The four cardinal points of marketing analysis.
This is an adaptation of an article that was published previously in Dentsu Aegis Network’s magazine Connected (2017).
Introduction.
Marketers nowadays have a wealth of data at their fingertips and they will need to make use of it in their everyday work, or risk missing out on valuable opportunities or worse, drawing faulty conclusions. However, most marketers are not data analysts and conversely, data analysts are generally not marketers and few companies have been able to bring their worlds successfully together. If you don’t know what the possibilities are, it’s hard to ask for analysis, but if you don’t give clear instructions, a data analysis is rarely going to give any unexpected insight.
This article attempts to provide a paradigm to think about various types of analysis that need to be in a marketers arsenal to discuss with his team of data analysts and data scientists. The Marketing Analysis Magic Quadrant, shown below is the general idea.
In this quadrant, I define 4 types of marketing analysis along the axis of rearward versus forward looking and channel focused versus people focused. There is a clear distinction between these four analyses in terms of the types of data required, but also in terms of expected outcome and application. Before choosing either of these options, you need to consider what the question is you are trying to answer.
Rearward looking means we are explaining what happened and interpreting the past, while on the right, we take it one step further and predict or influence what will happen going forward. In the top half of the image we are looking at individual paths of visitors, which means we need to know about each touchpoint. As a consequence, we can only work with online data. In the bottom half we also include offline data and create a model explaining the behavior and predicting future results, based on e.g. spend.
Attribution.
Looking back and interpreting in marketing a lot of times means attribution: dividing the value of the conversions over the campaigns we have had running, to understand the incremental value that each of those has had in our marketing mix. A lot has been said and written about attribution, yet a lot of marketers still struggle to successfully implement it in their workflows. Some of that may be caused by a confusing amount of types of attribution that all have their own set of requirements, methods, application and outcomes.
Please see below overview:
Specifications of all types of marketing attribution.
The first two types of Attribution, I don’t include in my Quadrant, as they are available without any form of analysis in most marketing and web analytics tools. Single Channel you will find standard in Adwords and most affiliate systems (with the 30 days cookie). Web analytics tools usually apply a last (non-direct) click model as standard. Typical for this type is that all value of conversions is attributed to one channel or campaign.
Rule Based Multi Channel is nowadays often an option in web analytics tools and lets you choose from for instance Linear Attribution (all touchpoints get the same value), Time Decay (the further from the conversion the lower the valuation), Position Based (different positions in the path give different valuations) and Weighted (the nature of and engagement caused by the interaction define the valuation). The last one will for instance give campaign impressions a lower value than clicks and channels that bring longer visits a higher value.
The first form of attribution that I do take up in my Quadrant, is Data Driven Attribution. This form is just like the previous two based on single source data. This means all touch points we analyse come from the same database and therefore we follow all visitors from their first through to their last visit. There difference is that in data driven, to be able to determine what works and what doesn’t, all touchpoints are analysed and not just the ones from converting users. From there a model is created automatically by the system each time the analysis is run, based on what the data suggests. The marketer doesn’t decide what is important, instant data analysis reveals what is and value is distributed accordingly. Models, mathematical formulas often used to describe this process are logistic regression, the hidden markov model and cooperative game theory. For a more in-depth description of attribution in all of its forms, I refer to my article Game Theory in Conversion Attribution.
Above version of attribution is as mentioned ‘single source’ and therefore only online activity can be modeled into the results. If we also need to model in offline campaigns, the method of analysis changes radically, as we cannot follow everyone’s touchpoints anymore. Therefore we look at spend, when did the advertiser air commercials for instance, or what week did they spend a great deal in outdoor, GRP’s are a great indication. The analysis that is then done is econometric, or statistical, meaning we will be looking for variance in the cause and the result metrics to then look for correlation and/or causation. For this, a significant amount of historical data - about 3 years is ideal - is needed to have enough data points, as the time series is usually as granular as on week level. The model thus created can be tested to see whether in the period that you have data for, it accurately predicts what the real values would have been.
In figure 2, you can see an outcome of an exercise like this. Data is modeled over the course of 1 year here, and the red line, the predicted value, follows the black line, the actual value pretty well. Then at the moment of the vertical line, a new campaign was launched, that was not taken up in the model and therefore all the positive effects of this campaign are shown as the difference between the prediction, the redline and the actual results, the black line.
Attribution of branded clicks following TV campaign.
Predictive and Prescriptive Marketing Analysis.
When you have created this model for multi source attribution and thus can predict historical revenue based on the input criteria, you can then use the same model to predict the results going forward. Not only that, but because you at that point have a good idea of the incremental contribution of each of the channels, you may want to take it a step further and optimize your media mix based on the budget you want to spend, or the sales targets you may have.
At that point you would be in the bottom right side of the Quadrant, doing Marketing Mix Modeling. There is some confusion every now and then around the terminology, whether to use Media - or Marketing Mix Modeling. Some people within Dentsu Aegis Network, prefer using the latter, so as to indicate that more than just media has gone into the model that predicts the optimal use of media. For instance macro economic data, weather data or competitive intelligence.
By taking all of these elements together and analysing them over about 3 years in one model, you create a baseline of sales that would have happened without any media buying at all. It makes a lot of sense that not all sales happens because of specific marketing campaigns and that amount is what the baseline sales represents. From there each channel is shown with its incremental value that it has contributed to the bottom line. This amount can then be compared to the amount that has been invested and the ROI of each of the channels compared.
A drawback of this way of working, is that only channels can be modeled in, that have had we have (a sizable amount of) historic data of. New channels can be attributed (as shown in figure 2) as long as they’re tried one by one, but not predicted.
The last form of marketing data analysis that is included in the Quadrant is not so much a data analysis as it is a data architecture, the foundation upon which continuous and partly automated data analysis can be performed. We call this data activation, because contrary to traditional setups, the data gathered here is immediately activated in order to optimize marketing campaigns.
Automation is key in this architecture, because the goal is to allow for real time segmentation and personalization. Someone doesn’t finish their basket session? Add them to a retarget segment based on the products in the basket and message them after a cooldown period. Someone did convert? Immediately take them out of the segment to retarget. Someone has shown a lasting interest in outlet products, put them in the Outlet segment, that is emailed and targeted with display and video ads when new discounts arrive. The number of segments you can create is endless.
At the heart of the architecture is a data management platform, a DMP, where data is gathered and stored on a per cookie basis. No PII is stored here (although what constitutes PII seems to be an everlasting discussion), but a unique identifier in combination with engagements that were performed as hits and events. On top of that, an Intelligence Layer is placed, where interpretation of these events happens, cookies are unified to unique users and they are segmented into marketing groups. These segments are the profiles built on the data points in the DMP. Links are made between users and devices so disparate cookie ID’s are stitched together. Preference for products, product types, product features are clustered. Average annual spending, buying frequency and recency are taken into consideration and if available, some demographic data is included (male / female, age and interest groups, etc). The rules upon which this segmentation happens is first done manually, but from then on machine learning can take over and optimize these segments. This is a continuous process.
Connectors with all major Demand Side Platforms (DSP’s) are created to sync the cookie ID’s with their audiences and find them on their advertising network. Within the DSP’s it is also possible to target Lookalikes - users with similar behavior, but who are not yet known in the DMP. That way also upper funnel marketing, like branding campaigns can be more intelligent and targeted and a lot of waste is cut out of the marketing spend. The cpm’s for targeted advertising are higher than for non-targeted, but on the one hand that can be funded by the budget that is saved from cutting out the wasted spend and on the other, it pays for itself in added intelligence you get about what works and what doesn’t.
Connectors with email marketing platforms allow for personalized emails to be sent out, as long as there’s also a connection with the CRM database, that contains the personal information. Within search advertising, the segments can be used to bid higher for the keywords that people in particular segments type in, that are deemed to have a higher chance of conversion. On the website itself, A/B testing tools can be fed with the data, to individualize landing pages and show products to people that according to their segmentation have a higher chance of conversion.
Which of the 4 cardinal points of marketing analysis is for you?
After reading this article, you should have a general insight in the purpose and distinctions of each of the types of marketing analysis described here. Attribution is a great starting point for understanding what your marketing efforts have resulted in and should be performed by any self respecting marketer. In case a sizable chunk of marketing budget is reserved for offline, you would automatically land at the lower half of the illustration. More advanced marketers will want to sink their teeth in making their models predict what will happen going forward and even prescript the ideal combination of channels in order to optimize the return on investment. Those inclined to understand their customers and audiences and aspire to the ideal of one on one conversations with them, should consider data activation, as this is undeniably the future of marketing. Do make sure however that you beef up your marketing department with some experienced data engineers and data scientists in order to succeed in this space.