System1 has a pretty similar view as well, from memory, where they find that ads that are good and might have high emotional resonance also deliver high sales spikes or something along those lines.

James Lawrence: Interesting.

Just in terms of wrapping up, if we fast-forward two years into the future, what do you think the best-performing marketing teams will be doing in terms of measurement versus the ones that are the laggards?

Henry Innis: I think the laggards, firstly, will be optimising per channel.

Marketers who operate in silos are going to increasingly be caught out. To some degree, your measurement system dictates how your behaviours work.

If your measurement systems are siloed and channel by channel, inherently your execution will be siloed and channel by channel.

I think marketers that bring everything together and treat themselves like capital allocators, or like hedge funds, are going to be the ones who get really smart across the ecosystem. They are able to distribute capital very efficiently, and I think they are going to be the ultimate winners in all this.

James Lawrence: Nice one.

Henry, thanks so much for coming onto the pod, mate. Good luck with the growth, both in Australia but also globally and into the States.

Henry Innis: Not a problem. Thanks for having me, James.

James Lawrence: Thanks, mate.