Skip to main content

Overview

In today’s multi-channel marketing environment, understanding the impact of each customer interaction on final conversions is crucial and incredibly complex, especially as consumers often engage with multiple marketing-driven interactions before making a purchase. While there have been considerable advances in marketing measurement frameworks, many marketers still rely on heuristic models. These models can inaccurately report marketing performance, leading to misinformed decisions in performance management. An in-house Multi-Touch Attribution (MTA) system provides a sophisticated approach to quantifying the effectiveness of each marketing touchpoint. This holistic view enables marketers to allocate budgets and design campaigns that maximise ROI. In this article, we will show you from the general concept of MTA to a step by step guide to build an in-house Multi-Touch Attribution.

Define Your Attribution Goals For Your Business

Before diving into the technical aspects, clearly define what you want to achieve with your MTA model. This could include understanding the impact of specific channels, optimising marketing spend, improving customer acquisition costs, or enhancing ROI.

It is also worth considering how your MTA model fits into your marketing measurement stack. Generally speaking MTA can be used for tactical optimisation and performance management for mid to lower funnel campaigns. Meanwhile incrementality testing and MMM models can be used to validate or calibrate your MTA model to take into account effects which are not modeled easily in MTA.

Data Collection

Collecting the right data is crucial for an effective MTA model. Ensure you have the means to track and store data from all customer touchpoints. This includes:

  • Web analytics tools: For tracking website visits, user behaviour, and interactions.
  • CRM systems: For customer data integration.
  • Ad platforms: For campaign performance data.
  • Social media analytics: For engagement and referral traffic.

Consider implementing a data layer on your website to consistently capture and structure the data you need.

Be aware that we can’t collect all interactions. Data privacy regulations, technology limitations, cross-device challenges, and offline interactions make it impossible to capture every touchpoint. Acknowledging these limitations is crucial for setting realistic expectations and continuously improving data collection strategies.

How to Choose the Best Marketing Attribution Model

Select an attribution model that best fits your business needs. There are many different types of attribution models, but they can be categorised into two categories as follows:

  • Heuristic attribution models - these are models that are decided in an arbitrary manner. The main heuristic models are:
    • Last touch attribution model - where we assign conversions to the last touch point
    • First touch attribution model - where we assign conversions to the first touch point
    • Linear attribution model - where we split the conversions equally amongst touch points
    Server-side tracking diagram
  • Multi Touch attribution models - this approach uses statistical methods or machine learning to split conversions amongst all touch points that led to the conversion. The most common models used are as follows:
    • The Markov attribution model: Modeling the removal effect in the Markov attribution model assesses the impact of removing specific touchpoints on overall conversion outcomes and credit allocation to other touchpoints. It helps marketers understand the direct and complementary roles of different channels in the customer journey, enabling optimisation of resource allocation for maximum conversion impact.
    • Shapley attribution model: The Shapley value model in attribution is rooted in game theory. By considering all possible combinations of channels and their respective contributions, akin to players in a cooperative game, the Shapley model provides insights into the relative importance of each channel within the marketing mix, creating a cooperative understanding of channel interactions and contributions.

Building a Multi Touch Marketing Attribution Model

1. Track returning users using first party server side tracking and fingerprinting

In order to build a MTA model, it is necessary to obtain data on how prospective customers interact with digital marketing channels. Tracking systems allow us to identify returning users so that we can build this data set. In recent years tracking at user level has become increasingly difficult, due to stricter privacy regulations, new browser features, and widespread use of ad blockers and privacy tools. Nonetheless, even partial data is useful in measuring marketing performance as it provides a strong signal to be used in modeling. By following best practices in tracking implementation we can maximise the percentage of returning users who can be identified. Two important ways to increase this percentage are:

  • Implement first-party server-side tracking. Server-side tracking enables tracking a higher share of users by bypassing browser limitations and privacy concerns inherent in client-side methods. You may still need to use first party cookies to identify a returning user, but this is preferable to using third party cookies as they are blocked less frequently.(ref 1)
Server-side tracking diagram
  • Use fingerprinting. When users do not consent to tracking, fingerprinting techniques can help to identify returning users using other non-personalised information. Using features such as browser type, operating system, screen resolution, plugins, fonts and whatever other information is available we can create a machine learning model to identify a user in a probabilistic manner. This allows us to track returning users even if cookies have been cleared or consent for tracking has not been granted.(ref 2)
Fingerprint Diagram

The output of this phase is raw data which allows us to identify a returning user and the marketing touch points which triggered the session


2. Create a data pipeline to structure data into user journeys

Now the data has been collected from your tracking system in raw format, it is time to structure the data. To do this we create a data pipeline to retrieve the data and then to process, clean, label and transform it into user journeys. As with any data modeling project a tool such as dbt can be used to efficiently structure the data. There are several important areas of focus in this step of building your in-house marketing attribution model.

In this article, we will use the following definitions:

  • Journey: The sequence of interactions and touchpoints a customer experiences with a brand from initial awareness through to conversion and beyond.
  • Session: A group of user interactions with a website that take place within a given time frame, typically ending after 30 minutes of inactivity.
  • Conversion: A key action that a marketer wants the user to take, such as a purchase, form submission, sign-up, or download.

  • Data stitching - This entails retrieving user level data (possibly from different data sources) and processing it into an event log of user touchpoints. This can be quite complicated, especially in some tech stacks where users might interact with multiple systems, each of which generates data in a unique manner and uses different identifiers. For example, a user might have a different identifier on the app and on the website, but these can be connected through an email address.
  • Decide time windows - various decisions need to be made on time windows to use. For example, the lookback window will define the amount of days to look back in time for user touch points. Generally companies use 30-90 windows. The window should be long enough to capture a high % paths completed, but be sufficiently short to make results actionable for the marketing team.(ref 3)
Conversion Window
In most cases, we typically recommend sticking to the default 30-day window.

  • Create sub channel inputs -  Many companies make the mistake of directly inputting campaign level data into the attribution model. This causes the data to be too noisy and the attribution model may fail to pick up patterns. It is a best practice to create sub channels that are more granular than channels but less granular than campaigns. For example, google ads could be split into branded search, generic search, competitor search. Usually marketing teams have this naming structure already in place which can be leveraged for creating the sub channel inputs.

Once finalised, we have a clean data set of user journeys and correctly named marketing touch points available for use in the statistical modeling phase.

Example 1
Fig. Example of cleaned data organised into an event log of user touchpoints

Example 2
Fig. Example of final data structure used.


3. Building MTA using Channel Attribution Pro

Once your data is cleaned and organised into user journeys, the next step is to build your MTA attribution model. Using ChannelAttribution Pro library will give you the opportunity to leverage a variety of sophisticated functions such as:

  • heuristic_models: Implements heuristic rules for attribution, such as first-touch, last-touch, and linear attribution models.

Heuristic Model

Heuristic Model

Heuristic Model

Heuristic Model
  • Markov_model: Utilises a Markov chain model to analyse the transition probabilities between channels and calculate the attributed value for each channel.

    Table: Result of markov model function, acv refers to average customer purchase value.

    Markov model
    Markov model
    Visualisation of the markov model output.

    Markov model
  • Shapley: Applies the Shapley value method from cooperative game theory to fairly distribute the total value (conversions) among all contributing channels based on their marginal contribution.

  • new_paths_attribution: Trains a Markov or Shapley model to perform attribution on new customer journeys, enabling real-time transaction-level attribution. This function is used after deploying your model.

  • choose_order: Selects the optimal Markov model order using an out-of-sample algorithm, maximising the average area under the ROC curve (AUC) through cross-validation.

  • combine_mta_mmm: Combines transaction-level attribution of a multi-touch model with channel-level attribution performed using a media-mix model.

  • budget_allocation: Allocates your budget to marketing channels based on different attribution models, including UAM as well.

Budget Allocation
  • compare_models: Compare the performance of first touch, last touch, linear touch, Markov model, Shapley value and logistic regression in a simulated traffic allocation problem based on real data.

    Compare Models
    Visualisation of the compare_models output

  • UAM: The Unified Attribution Model provides a unique approach to model performance of upper funnel campaigns or where tracking is limited. It takes impression data as well as touchpoint data as input and combines these. This model is unique on the model and not available in any other open source or proprietary packages.

To effectively use the ChannelAttribution Pro library, your input data should be structured and comprehensive. For detailed instructions on formatting your data, refer to the callout tutorial provided in the previous section of our documentation.

Once you have your dataset ready and you are prepared to use the library, check our tutorial on how to use the library for step-by-step guidance, or explore our documentation for a detailed overview of its functions.

Implementation, Communication, and Continuous Improvement

Validating your Attribution model

Once you have built your model it is important to validate before releasing it. This task is inherently complex as there's no definitive ground truth for validating an attribution model conclusively. Given this challenge, how do we determine which attribution model to implement? There are two main approaches which you can use: Experimentation or Simulation.

Fingerprint Diagram

Communicate with stakeholders

Effective communication with stakeholders is crucial when building an in-house model. Explaining the workings of the model, including its assumptions, methodologies, and limitations, helps stakeholders understand the decision-making process and the rationale behind the model's design choices. This transparency facilitates agreement on project timelines, deliverables, and resource requirements between the marketing and data teams.

Regular updates about the project's progress, challenges faced, and milestones achieved are essential to keep stakeholders informed about the model's development. Sharing intermediate results and preliminary findings allows for early feedback, which can be invaluable in refining the model. For instance, discussing with stakeholders whether the Return on Advertising Spend (ROAS) or Return on Investment (ROI) results from the Multi-Touch Attribution (MTA) model align with their experiences can provide critical insights. Additionally, evaluating whether the model's performance reflects the current strategy and supports stakeholders' goals is vital. By involving stakeholders from conception to implementation, you not only gain their support but also ensure that the final model is robust, reliable, and aligned with organisational needs.

Ongoing Refinement

MTA is not a set-and-forget system. Regularly update your model to incorporate new data, channels, and changing business goals.

Challenges and Limitations

Building an in-house MTA system is not without its challenges. Being aware of these challenges can help in strategising effective solutions and setting realistic expectations. The major limitations include:

  • Data Imbalance and Low Conversion Rates: MTA models rely heavily on comprehensive and balanced data from various touchpoints. However, when the data is strongly imbalanced or the conversion rate is exceptionally low (typically less than 1%), the model may face challenges in accurately attributing conversion credit. In such cases, techniques like oversampling may be required to rebalance the dataset, but these methods can introduce their own biases and must be carefully managed to ensure they do not skew the results.
  • Insufficient Channel Diversity: The effectiveness of MTA is contingent on having a diverse set of marketing channels and a significant percentage of cross-channel interactions. If a business does not engage with a wide array of channels, or if the interactions are not sufficiently integrated across channels, the attribution model may not provide useful insights. This scenario is common in smaller businesses or those just beginning to scale their digital marketing efforts.
  • Modeling value of upper funnel: MTA models rely on touchpoint data which is often not available for upper funnel campaigns. So for companies that have a significant share of upper funnel spend other methods such as UAM or MMM should be layered in.

Conclusion

Building an in-house Multi-Touch Attribution (MTA) model provides marketers with a powerful tool to analyse and optimise their marketing efforts by accurately assigning credit to each touchpoint in the customer journey. From defining clear attribution goals to collecting and structuring data, constructing a robust model involves several critical steps. It's essential to validate the model, communicate its findings with stakeholders, and regularly refine it to reflect new data and evolving business objectives.

While an in-house MTA model offers deeper insights into marketing performance, it is important to remain aware of the challenges such as data limitations, low conversion rates, and the difficulty in measuring upper-funnel activities. Overcoming these challenges requires a careful balance of technical accuracy and practical application, alongside continual calibrates the outcome using other marketing measurements such as Marketing Mix Model, Incrementmaality testing, Unified Attribution Model as marketing landscapes and technologies evolve.

Ultimately, implementing a sophisticated attribution framework allows for better resource allocation, improved marketing strategies, and, most importantly, a more precise understanding of how each channel contribute to business success. As marketing ecosystems become increasingly complex, in-house MTA models will continue to be a vital part of a data-driven marketing strategy.

Reference

  1. https://renta.im/blog/what-is-server-side-tracking/
  2. https://soax.com/blog/browser-fingerprinting-what-is-it-and-how-does-it-work
  3. https://growmyads.com/what-to-know-about-googles-default-lookback-attribution-window/