Conversion Delay: Marketing Performance

When it comes to quantifying the performance of marketing channels, the first thing that comes to mind is counting the orders per day that were placed per marketing channels and summing up their corresponding profit or revenue. To improve this direct approach, attribution models can be used to redistribute the impact of an order to the preceding interactions of the customer: The fact that the customer finally bought a product via marketing channel A is of course important, but so is the marketing channel B that caused the customer to start interacting with your shop after all – ultimately leading to spending money. Attribution models therefore reallocate (parts of) the revenue to the previous stations of the customer’s journey. This can be done either via attribution models that assign increased ratios to the first and last interaction(s), while giving less impact to those in the middle. Alternatively, dynamical attribution models determine these weights based on the data of all your customers, resulting in bespoke weights fitting their behaviour. If you are interested in more details on dynamical attribution models, you can read more in the corresponding separate Insights article on this blog.

Customer Journeys

Underlying these attribution models are the customer journeys, following certain rules like maximum length or whether they end after an order. This prescribed logics has direct influence on the attributed order shares / revenue: If a journey can have a maximum length of, say, 50 days, interactions of the last 50 may still get assigned a certain amount / revenue. If your journey logics allows for journeys with several consecutive orders, this holds even more. This circumstance is referred to as the conversion delay: some of yesterday’s interactions with your website may well only lead to revenue in the next upcoming days.

Conversion Delay

When you are interested in evaluating the overall impact of a marketing campaign, e.g., the attributed orders / revenue is already a good additional measurement. However, it can only consider orders that were already placed. If you want to monitor the daily performance of more recent days, too, you need to compensate for the conversion delay and with appropriate tools as well as suitable data, this is indeed possible!

The first idea of how to possibly fix the revenue gap due to the conversion delay is of course to use the historic rates of attributed revenue per interaction of the marketing channels. Knowing yesterday’s interactions per marketing channel, we can then estimate the revenue that is likely to be generated via these interactions over the next days. While this may already yield a reasonable result, it is not of much help when monitoring specific marketing campaigns where the customers’ behaviour deviates (as provoked!) from the usual one: the revenue per interaction is probably quite different on Black Friday than the average ratio and maybe even different enough before Christmas as compared to summer. The latter may be compensable by considering the average ratio of the last weeks only, but more temporarily limited scenarios like Black Friday still won’t be covered well enough.

Machine Learning

This is where machine learning models come into play. These allow us to draw on more input data, enabling the model to derive, e.g., whether a special situation like Black Friday is present or whether your customers show seasonal behaviour. When it comes to data acquisition, generating the training data for such a model can become rather complex, especially if the model should also use monetary values as input data: These should not be the ones the attribution model provides as they rely on the “perfect knowledge” of all future orders within the next 50 days after the reference date. If we want the model to consider the attributed revenue of the respective last days – to derive whether we are currently in a phase of a high rate of revenue per interaction –, the model should be trained on these values the same way we know them when evaluating the model today. This means that for all days of the past, we need to acquire the attributed revenues of the respective last days as we would have known them back then, on that day.