Challenge: The customer, a prominent US affiliate marketing network, maintains and fosters close relationships with its affiliate publishers. While the publishers were already carefully managed by Publisher Development Managers in individual portfolios, the high numbers of collaborators made it challenging and time consuming to reach out to each publisher with the right message.
We pioneered a cutting-edge recommender system that transformed the network’s affiliate marketing landscape. Our data-driven approach empowered publishers to effortlessly discover the most suitable and lucrative offers, while enabling publisher managers to focus on cultivating stronger relationships with less concern for publisher compliance.
This personalized experience increased affiliate engagement and earnings, driving additional network growth. Moreover, it benefited advertisers by ensuring their offers were surfaced to publishers similar to those who had already succeeded with the campaign, increasing the likelihood of strong performance across a wider range of publishers.
By harnessing the power of advanced analytics, the network solidified its position as the go-to platform for unrivaled affiliate success for both publishers and advertisers. Here is how we did it:
We have set up a data analytics architecture that involved connecting to the Affiliate Marketing Tracking Solution they were using, Everflow, to extract the underlying data.
Given the high volume of data, we have firstly brought the information in a MongoDB database and then we have flattened it into a MariaDB database, that we later connected to their Domo Business Intelligence platform.
We then built an RFM segmentation analysis to segment their affiliates according to the numbers of offers each publisher is running, the recency of the last conversion they have generated and the average monthly gross profit each publisher is generating.
Here is a preview of the underlying calculations:
Please see below a preview of the RFM analysis:
After performing the analysis, we noticed there was a segment of publishers that were running a small number of offers which were generating a high monthly gross profit. Our next step was to build a recommender system that generates highly personalized and targeted suggestions for them based on the offers that were performing well for other similar publishers.
In order to achieve that, we built an item-based collaborative filtering recommender system that runs once per week and determines should be recommended. It analyzes similarities in revenue patterns and campaign successes across different publishers to provide tailored recommendations.
In the next step, the generated recommendations are filtered to ensure that business and compliance rules are enforced. A few examples:
For new publishers with no historical data, our initial recommendations consist of the offers that have generated the highest gross profit.
We know machine learning algorithms are often seen as a black box and that users generally have a very hard time explaining the reason behind the recommendations. This is why, to ensure transparency and visibility, we have built a web app that the Publisher Development Managers can use to understand the underlying process and steps the algorithm has gone through to arrive at the final recommendations.
We have also implemented a feedback form for the Publisher Development Managers to be able to provide continuous feedback that we could use to refine the algorithm.
In order to allow the customer to use the recommendations and embed them in the Everflow platform, we have built an API that serves these recommendations.
Using the data provided by us, the customer then has built a “Suggested for You” Tab available for each publisher upon the login in the platform where they can visualize and apply for their personalized offers suggestions.
For attribution purposes, we have tagged and monitor the adoption and usage of the recommender system. As such, we can confidently show that, since the first six months from its launch, the recommender system has generated well over $50K in additional gross profit.
For reference, implementing a solution like this from scratch costs somewhere in the range of $30K – $50K, so the payback time of our recommender system is less than 6 months, which makes for a great investment. These are high level estimations, if you are interested in a personalized offer, do not hesitate to contact us.
MongoDB, Python Sklearn, Uvicorn, FastAPI, MariaDB, Domo
©2024 All rights reserved