Nerd Alert: The Bayesian Marketing Attribution Model

Summary

The Marketing Architects Podcast Episode Summary

Title: Nerd Alert: The Bayesian Marketing Attribution Model
Host: Marketing Architects
Release Date: May 15, 2025

---

Introduction

In the episode titled "Nerd Alert: The Bayesian Marketing Attribution Model," the Marketing Architects delve deep into the complexities of marketing attribution, exploring a cutting-edge Bayesian approach to better understand and quantify the impact of various marketing channels on consumer behavior. Hosted by Alina Jasper and co-hosted by Rob DeMars, this episode unpacks a significant study that promises to revolutionize how marketers attribute success across multiple channels.

---

Understanding Marketing Attribution

Alina Jasper introduces the core topic by highlighting the perennial challenge in marketing: accurately attributing customer conversions to the myriad of marketing channels employed. Traditional attribution models, often rule-based like first touch or last touch attribution, simplify this complex interplay but fall short in capturing nuanced dynamics.

“Attribution models... have been rule based. Those are things like first touch attribution or last touch attribution or heuristic offering oversimplified views of how advertising influences behavior.”
— Alina Jasper [02:00]

---

The Limitations of Traditional Models

Rob DeMars draws an analogy comparing simple attribution models to comfort food—familiar and easy but not necessarily the best choice for long-term success.

"It’s easy, it’s fast, it’s familiar, but it’s probably not the best for you... Last click just kind of sounds less scary than multi touch probabilistic modeling."
— Rob DeMars [04:12]

Alina concurs, emphasizing that reliance on easily measurable channels often sidelines more potent but harder-to-measure avenues, leading to suboptimal marketing investments.

"They’d rather invest in a channel that’s measurable with half the impact than a channel with double the impact where you can’t prove it."
— Alina Jasper [04:49]

---

Introducing Bayesian Marketing Attribution

The episode spotlight shines on the study titled "Bayesian Modeling of Marketing Attribution" by Ritwick Sina, David Arbor, and Ashled Manispooli. Alina simplifies the Bayesian approach, likening it to adjusting weather forecasts based on new evidence.

"Bayesian models... are like adjusting your weather prediction every time you look out the window."
— Alina Jasper [02:50]

This model stands out by accounting for:

• Direct Effects: Individual impact of each marketing channel.
• Time Decay: How the influence of each touchpoint diminishes over time.
• Interaction Effects: How different channels interact, either amplifying or suppressing responses.
• Customer Heterogeneity: Variations in baseline purchase likelihood among different customers.
• Uncertainty Quantification: Measuring the confidence in estimated parameters.

"The probability is they're driven by interpretable parameters like baseline, conversion likelihood, channel effectiveness, decay rate, and interaction strength."
— Alina Jasper [06:00]

---

Methodology of the Study

To validate their model, the researchers conducted two experiments:

1. Simulated Data: Utilizing synthetic customer journeys to test the model's ability to accurately recover known parameters.
2. Real-World Data: Analyzing a dataset from Adobe Analytics encompassing approximately 5 million users, 10,000 interaction events, and nine marketing channels within the travel and experience sector.

"They focused on a balanced subset of users who converted and those who didn’t."
— Alina Jasper [05:50]

---

Key Findings

1. Decay Parameters:

   • Channels with low direct effects showed flat, uninformative decay rates.
   • Channels with meaningful impacts had well-defined decay parameters, indicating how long their influence persisted.

   "There was a rapid decay of ad effects... their influence on consumers fades quickly."
   — Alina Jasper [07:00]

2. Interaction Effects:

   • High-frequency exposures led to negative interactions, highlighting ad fatigue where excessive impressions reduced conversion likelihood.

   "Overexposure reduced the likelihood of conversion."
   — Alina Jasper [08:15]

3. Attribution to Own and Offline Channels:

   • Unexpectedly strong attributions to proprietary channels and offline touchpoints, challenging the assumptions of traditional digital-centric models.

   "These results challenge common assumptions about what drives conversion and highlights limitations of models that ignore non-digital or indirectly measurable touch points."
   — Alina Jasper [09:30]

---

Implications for Marketers

Rob DeMars underscores the importance of model interpretability alongside predictive power, likening an uninterpretable model to a GPS system that communicates only in an unknown language.

"Predictive power without being able to interpret it right is like having GPS that only speaks Klingon."
— Rob DeMars [06:26]

Alina adds that interpretability fosters client trust and confidence, ensuring that stakeholders understand and buy into the attribution findings.

"If you're not walking through your methodology and it doesn't make sense, how are you going to feel confident bringing that number forward."
— Alina Jasper [07:00]

The Bayesian model's ability to provide a probabilistic and interpretable framework offers a more nuanced and accurate understanding of marketing effectiveness, enabling better strategic decisions.

---

Conclusion

The Marketing Architects highlight the Bayesian Modeling of Marketing Attribution as a transformative approach that addresses the shortcomings of traditional attribution methods. By incorporating time decay, interaction effects, and customer heterogeneity, and by ensuring model interpretability, this Bayesian framework offers marketers a more reliable and comprehensive tool for understanding and optimizing their marketing strategies.

"This is showing us traditional attribution methods... might be oversimplifying reality. And more advanced models like this could offer a deeper, accurate view of what's actually working in your marketing mix."
— Alina Jasper [10:00]

Rob humorously expresses enthusiasm for integrating advanced concepts like Bayesian models into everyday marketing conversations, signaling a shift towards more sophisticated measurement techniques.

"I'm definitely going to try to work the word Bayesian into a conversation today."
— Rob DeMars [11:22]

---

Key Takeaways

• Bayesian Models offer a probabilistic and interpretable approach to marketing attribution, overcoming the limitations of traditional rule-based models.
• Comprehensive Analysis: Accounting for time decay, interaction effects, and customer heterogeneity provides a more accurate attribution landscape.
• Enhanced Decision-Making: Understanding the nuanced impacts of various channels leads to better-informed marketing strategies and investments.
• Model Interpretability: Ensuring that attribution models are understandable builds trust and facilitates stakeholder buy-in.

For marketers striving to optimize their multi-channel strategies, embracing advanced models like Bayesian attribution can unlock deeper insights and drive more effective marketing outcomes.

---

This summary captures the essence of the "Nerd Alert: The Bayesian Marketing Attribution Model" episode, providing a comprehensive overview of the discussions and insights shared by Alina Jasper and Rob DeMars.

Transcript

[Rob DeMars]

Nerd alert. Learning is important, right?

[Alina Jasper]

Yes, exactly. What a bunch of nerds.

[Rob DeMars]

Nerd alert.

[Alina Jasper]

Marketing Architects. Hello and welcome to the Marketing Architects, a research first podcast dedicated to answering your toughest marketing questions. I'm Alina Jasper. I run the marketing team here at Marketing Architects. And I'm joined by my co host Rob DeMars, the chief product architect of misfits and machines.

[Rob DeMars]

Hi Alina.

[Alina Jasper]

Hello. We are back with your weekly Nerd Alert. Every week I'll take a deep dive into academic marketing research and translate its complex ideas into simple, understandable language for Rob and of course for all of you. Are you ready to nerd out, Rob?

[Rob DeMars]

Oh man, if marketing effectiveness nerdiness were a Marvel movie, I'm like ready for the post credits chaos to begin. So bring it. Bring it, Thanos.

[Alina Jasper]

I'm Thanos in this situation.

[Rob DeMars]

Yeah.

[Alina Jasper]

All right, let's get into it. As always, we'll link the research we cover in the episode notes. This week I read a study titled Bayesian Modeling of Marketing Attribution. This is by Ritwick Sina and David Arbor from Adobe Research and Ashled Manispooli from New York University. Before we get into what I'm sure is Rob's favorite topic, marketing attribution theory. Rob, I wanted to ask you this. Do you think marketers are more afraid of being wrong in their marketing measurement or of admitting that they're uncertain of it?

[Rob DeMars]

Admitting they're uncertain? You know, I nobody likes to look dumb. Unfortunately, I can't really help it for myself. But in marketing I know that nobody likes to have A thousand armchair CMOs chiming in because you feel uncertain about your opinion. So I think that's a showing. Weakness is just an invitation to have other people throw down. So I would say feeling uncertain.

[Alina Jasper]

Yeah, I agree with you. I think you can even see that in channels that marketers prefer to invest in. And it makes sense because if marketers have to, if they're held to such rigorous standards of dollar and what comes back, they end up probably investing more in certain channels, like performance channels, digital, where you can clearly tie that in. I think you're right. I think a lot of marketers maybe wouldn't want to admit this, but they'd rather invest in a channel that's measurable with half the impact than a channel with double the impact where you can't prove it. Because if you can't prove what happened, it's going to be hard to get buy in to move forward with anything. So that's one of the hard things is Some of the best marketing channels are not easy to measure, but they're the most powerful channels, so sometimes they get missed out on. But today we are going to talk about a study that attempts to bring a more rigorous, interpretable and probabilistically grounded framework to one of marketing's biggest challenges, which is multi channel attribution, that is.

[Rob DeMars]

Some multi channel vowels you got going on there. Syllables, multi channel syllables.

[Alina Jasper]

We're gonna go through them. So what this study does is it proposes a structured method for estimating the influence of different marketing channels on customer conversion outcomes. And they use this Bayesian approach, which I didn't know what that meant, but I guess it's a way of updating what you believe based on new evidence. So Bayesian models, they don't give just one fixed answer. It's kind of like adjusting your weather prediction every time you look out the window. So you start with the forecast, but if you see clouds, you revise your expectations. But back to the study. As most of us know, attribution models, they attempt to quantify how different marketing efforts, whether it's email, display, search, social, contribute to a consumer's eventual decision to purchase your product or service. And traditionally, models have been rule based. Those are things like first touch attribution or last touch attribution or heuristic offering oversimplified views of how advertising influences behavior. And we've talked about those a lot of on the show. Some of the downsides of those more recent data driven methods have made progress, but many still fail to incorporate important dynamics such as the time decay of ad effects, interaction effects between multiple ad exposures and heterogeneity, and consumer responsiveness. But Rob, before we go further, what's your take on why the industry continues to rely so heavily on simple rule based attribution despite all of this technological advancement we've had?

[Rob DeMars]

I think it's a lot like simple models are a lot like comfort food. It's easy, it's fast, it's familiar, but it's probably not the best for you. There's nothing I like more than curling up with a family sized box of Cheez Its, but that's probably not the best thing in the world. But it's simple and easy. Last click just kind of sounds less scary than multi touch probabilistic modeling.

[Alina Jasper]

Agreed. Like if, especially if you're first starting out or you haven't figured out like the complications, even invested in more complicated measurement models, it's like last click attribution seems like it's the easiest thing to measure. It's simple, it makes sense. But the problem with it is we talked about this on a few nerd alerts ago. They found, like Last touch attribution, it's not only inaccurate, but it leads to worse outcomes because you make decisions based on it that aren't accurate. So it seems simple and easy, but over time, it's leading to worse outcomes.

[Rob DeMars]

Like binging on a box of Cheez Its.

[Alina Jasper]

Oh, yeah, Cheez Its are good though. But you're right. So this paper, what it does is it presents this Bayesian likelihood model, and it incorporates a range of different marketing phenomena that includes the direct effects of individual ads and channels, the decay over time of those effects. We mentioned that earlier. That's capturing how the influence of an impression diminishes over time interaction effects between ads. So, for example, does the exposure to two channels amplify or suppress response? And customer heterogeneity, which allows for differences in baseline purchase likelihoods and finally, uncertainty quantification for all estimated parameters, which. I'll tell you what, I'm not sure what that sentence means, but it sounds important. So this model, it outputs a probability of purchase at each time point based on a customer's prior ad exposures, the channels involved, and their timing. So that's kind of. They take all that into account, and the probability is they're driven by interpretable parameters like baseline, conversion likelihood, channel effectiveness, decay rate, and interaction strength. And that word interpretable is important. When they say that, what they mean is you can easily understand what the model is doing and why it made a certain decision. So, Rob, how important do you think it is for attribution models to be interpretable and not just predictive?

[Rob DeMars]

Well, first I have to say interpretable.

[Alina Jasper]

I know I messed up that many times.

[Rob DeMars]

That's intimidating. You know, I have to imagine predictive power without being able to interpret it right is like having GPS that only speaks Klingon. It's cool. Like, that's really cool tech. But I. It's not really valuable in terms of helping with navigation. So I think it's really important to be able to be able to leverage predictive power, but definitely be able to understand what it means means.

[Alina Jasper]

Agreed. And I think we feel that with clients too, like, when we're walking through how a TV campaign performed, you could just put up, hey, here's your roas. But if you're not. If they don't understand how you got there, if you're not walking through your methodology and it doesn't make sense, how are you going to feel confident bringing that number forward. So I think yeah, it's super important. It's not enough just to have a predictive outcome like how did you actually get there? So let's talk a little bit about the methodology. So to validate this model, the researchers conducted two experiments. The first was simulated data. So they used synthetic customer journeys and applied the model to see if it could recover the original known parameters. And those results showed the model is able to find the right answers accurately. They had clear and confident estimates and very little error. And then in the second, they also used real world data. They used a large scale data set from Adobe analytics tied to a travel and experience brand. This DataSet included approximately 5 million users, 10,000 interaction events and nine marketing channels. What they did was they focused on a balanced subset of users who converted and those who didn't. And they used the model to estimate attribution values for each channel. And here's what they found. The model produced posterior distributions not just for attribution values, but also for the underlying dynamics of ad exposure. So some different important patterns emerged. When a channel had a low direct effect, the associated decay parameter often showed a flat uninformative distribution, which was expected. But when the channel had a meaningful impact, the decay parameter became well defined, which allowed analysts to understand how long that marketing influence lasted. There was a rapid decay of ad effects. So many ad impressions, particularly from display and search, showed extremely short half lives. This indicated that their influence on consumers fades quickly. And the model it uncovered negative interaction effects when consumers received a high number of ad impressions within a short window, AKA ad fatigue. This is something that we see with our clients too. So rather than reinforcing purchase intent, which you're often told over frequency does, the overexposure reduced the likelihood of conversion. So that suggested you're going to have diminishing returns or even adverse effects if you have a heavy frequency. The model also found unexpectedly strong attribution to own channels and offline touch points. These results challenge common assumptions about what drives conversion and highlights limitations of models that ignore non digital or indirectly measurable touch points. Which comes back to what we were talking about earlier, the channels with last touch attribution. You're missing all of these touch points if you only rely on that. Ultimately, what this study does, it's pretty complicated and to be honest, I had a hard time. I had a hard time recapping it, but I thought it was really interesting. It offers a more robust approach to attribution by leveraging Bayesian inference. The authors they're able to quantify the magnitude, the duration and the interaction of different ad effects. And they can also model how it varies between different customers and incorporate this degree of uncertainty that whether we like it or not, we're always going to have in our marketing attribution. So I don't think that you have to go and ask for a Bayesian model. Maybe learn how to say it before you do that. But I think there's a good takeaway here, which is just that once again, this is showing us traditional attribution methods. They might be oversimplifying reality. And more advanced models like this, they could offer a more deeper, accurate view of what's actually working in your marketing mix.

[Rob DeMars]

I am really excited for the robgpt.

[Alina Jasper]

I was about to say.

[Rob DeMars]

I was just going to say I've never wanted a Rob GPT more than now.

[Alina Jasper]

Well, let's see how it does. This study is like giving credit for a winning dish to the whole kitchen, not just the last person who sprinkled salt on top. Traditional models might say the last step caused the outcome, but this one looks at every ingredient, how much they combined, how long their flavors lasted, and whether adding too much made it worse. It's a smarter recipe for finding out what actually made the meal or in this case, the sale work.

[Rob DeMars]

That was really good. That was really good. I mean, you went full thanos on this one, Alina.

[Alina Jasper]

I confused my heavy.

[Rob DeMars]

That was some heavy nerdery.

[Alina Jasper]

I think it's interesting if anyone's in this sort of attribution challenge, it could be a good thing to look into or it's just another maybe inspiration or way to make the case for moving on from simplistic attribution methods. I mean, it could be a really big unlock for your brand and your company.

[Rob DeMars]

Absolutely.

[Alina Jasper]

If you can move away from that.

[Rob DeMars]

And I'm definitely going to try to work the word Bayesian into a conversation today. So thank you for that.

[Alina Jasper]

You're welcome. Good luck.

[Rob DeMars]

While I'm eating Cheez.

[Alina Jasper]

Its nice. That's it for this episode of the Marketing Architects. We'd like to thank Taylor De Los Reyes for producing the show. You can connect with us on LinkedIn and if you like the podcast, please leave us a review. Now go forth and build great marketing. Nobody's here to produce us, so hopefully this goes well. Nothing bad happens.

[Rob DeMars]

All right.

[Alina Jasper]

But we should be fine marketing architects.