Advertisers face the always carry heavier task responsible for the end-to-end management of the customer journey, across channels and devices. From the first moment of contact through the final acquisition to maintenance and service. Attribution analysis plays an important role.
The purpose of attribution is to determine to what extent and in what way the contact of clients with creatives affects the conversion. The idea is that a better understanding of the relative performance of example, search engine advertising, affiliate marketing programs and newsletters, as well as the interaction between the various advertising opportunities, advertisers allows to spend their budget more effectively.
But that's easier said than done. Indeed, it is complex, with many unknown factors. And use given the increased number of channels and devices to existing and potential customers for their decision and purchase process, the number of unknown factors will only increase.
Just managing data raises several questions:
We count clicks and / or ad impressions as a contact?
What time period we observe a typical purchase cycle when exporting data?
we can do a careful tracking cookies on all devices? And, it is possible for the customer journey thereby completely capture?
To what extent the gap between online and offline to be bridged?
What is the best model for the channels that provide customers effectively used a conversion score?
Figure 1: the customer journey from first contact with creatives until the time of purchase. How relevant are the touchpoints? How do they work together?
Figure 2: An example of a customer journey. This concerns, for example starting with the self entering a URL. This is followed by four email contact, a contact via a referring website, a new e-mail contact, and finally a contact using a search engine which ultimately results in a purchase transaction. Single touch 'models do not consider search engine ads, or with direct input from Internet addresses. They bring the customer journey mapping incomplete because channels like email are completely disregarded.
classic attribution
In practice, the most used simple heuristic models like ' first touch ' and ' last touch '. They provide, where necessary, virtually real-time results and their approach is simple to understand.
Both models indicate a conversion relatively simple manner to the channel a customer journey began or ended. This simplicity, however, brings a disadvantage with it. The ' single touch ' models show the intermediate classroom namely completely ignored, despite purchasing decisions will be affected.
'Multi-touch' models divide conversions across all channels that have contributed to it.
" Multi-touch ' models overcome this problem. They divide conversions across all channels that have contributed to it.
In the simplest case this happens proportionate. If the customer journey for example, takes you past eight touchpoints, each channel contributes to 12.5 percent of the result (8 x 12.5% = 100%).
However, a proportional distribution is not always realistic and effective. At the touch points that are at the beginning and end of a customer journey, often giving more relevance than at any intermediate stations. This type of differences is overcome, inter alia, by the position-based model. Usually get the first and last touchpoint assigned a weight of 40 percent. The remaining 20 percent is distributed to the intermediate, support channels.
However, there are alternative weighting imaginable.
In practice, even more approaches may be used. These are often variations on the rules-based attribution model, where the model has been adapted to personal preferences. The problem is that this kind of subjective approaches even be able to play an important role in, for example, the choice of the basic model, or the definition of the weighting factors. That is not conducive to an optimal outcome.
To score marketing channels for budget purposes at business and realistic, objective measures are needed. Furthermore, you could get a more complete picture if you take customer journeys in your analysis which did not lead to a conversion, while the classic models disregarded.
Here are algorithmic and probabilistic methods into play. These are terms that sound rather complicated. That this type of methods can still be used in an understandable way, demonstrated by a calculation which was presented as part of a research project. In addition, four shopping data sets were examined from different industries with a total of more than three million customer journeys.
Attribution according to a graphics-based method
In this model there is a transition chart a transition matrix centrally respectively (see figure below). Both model the likelihood of switching between touch points during the customer journey or the transition of a series of touch points to a different set of touch points. The graph displays it in the form of connecting lines. The matrix displays it in the form of entries.
All stochastic data can take simple post in tables based on process data that provide internet analysis tools (like Google Analytics).
For illustration depicted four customer journeys, one yielded no success:
Search> email> conversion
Search> Display> conversion
Search> email> Display> conversion
social
The overall conversion rate in the data set is less than 75 percent. The simplest Markov model, which will be discussed later, the transitions between the touchpoints following shows:
Figure 3: Examples of a transition matrix - charts for four customer journeys
In a representative customer journey dataset transition graph looks like this:
Figure 4: complex - the transition graph of a first arrangement for a representative shopping dataset
Overgangsmatrixen, such as the matrix which is displayed above the first graph, one usually read line by line, wherein the sum of each line is 1. The probability that the channel 'E-mail' (line 4) in the next step of the customer journey to "conversion" (column 6) results, thus amounts to 0.5 (= 50 percent).
'Start' and 'null' as artificial assumptions added to the model. Thereby showing off what channels often pave the way for a customer journey than other channels. Null represents a kind of model type final state that has no possible follow-up, with or without conversion. The value on the diagonal here amounts to 1 (= 100 percent), resulting in a loop.
Evaluation and interpretation: email is undervalued
In all shopping datasets being the popular attribution to a high conversion rate to the channel 'search engine advertising. Other channels, such as e-mail, "on the other hand were the ' last touch ' model and related approaches clearly undervalued.
Removal effects as comparison values
Provides such a model solution for attribution? Only as to calculate the value of the channel is.
A central question in marketingattributie is: what are the right decisions in the commercial field? Everything revolves primarily around the question of how the advertising budget is correctly assigned per channel. This is the so-called removal effect into play. This is to give the effect of the choices for certain channels.
The removal effect is a reliable indicator and works quite simply: by removing a channel entirely from a chart, the negative impact on the overall conversion rate the extent to which that particular channel contributes to the conversion. In extreme cases, it is removal effect equal to 100 percent of the total conversion ratio. In this case, there will be no conversion at all achieved without this one channel.
This model has a strong predictive value and is calculated by the multiplication of two probabilities:
The chance 'start' from the hub channel is visited.
The probability that the channel from the point 'conversion' is obtained in place of the point 'null'.
For 'search' is the removal effect in the above scenario amounts to 0.75 (= 0.75 x 1). It is equivalent to the overall conversion rate: three of the four customer journeys resulted in conversion.
This situation is directly visible in the chart: if one could reverse the line 'start' to the node search in the direction of the node null, there would be no conversion come about. The effect of 'social', however, is 0 (= 0.25 x 0). This channel therefore does not contribute to the conversion. 'Display' and 'email' are each from 0.5 (= 0.5 x 1).
Finally, in this analysis is a further step recommended. The proportion of an effect on the sum of all the effects removal, multiplied by the total number of successful conversions, resulting in the total number reached targets that can be attributed to the corresponding channel. Thus 'search' in theory accounted for approximately 42.86 percent (= 0.75 / (0.75 + 0.5 + 0.5)) of a total of three conversions. In both "email" and "display" this percentage is 28.57 percent.
In this way, the results of the Markov model are compared with those of the heuristic approach: relatively on the basis of the percentage share or absolutely on the basis of the channels distributed over the conversions. Here, the researchers found clear discrepancies between the solid results of the algorithmic and rules-based models ...
Evaluation and interpretation: email is undervalued
In all shopping datasets being the popular attribution to a high conversion rate to the channel 'search engine advertising. Other channels such as 'email', however, were the ' last touch ' model and related approaches clearly undervalued.
Figure 5: Third order of the Markov model (= knowledge on the current and historical data from the two touch points) versus the 'last touch', 'first touch' and linear model - search engine advertising 'by overvalued models based on rules, while 'e-mail and other channels are undervalued.
The results can serve here as comparative values a complex model. The complexity is measured by looking back based on how far the model within the customer journey.
In the simplest case (which is referred to as a "first order" model), only the current channel known. All probabilities are based solely on that one channel. At higher ordinations are not only at present but also the past. When the model of the second order transition probabilities are related to the current and previous touchpoint. When the model of the third organization they relate to whole sets of three channels and so on.
Models with a third arrangement not only being more accurate but also offer the advantage that the interaction can be studied between upstream and downstream channels.
The more complex, the more effective, but also more complicated. Models with higher organization let themselves easily convert to the first order. In addition, the model number of nodes increases exponentially. In this context, the researchers recommend the use of models up to one third planning to. These are not only more accurate, but also offer the advantage that the interaction can be studied between upstream and downstream channels.
Figure 6: attribution in two channel sets - the model of the second order is an analysis of the interplay of advertising channels preparatory channels
The fact that the 'start' condition considering the probability of closing the customer journey is most relevant (column 1), is due to the large number of customer journeys with one stage within the dataset. In addition, successive contacts often within the same channel a very strong effect.
It only 0.01 to 0 removal effect share an e-mail contact to another email contact follows (column 4, line 3) is close to 1 percent, while the effects to other channels are preceded (line 3) 27 percent hatch. Perhaps this reflects the channel preferences of many users.
Conclusion
The subject multichannel attribution remains fascinating. It is interesting for anyone who wants to attract customers through multiple advertising channels such as search engine advertising, email newsletters and affiliate partner programs.
The goal is to get an answer to the question to what extent these individual channels or the combination thereof contribute to the intended conversion rates. For whom 1,000 euros spending on search engine advertising, email marketing and affiliate programs and after three months finds that the revenue therefrom 6000, is 3,500 or 500 euros, will be choices or not reconsider in the next quarter.
It is too short by the bend to merely measure the advertisement channels based on the first or last contact in the sale process. Attribution should, where possible, all the channels involved, including the judge initiating, supporting and closing channels, honestly. To this end, Anderl and his colleagues in their whitepaper algorithmic and objectifying approach, based on the Markov chain, which is capable of large numbers of customer journeys efficiently process, whether or not led to purchase. This can be a useful supplement to traditional multi-touch models.
The fact remains that the most complex model is not necessarily the best model
A relativistic side note: the fact remains that the most complex model is not necessarily the best model. If the customer journeys of no more existence than an ad contact direct conversion, the first simple "first click'- and 'last click' rules should provide all satisfactory results. The results should always be seen in the light of the prevailing conditions. More ads means or budget limitations per channel for example, can yield very different results.
The idea of a sizable dataset quickly and completely in the manner described to assess nevertheless has a certain charm. Something is better than nothing. If an assessment not all contribute to the optimization of the media mix or the improvement of the customer relationships, then provides a basic review anyway, at least supporting information in the choice of the right traditional attribution model. Which let themselves after all, compare and analyze effectively with tools such as the ' Model Comparison Tool ' Google Analytics. This tool allows simple simulations and comparisons.
It is crucial to make the right, informed choice for attribution. Want conversion attribution to a higher level, then work towards a deeper model. But one that fits the activities, nature and customer journey within the company. And consistent with the level of the marketing department.
No comments:
Post a Comment