How to increase contextual advertising ROI 2.2x for an enterprise retailer

A question that’s familiar to every business is how to cut spending on contextual advertising while maintaining sales. If you purchase traffic to your website in an advertising system such as Google Ads or Facebook, a huge part of your spending goes to clicks of users who don’t end up putting items in the cart. And even if they do, they don’t always complete the purchase. So what can be done to optimize advertising campaigns?

In this case, we describe the solution provided by the OWOX BI team for the electronics and home appliances retailer that tested OWOX BI to calculate the probability of conversions.

Among the main performance marketing tasks for an enterprise retailer are growing efficiency metrics and reducing contacts with users who aren’t currently interested in purchasing. Data-driven solutions are the most promising instruments for optimizing ad campaigns. The OWOX BI attribution model uses machine learning technologies to calculate the purchase probability for each website user starting with the first visit. Using these purchase probabilities, businesses can determine if it’s worth spending on advertising to a user. Furthermore, the OWOX BI model can calculate the probability of a user’s performing any action.

Also, one of the mandatory tasks of a contextual advertising SEM specialist is segmenting website visitors into various groups depending on actions they perform on the website and the time passed since the last trigger. For example, the number of days since adding an item to the cart without a purchase. Manual analysis of such audiences and adjusting bid calculations take up a huge part of the work process.

To sum up, the goal was to increase the ROI of contextual ad campaigns while saving revenue from completed orders.

The OWOX BI machine learning solution was applied to optimize advertising campaigns, being trained on:

• historical data on the behavior of website users
• CRM data on redeemed orders
• aggregated and anonymized data from tens of thousands of projects by OWOX clients

What OWOX BI provided:

• For each website user, the purchase probability is calculated accounting for redeemed orders.
• Probability calculations are updated after each user action or period of inactivity. For example, if a user visits a website and performs a range of actions, they receive X% probability of making a purchase. If they don’t visit the website for four days after that, this probability decreases.
• 10 user segments divided according to purchase probability along a 10-point scale.

To assess the forecast accuracy and the performance of each segment, the team analyzed how many orders each segment has completed.

At the top of the table is the segment with the highest conversion probability, and at the bottom is the segment with the lowest.

1. Collected data about user's behavior on the website in Google BigQuery with OWOX BI. (It’s also possible to use Google Analytics 360.)

2. OWOX analysts trained the attribution model and started regular conversion probability calculations for the website visitors. Each visitor was assigned a conversion probability. The algorithm accounts for more than 60 parameters, including:

• number of sessions in hits within the conversion window
• actions on the website within a session
• pauses between sessions
• overall number of actions
• session device and operating system
• traffic sources the user had within a conversion window
• number of actions on each page within a session
• timing of a specific session and total time of sessions within a conversion window

3. The OWOX BI team gathered all this data into one table in Google BigQuery to evaluate the purchase probability for each user.

4. The audience data was transferred to the advertising service using an OWOX BI Pipeline.

5. The audiences were added to the advertising service.

The primary A/B test was conducted with an equal division of traffic among the examined groups for advertising campaigns:

• created manually.
• based on a product data feed with the help of an automation service.

After the successful experiment, the second test on campaigns with dynamic ads was started. Having gathered data according to its efficiency among advertising campaigns, bids for audiences were calculated and adjusted:

• Decrease bids for audiences with a low purchase probability (sometimes up to 90%).
• Increase bids for audiences with a high purchase probability.

The size of bid adjustments was calculated based on the data gathered about the conversion coefficient of each audience and target ROI. The current ROI of each campaign and audience was also taken into account. For each audience, OWOX BI calculated an appropriate price per click. Based on these calculations, it adjusted the current price to reach the target KPI. The adjustments consistently changed based on the results.

For two months, OWOX analysts worked with the client’s team to conduct two tests. The first lasted a month and a half and included five advertising campaigns with low volumes of traffic. Each of these campaigns was subdivided into two campaigns for A/B testing: one using OWOX BI audiences and one without them.

Campaigns using OWOX audiences showed 1.7 times better results both in terms of ROI and revenue for completed orders. For a retailer, this was a step in the direction of increased efficiency for advertising investments.

After the first experiment, the OWOX BI team ran a three-week test based on campaigns with large volumes of traffic according to the same principle — an equal number of campaigns using OWOX BI audiences and without OWOX BI audiences.

The result of the second test was 2.2 times higher ROI and 2.7 times higher revenue for completed orders in campaigns with OWOX BI audiences.

In both cases, the test group either brought more revenue with the same budget as the control group or brought the same revenue with a lower budget. Using OWOX BI to predict conversions, our client saved time on creating audience segments, all data on website users was processed automatically, and the resulting data was prepared for uploading to an advertising service where it could be used in advertising campaigns to adjust bids and run retargeting.

These test results were considered to be a successful start for further scaling of other performance and brandformance channels. As experiments with conversion probability calculations proved their efficiency, the next step was to scale the solution — to use it in all search campaigns, run retargeting by setting different bids depending on conversions in a segment, and apply it in other advertising services such as Facebook, and Google Ads.