Found.

We’re Found, the digital performance experts who help brands get found online. Working alongside our sister agencies, Disrupt (Influencer Marketing) and Braidr (Data Science & Analytics).

London, United Kingdom
Founded: 2005
Staff: 48
More
Skills
Search Engine Marketing
Digital Marketing
digital strategy
SEO
Performance Marketing
Google Analytics
PPC - Pay Per Click
Influencer marketing
Content Creation
Video

and 12 more

Clients
Randstad
Amazon Developers
Marshall Motors
Fender
Hoppy
Bonmarché
Deloitte
Unilever
Encore Tickets
Puma

and 17 more

Sector Experience
Financial
Retail
Travel
Ecommerce
Charity
Cosmetic / Beauty
Technology
Business to Business
food and drink
FMCG

and 4 more

Less

How Braidr helped a large e-commerce retailer generate 25% more ad revenue from thousands of dormant products

by Otilia Crizbasan

Situation

Our client has doubled in size yearly for the last two years and aims to achieve this again in 2022. The vast size of their product catalogue, a result of this growth, created a problem. The E-commerce retailer used automated biding on Google to promote their products to customers.

Simply put, Google spends a small amount of the overall budget on each product, sees how well it does and then decides whether or not it’s profitable to put more budget around it.

And here’s the problem - when you have the volume and velocity of stock our client has, then Google Ads' automated bidding technology takes too long to acquire enough data to push new products. Consequently, revenue from new products during their critical initial release period was much lower than it could be and often missed out on for prolonged periods of time (or even entirely).

As a result, the automated biding was withdrawing the budget from most products. Which became dormant or ‘zombified’ - not being bid on and not appearing in any ads.

Challenge

Our objective was to surface these ‘Zombie SKUs’ and deliver a 10% increase to the total revenue through Google Ads shopping activity. Whilst also achieving a return on ad spend (ROAS) that was no lower than existing shopping activity - creating genuine incremental growth from the existing portfolio.

There was no unique target audience as this campaign focused on delivering more inventory to all potential customers. Typically though, our target audience was 20-40 years old.

We started by analysing our client’s historical sales data to study how profitable the different products and variations (size, colour etc.) had been for the e-commerce retailer. We aimed to predict the gross merchandise value (GMV), the total value of merchandise sold over a given period.

The main challenge we faced was the volume of data that we had to analyse and train a model with. We were working with over 23 million data points from previous product sales.

Solution

To overcome this challenge, we leveraged BigQuery machine learning to engineer a deep neutral network powerful enough to predict how much GMV a product would generate in a given 30-day period, accurate to within £100. This significantly increased the speed of our model development and the level of complexity that we were able to build into it to achieve the desired level of accuracy,

In other words, we were now able to predict how much revenue could be made from dormant or zombified products by surfacing them through paid ads.

Models are meaningless unless they can be operationalised by marketing teams so we devised a phased activation strategy to guide bidding strategies. We made our model available to our client through a secure, easy-to-use web app. This provided:

Complete and transparent access to the models’ predictions

Empowering Found and our client to immediately determine the sales potential of new products

Wasting no time in data collection and enabling swift action

Our model was proven to reliably predict the next 30 days of revenue for any individual product (including colour & size variants) to an accuracy of roughly +/- £100.

The by-product of this was that we could see-under performing products, tier the scale of their underperformance, and treat them accordingly.

Three separate shopping campaigns, with unique bid strategies according to the expected potential of the products contained:

High performing products: if we had products that would predict to perform very well in the next 30 days we would be bid on them quite aggressively and put everything behind them

Mid-tier products: for the products that were not doing fantastically well the paid team built a granular campaign

Low performing products: for products that were predicted not to do anything then we put tight constraints on them just to make sure they don’t become unprofitable if we’re bidding on them

Results

We aimed to deliver a 10% incremental increase in total revenue through Google Ads shopping and maintain existing ROAS. And so far, we have achieved:

28.0% higher ROAS than all other shopping campaigns

4.5% uplift in the overall ROI of the entire Google Ads account

24.4% (> £1M) incremental increase in profits driven by model prediction

Tags

data science
deep learning
ad revenue
dormant products

Clients

Large e-commerce retailer