Too much spending, too little advertising impact. The complexity of the channel mix is increasing and, at the same time, the boundaries between offline and online are blurring: Marketers are facing ever greater challenges as the factors influencing the effectiveness of advertising measures are not only becoming more diverse and complex but also increasingly intertwined. To use the available, often hard-bargained advertising budget with the highest possible efficiency, smart and proven methods are needed.
With a holistic attribution model, those responsible for marketing benefit simultaneously from the valuable expertise available in the company and the constantly evolving capabilities of machine learning.
An attribution model is a method that quantifies the effect of influencing factors on a target metric. A holistic attribution model thus helps understand the effect of media in interaction with other drivers and, based on this, optimize the use of the advertising budget or allocation of other marketing resources. The word "holistic" here stands for "comprehensive". This means that all relevant influencing factors are included in the model as much as possible. The resulting algorithm can produce meaningful forecasts for a wide variety of scenarios and offer the possibility of selecting the best scenario for implementation in each case.
To put it somewhat more abstractly: A holistic attribution model quantifies the impact of individual factors taking their cross-effects into account and enables valuable statements about the future development of the target metric. These insights help to allocate advertising budgets more reasonably and efficiently.
While well-known attribution models such as Last Click or First Click assign the complete or partial value of a conversion to a specific touchpoint, a holistic attribution model quantifies the effect of all relevant influencing factors. This helps avoid over-interpreting the impact of media support for a particular touchpoint.
With the help of meaningful impact forecasts for different scenarios, informed decisions can be made to select the most effective scenario. In addition, continuous monitoring of the actual development against the forecast is possible. Marketers can therefore react quickly to unforeseen events and thus get more out of their advertising budget.
There are also further optimization opportunities beyond media, as every influencing factor considered in the model, belonging to the company's scope of control, can be optimized in favor of the company's goals.
To develop a holistic attribution model, marketers should first set a desirable and at the same time ambitious goal, such as getting the maximum advertising impact out of a predefined budget. This overarching goal is then used as a basis to derive an easier-to-achieve milestone, such as increasing the advertising effectiveness by 15 percent with a set budget. This first milestone can then be achieved by going through the following seven steps:
To start, it is a good idea to choose a simple KPI that has already been tracked over several years and is important for the company's success.
A multi-channel retailer operates an online store, an app, and several brick-and-mortar outlets throughout Germany. Possible target KPIs to start with here could be:
- Number of online store visits
- Number of app starts
- Number of visitors in brick-and-mortar stores
- Alternatively, the total of the above
If a simple target KPI was selected in the first step, it should not be difficult to identify the factors that significantly influence this KPI. These can usually be determined based on the company's internal expertise. An important comment here: It is sufficient if only a few influencing factors are defined at this stage. It is also okay if it eventually turns out that a certain influencing factor does not affect the target key figure. The aim here is merely to generate an initial set of factors to structure the process of data collection.
The above-mentioned multi-channel retailer selects the traffic in its brick-and-mortar stores in Hamburg as the target metric. The potential influencing factors for this target metric can be:
- Calendar information (date, day of the week, month, year)
- Special dates (holidays, school vacations, or quarantine days, etc.)
- Socio-demographics of the region (size of population, age, income, purchasing power, etc.)
- Weather (hours of sunshine, hours of rain, temperature, etc.)
- Advertising and marketing measures (brand or performance campaigns)
- Sales measures (discounts, promotions)
- Market offer (price, product mix, its width and depth, etc.)
Data collection depends on the granularity of the data for the selected target key figure. If the data is available on a weekly basis, weekly values of influencing factors will be required. The following applies in general: The more measurements of the target metric are available, the better models can be trained. It is therefore preferable if the target metric is available daily rather than weekly because this alone would increase the number of measurements by a factor of seven.
The multi-channel retailer has had a daily visitors count over the last three years. Calendar information such as weekly breakdowns, public holidays, and vacation days as well as weather data are freely available on the web. Data on advertising and marketing activities can be provided by the relevant department in the company. This means that 3 years * 365 days, i.e. 1095 measurements/experiments per stationery store are available for model training.
Since the format of the data may vary, they must first be transformed into a uniform data set before the model can be developed. Here, the most pragmatic way to get the result should be preferred. It is important that in the end all the data are consolidated in one data set. The structure of the data set is defined by the target metric. If the input is available on a weekly basis, a table with the columns “calendar week”, “target metric”, “factor 1”, “factor 2”, etc. is required.
The multi-channel retailer has the daily visitors count per store as the target metric. Accordingly, it adds another column to the data set for each available driver. Since out-of-home placements always have to be booked for at least ten days ahead, the costs for the ten days are distributed equally over the target period (in this example, € 1000 per day).
Once the consolidated data set for the first iteration is available, the data is used to train the initial model by applying machine learning techniques. The results include the following:
The better the model is able to match the actual development of the target metric, the more certain we can be that all the relevant influencing factors have been taken into account, and the more meaningful the available results will be - in other words, the more suitable the current model is for forecasting future development and assessing various scenarios. It is advisable to aim for the model accuracy above 75 percent so that the actual development is explained sufficiently well by the model, and it can be used for operational purposes with a clear conscience. The following must be taken into account: The higher the achieved model accuracy is, the more difficult it will be to increase it further. The increase from 75 to 80 percent accuracy is much easier to achieve than the rise from 90 to 95 percent. The existing internal expertise of the company should also be used to challenge the model results, especially at the initial stage. This approach will help better assess the data quality and generate valuable ideas for the next iterations. In most cases, several iterations are necessary to produce an applicable model.
The model from the first iteration shows that the weather factor has a decisive effect on the number of visitors in stationery stores. In particular, on cold rainy days and very hot sunny days, the number of visitors drops, whereas on dry days with moderate temperatures it rises sharply. Since there are still some outliers in the development of the target metric that cannot yet be explained by the model, it has been decided to take sales information (discounts, promotions, etc.) into account in the next iteration.
Once an applicable model is available, we can project the possible development of the target KPI when, for example, the current media plan is implemented. To do this, we need information or assumptions about the expected values of the drivers during the planning period. The dates of public holidays and school vacation days are already known. As for the weather, we can assume that it will be the same as last year. The information about the planned marketing and sales measures should be available from the relevant departments.
The multi-channel retailer realizes that it will not achieve the targets of the corporate planning with the current media plan. It has been decided to create two additional media plan scenarios.
The resulting holistic attribution model can now be used in a variety of value-adding ways. For example, it can be applied to optimize media planning. But at the same time, it can also help boost the impact of discount-based promotions.
The multi-channel retailer estimates the visitor traffic for the two additional scenarios. In the end, it finds out that the combination of simultaneous advertising and discount promotions leads to a particularly high number of visitors per day on days with moderate temperatures. During the periods without attractive discount promotions, the advertising is scaled down. Advertising campaigns are particularly strong in the spring, as many dry days with moderate temperatures are expected during that time of the year.
A holistic attribution model is a powerful tool to increase advertising effectiveness and allocate the advertising budget more efficiently. It is therefore advisable to use this approach as a standard tool in every company with significant advertising budgets. Generating a perfect model in a short time is not as important as gradually rolling out this tool across the company to achieve long-term benefits.