7 eCommerce Attribution Models. What Did We Learn From the Real Life Sales Data?

Eradium Ecommerce Attribution Modeling Feature Image

7 eCommerce Attribution Models. What Did We Learn From the Real Life Sales Data?

Multichannel Marketing and Attribution Models

In this blog, we take a live example of e-commerce sales data to learn how each of 7 eCommerce Attribution Models changes our evaluation of the channel marketing ROI and ROAS (return on ad spend). This kind of learning is crucial for our future omnichannel marketing planning and ROI optimization.

For this analysis, we took traffic and online sale data for two months randomly selected period in 2015. We will not disclose the name of retailer client and the exact dates range. The data set is from a retail business that has both physical and online stores. The brick and mortar stores were contributing almost 90% of total revenue while online share was growing rapidly. We analyzed only the online sales portion and did not include sales that have been initiated online but completed in the physical stores.

The sample includes 2,759 transactions with total revenue of $383,161.

Average revenue per order (ARO) is about $139. The product portfolio is a mix of products in different price ranges: from under $10 to some of them over $1000.

It is evident that the customers would spend more time to research items that cost ten versus few hundred dollars. The buyers of more expensive products were more likely to visit the site multiple times before placing an order.

What is An Attribution Model?

An attribution model (AM) is the rule or set of rules, that determines how credit for sales and conversions is assigned to touchpoints in conversion paths. For example, the Last Interaction model in Google Analytics assigns 100% credit to the final touchpoints (i.e., clicks) that immediately precede sales or conversions. In contrast, the First Interaction model assigns 100% credit to touchpoints that initiate conversion paths. (1)

Multi-Channel eCommerce Attrition Modeling is a significant challenge in commerce marketing analytics and channel budget allocation planning.

Avinash Kaushik wrote in his blog Occam’s Rasor (2): “There are few things more complicated in analytics (all analytics, big data and huge data!) than multi-channel attribution modeling.”

We cannot ignore Multi-Channel Attrition Modeling and rely on oversimplified approach when we credit only one channel for each sales order. If we assign 100% of the credit to the last interaction before the purchase was made then, we make inaccurate and often-misleading conclusions about marketing channels contribution to revenue. As we can see from the data below most of the customers interacted with more than one channel before making a purchase.

How big is the share of sales that are the result of interactions with multiple touch points?

Our data show that 1,888 from 2,659 orders – are outcomes of assisted conversions.

Assisted conversions are the interactions that a customer has with a website leading up to a conversion, but not the final interaction. We see from these numbers that 71% of total sales orders are results of assisted conversion.

Top Conversion Paths

A conversion path is the sequence of interactions (i.e., clicks and referrals from channels) that led up to each conversion and transaction.

How can we find which of the conversion paths contributed the most to our sales? Google Analytics provides Top Conversion Path Report that shows (see our sample below) all Multi-Channel Funnel (MFC) grouping paths and ranks them based on the contribution to the sales revenue.

Eradium Ecommerce Attribution Top Conversion Path

We see from the report that contribution of each path is relatively small. None of the non-direct conversion paths contribute more than 5% into the total number of sales conversions.

How can we evaluate the role of each channel combination in overall conversion? The Multi-Channel Funnels Overview report shows how different channel combinations contribute to the total sales; regardless of the sequence of steps in the conversion path.

You can see below, for instance, that 9.71% of conversion paths included Paid Search and Direct interactions.

Eradium Ecommerce Attribution Assisted Paid Search Direct

The report tells us also that 5.26% of the conversion paths included Paid Search Advertising and Organic search

Only small percentage of total conversion paths (0.62%) included the combination of 4 channels: Direct, Organic Search, Referral, and Paid Search.

Eradium Ecommerce Attribution Assisted Paid Search Organic
Eradium Ecommerce Attribution MultiChannel Conversion

Assisted Conversions Report

The Assisted Conversions report helps us understand which channels mostly convert at the latest step and which channels mainly assist to conversion.

We have to check the values from the last column in this report: Assisted/Last Click or Direct Conversions.

If the values are less than one, than the channel mostly drives the last click conversions.

If the values are greater than one than the channel primarily assists to conversion and it is more dominant at the earlier steps of conversion path. We see from our data that the referral and the direct channels completed more conversions than assisted.  On another hand, the email channel mainly contributed to conversions by triggering the start of the purchase journey. The paid and organic search channels were more likely to assist conversions than convert. It means that customers were more likely to interact with these channels at earlier stages of the conversion paths.

Eradium Ecommerce Attribution Assisted Vs Direct

Analyzing How 7 Attribution Models Represent our Conversion Data

Below we compare seven attribution models to see how an individual model may change our understanding of each channel performance and what can we learn to improve it.

1.     Last Interaction

In the Last Interaction attribution model, the last touchpoint would receive 100% of the credit for the sale.

Below are the results based our retail sample with Last Interaction AM model. We see how the channel ranking based on their contribution to conversion and sales revenue, Cost Per Acquisition (CPA) and ROAS (return on ad spending) changed.

Where ROAS = revenue from ad campaign / cost of ad campaign

This model is the most traditional and in most cases oversimplified view on the online marketing attribution.

Eradium Ecommerce Attribution Last Interaction

2. Last Non-Direct Click

In the Last Non-Direct Click attribution model, all direct traffic is ignored, and 100% of the credit for the sale goes to the last channel that the customer clicked through from before converting. The data shows that channel revenue contribution ranking, Cost Per Acquisition (CPA) and ROAS (return on ad spend) changed from the Last Interaction Model.

We can see that Paid Search CPA has been reduced from $10.47 to $6.99 and ROAS increased from 1,437% to 2,101.87%. The Last Non-Direct Click AM calculates the contribution of the Email channel higher than any other model.

Eradium Ecommerce Attribution Modeling Last Not Direct Click

3. Last AdWords Click

In the Last AdWords Click attribution model, the last AdWords click—in this case, the first and only click to the Paid Search channel —would receive 100% of the credit for the sale.

This model shows that Paid search CPA is dropped to $4.74 and ROAS increased to 3,178.16%.

We have less than half of CPA and more than double ROAS if we compare the Last AdWords Click model with the last interaction model.

Eradium Ecommerce Attribution Modeling Last AdWords Click

4. First Interaction

In the First Interaction attribution model, the first touchpoint would receive 100% of the credit for the sale.

We can see that ranking again changed, with Paid Search now in the third position and CPA and ROAS are close to the Last Non-Direct Click model. The Organic Search shows the highest conversion numbers than all other models.

Eradium Ecommerce Attribution Modeling First Interaction

5. Linear attribution

In the Linear attribution model, each touchpoint in the conversion path would share equal credit for the sale. We can see that the channel ranking is the same in this model as in the Last Interaction model. The Paid Search ROAS is about 28% higher than in the Last Interaction model.

Eradium Ecommerce Attribution MultiChannel Conversion

6. Time Decay

In the Time Decay attribution model, the touchpoints closest in time to the sale or conversion get most of the credit. In this particular transaction, the direct channel would receive the most credit because the customer interacted with it within a short period before conversion. Other channels would receive less credit. In cases when the Paid Search interactions occurred earlier, this channel would receive significantly less credit.

The interesting finding is that over 25% of conversion paths take 12 -30 days to convert.

Eradium Ecommerce Attribution Time Decay
Eradium Ecommerce Attribution Time Lag Report

7. Position Based

In the Position Based attribution model, 40% credit is assigned to each the first and last interaction, and the remaining 20% credit is distributed evenly to the middle interactions.

Our findings for the Position Based AM are relatively similar to the Time Decay model. The email and direct touchpoints have slightly more differences in comparison to the Time Decay model

Eradium Ecommerce Attribution Position Based

Lessons Learned

  1. Why do we need attribution modeling? Based on our data that represents a typical online retailer – only 30% of sales are results of one channel interaction. We have to use attribution modeling to understand our multi-channel marketing performance for 70% of our sales. Ignoring this fact would lead to an entirely wrong decision about our marketing budget allocation.
  2. Attribution modeling is both art and science at the same time. The ideal attribution model does not exist. We need to look at the models in the context of our products segment and conversion paths patterns.
  3. We practically cannot focus only on one main conversion path because the contribution of each path is relatively small (less than 5%).
  4. The Attribution Modeling helps us evaluate ROI of the channels that are more assisting to conversion than actually converting and find ways to improve overall conversion at later stages.
  5. We discovered that we would underestimate the email channel significantly if we don’t use attribution modeling. Email played much bigger role in starting a journey through conversion path steps.
  6. Organic and paid search advertising contribute more at the earlier stages of conversion paths and are underestimated with the last click attribution.
  7. We can improve overall conversion if we supplement more initial stage channels with display ad remarketing or conversion focused emails.
  8. The last but not least lesson – we must continually experiment with our channel mix, measure results and use attribution modeling to improve our outcome. If we make a daily improvement of 1%, it will lead to 3778% increase in one year.


  1. Attribution modeling overview, https://support.google.com/analytics/answer/1662518?hl=en&ref_topic=3205717
  2. Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models, http://www.kaushik.net/avinash/multi-channel-attribution-modeling-good-bad-ugly-models/
Igor Nesmyanovich, Ph.D., CISSP

Igor Nesmyanovich, Data Scientist and CEO of Eradium. Igor began his career as a space scientist, and for the more than two decades applied the unique art of science to the emerging digital world. Igor is a Certified Information Systems Security Professional since 2004. Today Igor’s focus is on helping clients to realize business value from using data science and technology in their core business operations and marketing.

To start a dialog about your next eCommerce initiative