Customer journeys on an advertising champaign. Each journey is composed by a chronological sequence of actions by a user on three advertising channels, including display ad, social ad and paid search ad. Source:

How to evaluate marketing attribution models ?

Krishna Kumar K


Marketing attribution is the process of determining which interactions influence a customer to purchase from your brand. There are many attribution models available starting from simple last-click models to more advanced multi-touch attribution models.

Types of marketing attribution models —

How do we decide which model to pick ? How do we decide if the complexity of evaluating and implementing attribution models is worth the effort ?

As a business, you allocate your budget across channels to maximise certain goals (eg. sales, conversions, customer acquisition, account creation) with certain guardrails (CostOfSales, CostPerConversion, CustomerAcquisitionCost etc)

There are some organisational and political considerations that drive many decisions. For example, marketing team wants to maximise their budget and spend it completely. Teams that own individual channels also want to maximise their budget and spend it completely. These are very important to consider when we have to navigate big organisations. Let’s imagine a theoretical business (an e-commerce site) where all decisions are data driven and the entrepreneur has visibility to most marketing-mix decisions.

Step1 : Allocating marketing budget, aligning on goals

This step depends a lot on the strategy that the company wants to pursue based on the stage of its business, ability to raise funds and state of the economy. It could be

  • Grab marketshare. Maximise revenue at all costs


  • Maximise revenue keeping profitability into consideration.

Let’s assume that it is the latter considering the sentiments prevailing in current market conditions.

If you have an estimated curve on spend and gross profit, you can pick a budget that maximises (gross profit — marketing spend).

Here a valid question is — do you estimate overall Gross profit or the one that is attributed to paid marketing channels ?

  • Ideally the business has to maximise overall gross profit hence estimating that value is better

Another valid question is — How do you estimate the curve ?

  • Based on past performance data of various channels
  • By doing small experimental campaigns on various channels

You don’t need the entire curve to set a budget, but few points in the relevant areas would be sufficient.

If this sounds too complicated, the team can fall back to the popular method of allocating a percentage of sales revenue to marketing.

Step 2: Maximising gross profit for a budget / Increasing the efficiency of the curve

Now the next step is to maximise gross profit for a given marketing budget, by optimising allocation across channels.

It is important to note here that, each channel will have its own performance curve and saturation point. In addition, the channels also will have interactions with each other.

A common rule of thumb is to allocate (incremental) budget to channels that has maximum (incremental) ROI = (inc)Gross Profit /(inc)Spend. Similar to gradient descent method of optimisation. (except that this is an ascent)

Depiction of gradient descent

It is here that marketing attribution models can make a difference. To measure the ROI of a channel, we need to measure the sales that can be attributed to it. As described in the image, a customer’s interaction with different ad copies across channels add up to the final decision of buying a product from the business.

This illustration depicts the business value of a better attribution model quantified as the increase in gross profits for the new optimal allocation over the old one.

However the attribution model is just one ingredient in the entire system. For researchers who develop new attribution models, they need to pick a methodology that can be used to create the attribution model as well as independently validate it before it is used in the allocation decision.

Developing and evaluating algorithmic attribution models based on conversion performance

Algorithmic attribution models try to model conversion performance based on the user’s interaction with various channels. For example a very commonly used model — logistic regression can be used to generate a prediction model of the user converting.

In this case, the evaluation is also based on the accuracy of conversion prediction.

Organisational challenges

As we can easily infer, the choice of attribution model can change the sales attributed to each channel. So if the attribution model is picked by the paid marketing team, they could be incentivised to pick one that attributes more credit to their channels (eg. last paid click).

Secondly very few decision makers want to take the pain to understand the attribution models and calculate the ROI of implementing a better one. In fact, the optimal marketing allocation strategy can be overruled due to various other strategic and personal reasons.

The verdict

As more and more firms are looking at profitability, it is becoming increasingly important to measure and spend each dollar. Attribution is an important part of spend optimisation that is often overlooked.

About me:

I’m a product leader and technologist with experience in ML/AI and analytics based products on search ads space (Microsoft). DM me on LinkedIn to collaborate on related topics.

Similar articles:




Krishna Kumar K

Product Guy. (Worked at Indeed, Microsoft ...). I write about product management, startups, analytics and machine learning. Occasionally I digress...