Good news: You’ve decided to abandon last-click attribution measurement and made a firm resolution this year to step up your media attribution analysis game. Was losing third-party cookies and your ability to measure channel performance the reason?
Or, you’re just think it’s time to change the way you’re evaluating upper funnel media, then media mix modelling can be a solution. Why? The primary reason being that it can report on performance without the same requirements for impression and click-based journey tracking.
Media mix modelling is not exactly in its prime youth
It is indeed a 50-year-old practice based on the two following principles:
- The sales of a product are driven by a mixture of factors. These include variables such as its price, its availability, the number of competitors, the size of its market, and the market’s awareness and desire for that product – driven by advertising and promotional activity.
- As the different factors of the mixture are changed, the effect on sales can be measured. The impact of different factors in the mixture can then be calculated using statistical approaches.
In short, media mix modelling focuses on understanding the relative contributions of different channels, or different funnel stages of activity, on performance of a media campaign.
Media mix modelling can provide insight into base and incremental conversions and revenue. Simply put, incremental results are the percentage of media-touched conversions that would not have occurred without marketing exposure. The analysis of base and incremental results can reveal the underlying brand awareness and loyalty over the long term, as well as the effect of shorter-term advertising activities and competition levels.
Traditionally, established statistical techniques such as linear regression analysis have been used to perform these calculations, but more sophisticated tools are emerging to take media mix modelling to the next level.
How machine-learning is elevating media mix modelling in 2023
Since the 1960s, statisticians have patiently performed media mix modelling analyses using traditional econometric tools such as linear regression, sometimes even by hand. But media data have become more complex over the years, and the shortcomings of traditional methods have become increasingly apparent. Similarly, media data has become increasingly available and granular, with the multiplication of first-party data enriched by second- and third-party data. Finally, the computing power of machines having been overmultiplied and all the conditions were met to allow new machine learning methods to be applied in the framework of media mix modelling analysis.
Some of these approaches, such as Robyn—Meta’s ML-powered and semi-automated marketing mix modelling open source package launched back in 2021—seek to simplify the technique, allowing users to “pour” their cost and conversion data into a data science workflow and generate a selection of models. In short, machine learning simplifies the process: It is able to iteratively and rapidly tweak the model for better performance.
So now you only need to plug in the data and voilà?
Did you really think we were going to tell you that a simple (or not so simple!) machine would do the job perfectly?
Even if it is technically possible to simply put the data into the machine and cross your fingers that everything goes well, we are convinced of the importance of the strategic contribution of data scientists, statisticians, and subject matter experts to develop a truly relevant and efficient model.
Here are some situations where deeper investigation would be more than welcomed:
- Different datasets may be better suited to different analysis using different model types
- Different datasets may benefit from manual ‘tuning’ of some of the parameters required by the algorithm
- Model outputs may simply not be actionable: perhaps the model’s recommendations are not aligned with your business objectives or the recommended tactics cannot be effectively deployed in the marketing channels used
- Initial exploratory analysis of the data may allow the building of more accurate models
Let’s explore this last case in a little more depth. For example, you could have decided to include “paid search” as a single channel in your analysis. This seems like a logical thing to do, but within each channel there may exist different behaviours of the data, and these vastly different patterns may limit the accuracy of the model:
Exploring the data before the model build may uncover these differences, such as between branded keywords and generic keywords campaigns or variation of performance between regions, which can inform the model building process and improve model accuracy:
The myopia caused by hidden patterns in the concatenated data can lead to questionable decisions that result in unnecessary risk-taking and potential disastrous consequences on campaign profitability.
What can happen if you trust the machine blindly and how to avoid the pitfalls
Among the potential negative impact of over-reliance on machine learning media mix modelling tactics, you could for example end up setting your budget splits completely inappropriately, overinvesting in irrelevant channels or underfunding others. You could also end up with a model output that is incorrectly interpreted—that would be the case if you are provided with an aggregate model for the whole year but are in fact in a highly seasonal business.
The solution? Common sense, first and foremost. Asking yourself if what the model presents to us makes intuitive sense, based on your knowledge of the industry, the company, its products, its past successes.
And the second option, complementary to the first: The implementation of a robust process of testing and experimentation, to give yourself the opportunity to explore all scenarios, even the most unexpected, in a controlled environment.
All aboard the media mix modelling train?
Is media mix modelling version 2023 for every company and every industry? Well, it is for you if:
- you have enough data and that your data is in good shape, because machine-learning is feeding of big quantity of clean data;
- you understand that platforms are evolving over time and that historical data may not always be the best representation of how a platform is going to perform in the future—for every company, including your competitors;
- or you are looking for high-level, strategic insights into the impact and effectiveness of marketing channels. If you are looking for more granular, tactical insights into the role of different touchpoints at different stages of the journey, you might not want to say goodbye completely to clickstream data analysis yet.
We hope that this first excursion into the fascinating world of media mix modelling has helped you ask yourself the right questions. Do not hesitate to contact our experts to see if this technique is the best option for your company.