Summary
Ardennes-Etape made it their mission to help people discover the marvelous Belgian Ardennes. They do this by connecting local homeowners and vacationers in order to create a magical holiday experience. 39% of the Belgian population books their holiday accommodation online. You envision the holiday shoppers sitting at home in their pajamas tapping on their phones. For marketers, this means that competition for online shoppers is intense. For Ardennes-Etape, it means maintaining dominance versus the big international players like airbnb, booking.com, Belvilla and others. Yet, that doesn’t mean Ardennes-Etape should throw money at this problem though. Semetis helped Ardennes-Etape to win the holidays by executing data-backed media buying while everyone else is throwing money around.
Why Ardennes-Etape needs to become a smart holiday advertiser
Let’s start with the context. Even though Ardennes-Etape is the #1 player in the Belgian market, they see increasing competition from larger (international) players such as Airbnb, booking.com and others. In order to remain in their top position, Ardennes-Etape had to figure out how they could compete with the big boys knowing they only have a fraction of the marketing budget. To tackle this challenge of becoming a smart holiday advertiser and thus not tolerating wasted advertising spending, Semetis built a cloud infrastructure on top of their adtech.
Borrowed from corporate finance, we decided to use the incrementality concept as a problem-solving approach that applies buying intent information to decision making. What we see is that website visitors that don’t complete their booking are all too often seen as “hot leads” that should be chased. We asked ourselves the question if they really are hot leads and if they actually should be chased (often till the end of their days). We built an incremental analysis that considers the opportunity costs, being the missed opportunity when choosing to chase the lead or not. The important end result of the incrementality analysis is that it helps us avoid delivering ads to users who are not predisposed to book their stay with us, thus pursuing only the most favorable leads.
It is hard to tell who’s ready to go down the funnel and who is not, but signals do exist in the behavioral data from individual users. Yet, using those signals for your incrementality analysis is not a walk in the park. Looking at visitor data we have millions of data points. The data overload is simply too big to process using spreadsheets or local machines. Moreover, the large amount of dimensions representing the behaviour of the customers throughout their sessions makes it nearly impossible to be manually investigated and represented in a comprehensive manner. And then we have not even talked about activating this data yet.
In a world where less factors need to be taken into account, decision making would be simpler. We could easily script a rule stating that if a user looked at 2+ pages, he or she would be flagged as predisposed to buy or not (hot or cold lead). However, today we’re collecting vast amounts of granular data which renders this impossible.
Therefore, we used a complex data infrastructure that was architected in the Google Cloud Platform which helps us understand underlying patterns that were invisible to the human eye. All of this was done by leveraging artificial intelligence capabilities. And the result, a model that is able to accurately predict if a user will be converting or not in their next visit to the website.
Adding Google Cloud functionalities in the adTech ecosystem to conduct incrementality analysis
Incorporating a robust data architecture easing data analysis was a real game-changer for Ardennes-Etape (see annex 1). The immediate output of this infrastructure is a solution we call user scoring.
At the end of each business day the user data of that day is run against our predictive master model. When the data is ingested, the model returns a score for each user between 0 and 1. Here, 0 indicating the chances of a user converting to be very low or 0% and 1 indicating the chances of a user converting to be 100%. This score can be interpreted as the propensity for each user to convert. The higher the score, the hotter the lead (see annex 2)
Afterwards, this predictive user score is pushed into Google Analytics where 2 different audience segments are created. On the one hand a segment that includes the hot leads with high chance of converting (score of 1) and on the other hand a segment for cold leads with low conversion likelihood (score of 0).
Now, having this information at our fingertips opens up numerous possibilities for activation. Within the field of media activation we have succeeded to impact performances in all layers of the funnel:
- Conversion: Cut the fat in campaign remarketing spending - Activation of the audience segments in paid campaigns to increase effectiveness of marketing spending in remarketing. Because of the API connections we can push the audiences in all digital advertising channels being Google Marketing Platform, Google Ads, Facebook and Instagram
- Consideration: Prequalifying Acquisition traffic in campaigns - By using the high likelihood converter audiences as a seed, we can find new profiles with similar characteristics, reducing greatly the noise before people even enter our website
- Awareness: Building data-driven personas for higher funnel marketing - Based on the user segments we are able to define the customer personas better giving us a more accurate direction for media mix selection, content creation & message personalization.
Embedding Semetis data science teams into Ardennes-Etape marketing teams
What was our contribution of media expertise? Due to the technical skill set and the data scientists at Semetis, we succeeded in activating predictive modeling in our media buying. We build a Cloud infrastructure bridging all platforms together: web analytics, CRM and advertising channels. We started by collecting raw data in Google Analytics 360 from the dataLayer and then exported it to BigQuery for storage and analysis. After, BigQueryML is used by our data science team to build and run custom data models and algorithms driven by machine learning capabilities. These models correlate the usage of filters to the booking of a holiday vacation home. In addition, Google Cloud Feature Importance allowed our data science teams to generate predictions about the probability that a user would book a stay based on the filters they used in the booking engine. All of this resulted in a label being attributed to the user where audience segments could be created on. By leveraging the different APIs of the advertising channels, these audiences could then be activated.
The integration of predictive modelling in our media buying was done gradually. We assessed where the user scoring audiences could generate the largest impact. In this case, search campaigns were selected as the first testing grounds. Knowing that search advertising is a highly converting channel that focuses on generating conversion close to the bottomline (here, bookings), these channels are an absolute priority for us and have a direct business impact by optimizing our marketing spent here.
We applied the methodology of A/B testing to measure the incrementality. We created test versions of our campaigns where we excluded the low likelihood converters (cold leads) audiences and ran the test versus a control group consisting of our normal campaigns. This allowed us to accurately see if there was an effective uplift in performances using the predictive audiences.
The results: getting much more from your ad budget
The activation of the user scoring audiences enabled us to accurately remove 159.000 (19%) non-converting users from our remarketing audiences. This means 1 in 5 people are now effectively labeled as ‘cold leads’ and excluded from our campaign, thus not receiving any more ads from Ardennes-Etape.
Following the search campaigns experiments we concluded that optimization of our search marketing spend by removing those website visitors impacted our campaign performances to an amazing extent. Doing so, we were able to:
- Reduce the investment on the campaigns by 9% while actually increasing its conversions by 7%.
- Increase our conversion rate by 11% and decrease our cost per acquisition (CPA) by 15%.
- Increase return on Investment (ROI) by 17% on the bottomline.
All of this resulted in over 40.000 EUR in marketing spend saved on a yearly basis which is huge. Re-investing this amount further into the higher funnel campaigns ultimately drives our ROI up even further.
This is however not the end. Given that these audiences are built within Google Cloud Platform, it makes it possible to push them through the APIs of other marketing channels such as social and display advertising. The incrementality analysis is thus an effort that is scalable across channels and the whole customer journey.
Why does the case deserve the best use of data & performance of the year?
The project presented here above showcases the perfect example of how Ardennes-Etape and Semetis were able to set in place a robust data architecture linking website data and machine learning to tangible activation and performance improvements in our media buying. By leveraging cloud functionalities and a data-driven way of working we were able to act and improve the effectiveness of our campaigns fast, resulting in a significant positive impact on the bottomline of Ardennes-Etapes business. Moreover, this case is picked up by Google EMEA as a flagship case in the field of cloud for marketing and it is a first real example of how cloud computing can drive results in the domain of marketing.
Jeroen Cools
| LinkedinThis email address is being protected from spambots. You need JavaScript enabled to view it.
