Home Blog Marketing Mix Modeli ...

Marketing Mix Modeling, the future of measurement

Filippo Trocca

Over the past few weeks, there has been a lot of talk about data management, privacy, web cookieless: but what does actually change for advertisers? And how will they map the user journeys if data is decreasing while channels are increasingly fragmented and hybrid?

In this scenario, Marketing Mix Modeling could truly help marketers and companies.

 

Once upon a time, there was the Multi-touch Attribution Model 

As we have previously explained, the Next Advertising Era will let advertisers accurately measure their campaign performance only when certain thresholds are reached.

Both Apple’s SKAdnetwork and Google’s Attribution API introduce the concept of “hierarchy”, which unlocks a more detailed tracking as the number of conversions increases, in order to prevent the recognition of individual users.

Obviously, these frameworks only affect the measurement capabilities of advertising platforms, not of Digital Analytics tools such as Google Analytics, Adobe Analytics, Matomo, Piwik Pro, etc., which will continue to measure the same way they do, even if they will still not be able to measure the post impression conversions.

Therefore, multi-touch attribution will work only post click and the user identification will be delegated to the digital analytics platform (Google Analytics 4 promises to recognize the user using Google data when Google Signals is active) or the CMS itself.

In other words: the holy Grail of multi-touch attribution is more and more distant, quite impossible to reach.

As a matter of fact, the multi-touch attribution allows the platform to follow the user in all its touchpoints, so that it can piece together the path that led the user to the conversion.

By analyzing all user-generated paths, the model will be able to determine how each touchpoint – and therefore each marketing channel – counted in determining the conversion, and provide advertisers with a specific analysis of how successful the marketing planning was, split by channels.

When it works properly, a multi-touch attribution model recognizes the user cross device, in order to be able to truly retrace the same paths.

Here’s an example to make this clearer.

Mark sees an Ad on TikTok while browsing his feed, then he sees an Ad from the same company on Instagram; intrigued by the Ads, he searches for the product name on Google (he does all these activities on his smartphone) and finds a Google Shopping Ad, but he visits the company website by clicking on an organic result, he’s interested in the product so he subscribes to the newsletter to receive offers and more information.

Two weeks pass and Mark receives an email from the company, reminding him that the new version of the product he liked has been released, in new colors as well, so he comes back to the site by clicking on the newsletter link to check out the new collection.

The next day, Mark sees the company’s advertising campaign on the bus. At the office, he visits the company’s website again by searching for the name on Google, but this time from his laptop, because he wants to see the images of the product on a larger computer screen.

Three weeks later he decides to buy that very product as a birthday present. Sitting on the sofa, he purchases the product from his tablet, after searching for the brand name on Google and clicking on an organic result.

 

Let’s make a summary:

  • Mark used 3 different devices: smartphone, laptop, tablet
  • He interacted with 7 different channels: TikTok, Instagram, Google Ads, newsletter, OOH, Google Organic
  • Conversion time: about 6 weeks
  • Type of interaction: Impressions, click, session, OOH

 

Now the question is: do we have a tool that could follow Mark’s complete path, even through 3rd Party cookies? The answer is no, we don’t, mainly because the offline component would be impossible to piece together, but also because the impressions of TikTok or Instagram cannot be reconciled in a single tool, even if they are both digital channels.

This simple case study is basically what we all experience as users on a daily basis, yet it shows the complexity of marketing when a company tmixes online and offline advertising, and it highlights how even the digital tools cannot be correctly mapped and attributed as touchpoints, because there is no access to that type of data (in this case, impressions from social networks).

An attribution model created on a Digital Analytics tool such as Google Analytics or Adobe Analytics will tend to consider the Advertising activity on TikTok and Instagram useless, but that is simply because it does not record their ad views.

Therefore, even today the multi-touch model must face strong limits and in the future it will be even worse, and this is why we should really start thinking about new methodologies.

 

Marketing Mix Model: what it is and why it is coming back 

Just like Artificial Intelligence, whose concept of a neural network was already conceived in the forties, the mathematical models born in the past still helps us today when it comes to attribution, now boosted by a greater computing power that was not available back in the days.

In the past, in order to calculate how effective an offline marketing investment was (TV, radio, newspapers, OOH, events, etc.) you could rely on the so-called econometric models, albeit with some flaws:

  • long implementation and optimization times, as well as high effort, since the data used to be manually collected and applied to campaigns that could start even after 6 months;
  • inaccuracy due to the analyst biases, therefore beliefs, errors and experience of the individual operator largely used to determine the success of each channel in the marketing mix, being the models created manually.

Actually, this is the way they still work. 

To overcome these problems, the process needed to be automated.

This was the challenge for the first Marketing Mix Model open source project: to start from the statistical basis developed in the past, apply the analytical skills of modern computing, which allow you to analyze millions of models all together in order to identify the most relevant, leaving the final decision to the analyst, even if greatly supported by technology. In addition to this, we should also consider the forecasting capacity developed by artificial intelligence in the analysis of numerical series.

This is the simplified description of Meta’s “Robyn” project: the most advanced open source marketing mix model on the market today.

 

How does a Marketing Mix Model work?

To process its results, a MMM does not need to follow the user, actually it is privacy-first by design, because it only uses aggregate data, such as clicks, impressions, sessions, conversions, costs and revenues.

This means that you can feed the algorithm with the performance of the different marketing channels, using clicks and impressions as metrics, and collect cost information and results in terms of revenue: this is the kind of information that will always be available, because it will never be influenced or limited by new privacy laws or institutional decisions. You can keep collecting these metrics  without the user permission, since they cannot let you identify or profile users, in any way.

Furthermore, a marketing mix model needs context information to understand the macro-economic trend: for example, Google Trends data are perfect for this purpose, while you can integrate data relating to competition and demand for a specific product/service.

 

What do you get from a Marketing Mix Model?

In the case of Robyn, you get a model based on historical data, in order to evaluate whether and to what extent the individual channels of the marketing mix contributed to generating that business result, but also what the optimal budget distribution would have been, since it calculates the saturation achieved in the various channels. From a future perspective, this is useful because it suggests when you should increase the investment and expect a better result, or on the contrary, if you cannot expect any improvement, even increasing the budget.

Knowing that many various industry experts recognized Robyn as the most advanced Marketing Mix Model project, as recognized by, Google didn’t stay and watch. 

It is obviously working on its own model and has just published the first open source libraries, called Lightweight.

Currently, the Mountain View framework cannot compete with Robyn either for functionality or for analytical skills, since it focuses only on digital marketing channels, but Google is working hard on this project, investing relevant resources.

The Marketing Mix Models are not simple tools to use, it will be difficult for Google or Facebook or others to develop an all-encompassing attribution tool that provides companies with easy-to-interpret turnkey results. 

At the moment, they are relatively simple frameworks in the data ingestion phase, but an expert analyst is needed when you have to interpret the different models provided in order to choose the best one, and then continue to work on them over the months and interpret the results, adjusting the shot when needed.

Of course, Marketing Mix Models also have their weaknesses: one above all, their lack of flexibility. Being based on historical data, they are able to evaluate future results only when no mutation occurs. 

For example, analyzing our Social campaigns, the model could warn you when a saturation point is reached: however, this does not refer to the channel, but to the audiences you are using at that particular moment. Hence, the MMM will not be able to tell us if you will get better results adding new audiences. Nevertheless, a marketing mix model that updates its analysis on a weekly basis truly helps, because over time the model will learn and give you feedback on how the changes have positively or negatively impacted your strategy.

MMMs are complex tools, however, they are becoming more and more manageable, within the reach of PMIs and not just Corporates. My advice is to start getting information and invest in experimenting with these new measurement models, which could be of great help in the current economic situation, where you must foresee and control future scenarios in order to survive.

Interest in Marketing Mix Models is growing: Uber has also released a new Bayesian Time Varying Coefficient Model feature in its library for time series analysis, Orbyt, which can be used for MMMs.

Together with incremental analysis, the marketing MIx models are the hot topics of the moment, and every marketer should inquire and understand if they are or will be useful for the projects they manage.