"Attribution is dead they say"

You know what and to be blunt, hearing people say 'attribution is dead' really gets my goat.

And before I go all in, let me make the distinction required for the theme of this article between ‘attribution’ of marketing spend to success/failure, and ‘attribution modelling’, one of the multiple approaches to monitoring and measuring marketing effectiveness.

As marketers who have a budget of X to achieve targets of Y, surely we want to know the effectiveness of our well-fought-for money coinage; surely......?

Is that content, on that channel, targeting those audiences, with that spend actually working?

Hmmm - define 'working'..... I'll come back to this in moment.

Here's the thing - I don't think attribution modelling is dead, I think with all the data points we have these days as marketing spends continue to migrate from offline to digital, and where offline has and continues to be smarter in its tracking, I would reframe this sentiment.

How attribution modelling is used - and often in silo to other models - is often wrong, misleading, and disjointed, and, depending on the attribution model one wishes to use at a given time, can often help feed a narrative that justifies a spend or activity; genuinely or manipulatively

So, by definition, what is attribution in a sense of marketing effectiveness?

I asked my AI personal assistant (sounds way more fancy than it actually is, but on a side note, highly recommend using Monica) what by definition it is.

"In the context of marketing, attribution refers to the process of identifying and assigning credit to the various touchpoints or interactions that a customer has with a brand before completing a desired action, such as making a purchase or signing up for a service. The goal of attribution is to understand which channels, campaigns, or touchpoints are most effective in driving conversions and to allocate marketing resources more efficiently.

Here are some key aspects of attribution:

1. Touchpoints: These are the interactions or engagements a customer has with a brand across various channels (e.g., social media, email, paid ads, organic search).

2. Conversion: The desired action that a customer completes, such as a purchase, sign-up, or download.

3. Attribution Models: These are the frameworks used to assign credit to different touchpoints. Examples include first-touch, last-touch, linear, time-decay, and data-driven attribution models.

In a broader sense, attribution can also refer to the act of ascribing a particular outcome or effect to a specific cause or source. This can apply to various fields, such as psychology, where it involves determining the reasons behind people's behavior, or in general use, where it involves identifying the origin or cause of an event or action."

Ok, that makes sense. Why wouldn't we want to understand the attribution of our marketing spending and efforts?

Having worked in the sector for ~20 years; brand side, agency side, and product development side, I've always been accountable for marketing spending to achieve a desired outcome. Attributing spend to outcomes has played a huge role in my decision-making process. It has had holes, and been somewhat flawed at times, but layered with some common sense, testing various things over time, and understanding where the holes in the analysis could be, it has always given me a steer towards further success and continuous improvement.

We all know the purchase process is often non-linear, across multiple touchpoints, on multiple devices over time. It could be researched by one person and purchased by a partner. The combinations are immensely complex and different.

On a grander scale, there are patterns in the role of a said channel versus another, or an ad format versus another but layer in 'brand' and the compounding effect of this (good and bad), word of mouth, creativity and messaging in terms of resonating and grabbing attention and interest, and a perfect way of 'attributing' one channel versus another is nuanced at best, and complex full stop.

But I didn't say it was dead did I, because it isn't!

So, what are the various models for understanding the effectiveness of your marketing I hear you say?

There are three main approaches, as per the below.

• Attribution Marketing Modelling (AMM)

• Marketing Mix Modelling (MMM)

• Incrementality measurement

Attribution Marketing Modelling (AMM)

Attribution Marketing modelling (AMM) - the digital natives amongst us will be highly familiar with this approach.

There are a number of different attribution models within that are used and debated. All of these have their pro’s and cons. Summaries and use cases below for reference.

Last click attribution

• Description: Assigns all credit to the first interaction that led to the conversion.

• Use Case: Useful for understanding which channels are effective in generating initial interest.

First click attribution

• Description: Assigns all credit to the first interaction that led to the conversion.

• Use Case: Useful for understanding which channels are effective in generating initial interest.

Linear attribution

• Description: Distributes equal credit across all touchpoints in the customer journey.

• Use Case: Provides a balanced view of all interactions, but may not accurately reflect the impact of each touchpoint.

Time-decay attribution

• Description: Gives more credit to touchpoints that occur closer to the time of conversion.

• Use Case: Useful when the timing of interactions is crucial, reflecting the increasing influence of recent touchpoints.

Position-Based attribution

• Description: Allocates 40% of the credit to both the first and last touchpoints, with the remaining 20% distributed among the middle interactions.

• Use Case: Highlights the importance of both the initial and final interactions while acknowledging the role of middle touchpoints.

Data driven attribution

• Description: Uses machine learning algorithms to analyse data and determine the contribution of each touchpoint to the conversion.

• Use Case: Offers the most accurate and dynamic insights by adapting to changes in consumer behaviour and marketing strategies.

In summary - it can be excellent for granular, real-time insights and optimising individual touchpoints but can be complex, short-term focused, and not - through GDPR, technical and offline/online touchpoints, give a full and accurate picture.

Marketing Mix modelling (MMM)

Marketing Mix modelling (MMM) offers a holistic, long-term view of marketing effectiveness, including offline channels, but may lack granularity and rely on historical data.

As with attribution marketing modelling, there a number of different models within this, all of which have their own strengths, weaknesses, and use.

Econometric modelling

• Description: Uses econometric techniques to model the relationship between marketing activities and sales. It often involves regression analysis to quantify the impact of various marketing inputs.

• Use Case: Suitable for businesses with access to detailed historical data and a need for precise measurement of marketing impact.

Bayesian modelling

• Description: Employs Bayesian statistical methods to incorporate prior knowledge and uncertainty into the modelling process. It allows for more flexible and dynamic models.

• Use Case: Useful when there is uncertainty in the data or when integrating expert knowledge into the model.

Hierarchical modelling

• Description: Involves building models that account for different levels of data, such as geographic regions or product categories, to capture variations across these dimensions.

• Use Case: Ideal for large organisations with diverse product lines or operations in multiple locations.

Time-series analysis

• Description: Focuses on analyzing data over time to identify trends, seasonal patterns, and the impact of marketing activities on sales.

• Use Case: Effective for understanding the temporal dynamics of marketing effects and planning future campaigns.

Constrained optimisation

• Description: Integrates optimization algorithms with MMM to determine the optimal allocation of marketing budgets under specific constraints (e.g., budget limits, channel capacity).

• Use Case: Helps businesses maximize ROI while adhering to real-world constraints.

Machine learning approaches

• Description: utilises machine learning techniques to enhance traditional MMM, allowing for more complex modelling and the ability to handle large datasets.

• Use Case: Suitable for organizations looking to leverage big data and advanced analytics for more accurate predictions.

Incrementality Measurement

Incrementality measurement aims to determine the additional value generated by a marketing action that would not have occurred without it.

A/B Testing (Randomised Controlled Trials)

• Description: Involves splitting the audience into a test group (exposed to the marketing action) and a control group (not exposed), then comparing the outcomes.

• Use Case: Provides a clear measure of incrementality by isolating the effect of the marketing action.

Geo-Testing

• Description: Similar to A/B testing but conducted across different geographic regions. Some regions receive the marketing treatment, while others serve as controls.

• Use Case: Useful for assessing the impact of marketing activities that can be geographically isolated, such as regional promotions.

Matched Market Testing

• Description: Involves selecting similar markets or segments and applying the marketing action to one while using the other as a control.

• Use Case: Effective when it's challenging to randomise at the individual level, allowing for a more natural experiment setup.

Pre-Post Analysis

• Description: Compares outcomes before and after a marketing intervention within the same group. Adjustments are made to account for external factors.

• Use Case: Useful when randomisation is not possible, though it requires careful control of confounding variables.

Synthetic Control Method

• Description: Constructs a synthetic control group using a weighted combination of untreated units to estimate what would have happened without the intervention.

• Use Case: Effective for single-case studies where traditional control groups are not available.

Regression-Based Approaches

• Description: Uses statistical techniques to model the relationship between marketing actions and outcomes, controlling for other variables.

• Use Case: Suitable for complex environments where multiple factors influence outcomes.

Uplift modelling

• Description: A machine learning technique that predicts the incremental impact of an action on an individual's behavior.

• Use Case: Allows for personalised marketing strategies by identifying which individuals are most likely to be influenced by a campaign.

Each method has its strengths and limitations, and the choice of technique depends on factors such as the nature of the marketing activity, available data, and the ability to create control groups. Below is a comparison of the pros and cons of both AMM and MMM.

Digital lured & seduced us….

The thing is when digital became 'a thing', it lured and seduced us with its real-time insights and instant gratification of what we (at the time) were able to deem as a success or failure. This, at the time, was often at the envy of its marketing-mix peers on other channels; radio, TV, papers, PR etc.

I have definitely experienced the short-term'ism of 'attributed' success/failure of a channel or marketing activity versus the longer burn often required to shift brand positioning and consumer perception/behavior.

I read recently that it was felt by a leading authority on the subject of attribution, marketing mix modelling, and econometrics, that the majority of the 'profit' from a marketing mix isn't actually really known or understood - that is, what we can't immediately 'see', doesn't mean hasn't had a positive (or negative) influence on the consumer and / or brand over time. If we spent more time considering, studying, and observing the longer-term 'more than one campaign long' effect of marketing channels, we may actually find additional valuable insights.

There are two key thoughts that come to mind for me:

1. Marketers (IMO) have become reliant on short-term shotgun tactics to drive conversions and immediate revenue (and often introduce a whole host of further tactics like discounts and frequent sales which damage profit margins and price elasticity for the hope of generating quick win conversions. Attribution modelling has helped support/fuel/justify this to a point.
   
   

2. Other models have then evolved - and rightfully so - to help both understand and also justify spends on other channels that historically have been 'hard' to attribute success metrics too. They have - rightfully so - been put in place to better understand the bigger picture, the shift of the brand awareness and consumer perception; over time.

And, are we neglecting the other P’s aside from Promotion when we are planning and strategising our next big campaign, and then reporting the success/failings of it - probably yes, but that is a whole other article; coming soon!

Closing thoughts

I personally think that, like anything in life, there are pros and cons associated with all of these methods and approaches.

Case in point: I am going camping soon and need to fit more into a car that keeps having less and less space (blame my three boys who are under 8 and keep growing for that!). I find myself searching more and more for 'compact' camping chairs, camping beds, etc., to replace their bigger counterparts sitting in the garage. Even with the number-one recommended products on the review sites, there are always cons associated with them. So, it's about preference and purpose at a given time for a given need—those preferences and needs can be interchanged when and where appropriate

On the subject of ‘pre-testing’ - I’ve heard so many people who do pre-testing that they ultimately are optimising for the green light of pre-testing, which once achieved gives a short-lived shot of serotonin along with a celebratory drink (or two) only to quickly relapse at its far too frequent failing predicted trajectory once out in the real world. 

Case in point - no model is perfect.

Linear or last-click attribution may suit your immediate attribution needs well. Likewise, MMM or incrementality measurement may be better (if you have the means and/or the budget).

Let me explain.

On the one hand, it is trying to understand and isolate the effectiveness of an individual marketing channel while also understanding its place in a wider marketing mix. I've seen it time and time again where an 'underperforming' channel is pulled from a marketing mix because what is being attributed to it isn't passing the 'success' markers that are set, only for an immediate and negatively geared domino effect happening on overall sales. I don't think that's attribution itself that is at fault, again, it's how it is being used, and the hypothesis that are created to plug a hole in the modelling not being tested.

In summary

Choosing between AMM, MMM, and incrementality measurement depends on your specific needs, such as the level of detail required, the importance of real-time data, and the ability to measure long-term impacts. The accuracy and usefulness of the outcomes are limited by a number of factors—know and understand the limitations, and like anything in life, what you put in, you get back. People often wonder why a campaign that has used the outputs of a qualitative session or survey hasn’t worked, only to overlook the flaws in the questions and format of these sessions. It was doomed before it started.

All of the various models and approaches have pros and cons. Some give you the short-term summary that is necessary, while others provide a bigger picture view and show how the marketing is moving brand awareness, brand positioning, and consumer behavior over time. Some of the models utilise highly complex mathematics across hundreds of touchpoints—not everyone has the means, experience, time, and budget for this. In fact, most don't.

• Use what you feel works for you BUT know it isn't perfect, it never will be, perfection in 'attribution' is literally impossible

• Try isolated tests where possible

• Test singular elements so not to muddy the water and distort the clarity of the picture being presented

• Track everything over time - don't just focus on the short-term goals. It takes a while to shift brand perception and consumer behavior

• Validate through quant and qual where you can with your actual audience before, during and after an activity

Just please don't jump on the cool kids wagon and say “attribution is dead” without really thinking about the implications of what you are saying. You are literally talking out of your backside. 

Attribution (of a marketing channel to a success/failure metric) in some form or another is here to stay and evolve; warts and imperfections too. Use the various models available to you, they all have their strengths and weaknesses, know and understand their flaws so you can reap their individual rewards. 

Thanks for reading.