Who can say they never impulse buy? You’re looking for a set of power tools online, only to have the recommendation system suggest a model airplane, like the one you once built as a child. How did the store know you would be interested in buying one? If you want to learn about the power of collaborative filtering algorithms and an AI powered recommendation system, keep reading. Find out how to use Recommendations AI in your business to drive up sales.
Online shopping and impulse buys
Moving from stationary stores to e-commerce means a change in customer needs and behaviors. It also requires retailers to adapt.
Instead of driving up sales by using clever displays and in-store placement that would capture your customer’s attention, you need to meet the customer in their own home, in front of a screen. Oftentimes it is a screen as small as the mobile phone, with little room to display relevant content.
Luckily data science is at your service with machine learning models that analyze user’s preferences and user’s past behavior.
In order to increase both sales and customer satisfaction from their online shopping experience, companies analyze user data and use recommendation systems. They increase user engagement and ultimately, sales.
The power of personalization
Personalization is at the forefront of these changes – up to 90% of consumers say they are more likely to purchase products from brands that provide them with a personalized online experience.
For owners of online stores this is a huge opportunity. Automatic recommendations reduce the need to navigate between shelves, whether in a stationary or online store, save shoppers’ time and increase their loyalty and conversion rates.
The key is an appropriate behavior prediction engine that will take into account the customer’s route, the actions he is currently taking and compare them with the goods in stock and current promotions. And ideally, it should react to changes within a few seconds, as well as account for new users. Google Cloud’s Recommendations AI is one of the most commonly used recommender systems. Let’s take a closer look.
What is Recommendations AI?
Recommendations AI is a recommendation engine available on Google Cloud infrastructure, combining data on user behavior with artificial intelligence models. The service uses data on product availability and changes in the offer, promotions, historical records of user actions and information on current actions. Machine learning algorithms allow it to recognize patterns in individual users shopping behaviors. It continuously analyzes the data, serving tailored suggestions within seconds in response to customer behavior.
Recommendations can be displayed through various channels, including: inside the website, in mailing campaigns or online advertisements.
For years, Google has been committed to providing the most relevant results to its users – whether in the most popular internet search engine, on YouTube or in Google Ads. The same mechanism has been made available to owners of e-commerce websites.
A fully-managed service
Recommendations AI is a fully managed service – to use the engine, you do not need to write code or prepare your own machine learning models from scratch. Google Cloud ensures proper training and tuning of models, and also deals with load balancing, load handling, scaling and updating. The service is launched in three steps via a transparent and intuitive panel in the GCP console. The engine can be integrated with data from other Google services (e.g. Google Tag Manager) or imported from an external database.
Your own recommendation engine
The example shared by Kathy DePaolo – VP of Engineering at Disney during the Google Cloud Next conference shows the power of recommendation systems. Kathy compared the results presented by Disney’s proprietary engine with the results of Recommendations AI.
After adding a sleep blindfold to the cart, the Google model read the user’s intention and suggested sleep-related products – pajamas, slippers, stuffed animals. However, the original engine limited its recommendations to the category of carnival masks and suggested products that, as Kathy said, “no one would probably want to sleep in.”
This is quite a vivid example of how difficult it is to create your own recommendation system. Of course, building and training your own artificial intelligence model is possible – but it requires a ton of data, time and a sizable budget.
What makes it work?
Below is a comparison of the mechanisms recommender engines use.
“Legacy” engines most often propose similar items from the same category or are based on the choices of other customers.
The Recommendations AI engine also takes into account the user’s search history and previously browsed offers. After adding a dress to the cart, the recommendation of a watch may seem inappropriate – but not if the customer has previously browsed this product category. The Google engine focuses on the user’s intention and history of contact with the brand, while other engines focus on the physical relationships between products.
Personalization is not easy
Why is it so difficult to create your own behavior prediction engine? WITH surveys conducted by Yieldify among 400 e-commerce directors and senior marketers shows that the most common difficulties in implementing offer personalization are:
- insufficient skillset among employees – 37%,
- limited tool functionalities – 36%,
- insufficient time – 35%,
- too expensive tools – 34%,
- difficulties in obtaining valuable data – 34%,
- fear of excessive, artificial personalization – 29%,
- difficult to determine return on investment (ROI) – 24%,
- the complexity of the project – 23%.
Benefits of Google’s Recommendations AI
Google’s proposal appears to solve most of these problems. Many areas related to the recommendation system can be automated – for example, periodic model tuning or importing assortment data. The tool is launched from an intuitive interface, where you can also make changes and track the results. The cost of the service is predictable and the settlement method is transparent.
And the return on investment? Google Cloud indicates that, compared to other prediction engines, Recommendations AI generates increases in metrics such as click-through rate (CTR), conversion and income:
Results of the service compared to a simple recommendation engine among pilot projects, obtained by A/B testing. These numbers are not guaranteed – results may vary between e-commerce sites.
Examples of companies using Recommendations AI
Sephora – increasing CTR by 50% and conversions by 2%
Sephora is an international chain of stationary stores and a global e-commerce website with cosmetics and personal care products. The brand ensures a high level of experience for its customers – in-store, employing service specialists and consultants, and in the digital channel, using Recommendations AI.
According to Jaclyn Luft, Manager Site Personalization & Testing at Sephora, the implementation of the Google engine resulted in a 50% increase in the click rate on product pages and a 2% increase in the entire conversion rates on the home page compared to the previously used mechanism. Currently, Sephora is testing GCP’s machine learning capabilities at other customer touchpoints. This includes the transaction process or e-mail marketing campaigns.
Hanes Australasia – from testing to production in a month
Hanes Australasia is an e-commerce company that brings together many popular Australian clothing and lifestyle brands. The platform’s developers initially implemented Recommendations AI for over 10,000 products, and the transition from the first launch of the service, through training and tuning the models, to the release of the production version took less than a month.
Peter Luu, Online Analytics Manager at Hanes Australasia, says that after introducing ML technology, they quickly saw a “double-digit increase in revenue per session” compared to the mechanism they previously used. Peter also says that the engine is excellent at handling new products that have just been added to the service’s offering, and that the service provides a better understanding of customer needs and habits, which is valuable to Hanes Australasia contractors and partners.
Qubit – 5% more revenue on each transaction
Qubit provides advanced e-commerce solutions that support sales and build customer loyalty through personalization. He cooperates with global brands, including Topshop, River Island and MAC Cosmetics.
While working with one of the clients, they tackled, as Qubit CEO Graham Cooke says, the most difficult area to increase conversion on the customer path – the shopping cart, the last step before making a payment. Graham says that this point in the path is so critical that an incorrect recommendation can even be harmful. It can distract the customer who was ready to proceed to payment a moment ago.
The implementation of the product carousel in the basket with Recommendations AI suggestions not only did not harm, but also increased the website’s results. Qubit’s CEO claims that this solution, regardless of the website or industry (fashion, cosmetics, jewelry, luxury brands), has always equaled an average 5% increase in revenue on each conversion.
Starting the service in 3 steps
Enabling the service is done via the transparent GCP interface and consists of three steps:
- Import data,
- Create a model,
- Select where to display recommendations.
1. Import data into Recommendations AI
The first step is to import data regarding the product range and user behavior. If you already use Google tools – Google Tag Manager or Google Merchant Center – just integrate the engine with these services and the data will be downloaded automatically. You can also import data from Cloud Storage or BigQuery.
If you have information about the assortment and events in external tools, the data should be imported using the catalogItems.create method or the catalogItems.import method for large catalogues.
2. Creating a model
The next step is to select the type of model, adapt it to the needs of the website and set efficiency indicators.
We have three types of recommendations to choose from:
- Others you may like – the model predicts which product the customer is most likely to interact with next by analyzing the history of previously viewed and added products. The model requires 10,000+ homepage views and 10,000+ product page views or 10,000+ add-to-cart events within 90 days. Google recommends displaying these recommendations on a specific product card.
- Frequently purchased together – displays products that were frequently purchased by other users during one shopping session. The model requires a list of 1000+ purchase-related events throughout the year. It is best to include recommendations in the message after adding the product to the cart or at the stage before making the payment.
- Suggestions for you – the model, using information about the history of products searched and viewed, displays product recommendations to the user before he or she starts shopping. Requires 90 days of data with 10,000+ product card views and 10,000+ home page views.
The service offers three ways to measure effectiveness. Depending on the model, these are:
- Click-through rate (CTR) – a measure of user engagement through the number of views of recommended products (available for the “Other you may like” and “Suggestions for you” models),
- Conversion rate (CVR) – percentage of recommended products added to the cart (for the “Other you may like” and “Suggestions for you” models),
- Revenue per session – revenue generated as a result of proposing products (indicator available only for the “Frequently purchased together” model).
3. Select where to display recommendations
In the last step, we indicate where recommendations should appear on the e-commerce website. This can be the home page, a product page, a popup after adding an item to the cart, or the cart itself.
Recommendations can be included in email campaigns using a script. Here is the relevant documentation describing the process: Using recommendations in emails.
How much does Recommendations AI cost?
The service incurs the costs of training and tuning models and the costs of handling recommendation requests. There is no fee for importing event data or product catalogues.
Model training and tuning is paid per hour per node and costs $2.5 per node per hour. The fee is charged for working time; if training/tuning is interrupted or the model is deleted, you only pay for the computing power used. When work resumes, fees are charged again at a fixed rate.
The cost of handling prediction requests is charged in batches of 1,000 recommendations and is divided into three levels depending on the monthly sum of all requests:
- up to 20,000,000 requests per month = USD 0.27 per 1000 requests,
- another 280,000,000 requests per month = $0.18 per 1000 requests,
- over 300,000,000 requests per month = USD 0.10 per 1000 requests.
Prices may vary depending on the region or currency.
To better understand the service billing models, below are examples of monthly costs in large and medium-sized e-commerce.
Example of settlement in a large e-commerce
Let’s assume that a popular e-commerce website is visited by several dozen million unique visitors per month. The service will serve approximately one billion recommendations to suit the taste of every customer.
The website owners launch three models for training. The models are trained once a day, and in a month this is the sum of 500 hours of node operation:
- 500 node hours = 500 * $2.5 = $1,250
The trained models are fine-tuned once a quarter and in total it takes 300 hours of node operation. For calculation purposes, let’s average it to 100 hours per month:
- 100 hours of node operation = 100 * USD 2.5 = USD 250
The settlement for handling one billion requests will look like this:
- first 20,000,000 recommendations = 20,000,000 / 1000 * USD 0.27 = USD 5,400
- next 280,000,000 recommendations = 280,000,000 / 1000 * $0.18 = $50,400
- remaining 700,000,000 recommendations = 700,000,000 / 1000 * $0.10 = $70,000
Total monthly costs – training, tuning and returning recommendations – will be $127,300.
Example of settlement in medium-sized e-commerce
Assuming that the website is visited by nearly 100,000 unique users per month, the service should return approximately 10,000,000 recommendations.
The website owner runs one predictive model. Training takes 150 hours of node operation:
- 150 node hours = 150 * $2.5 = $375
The model is fine-tuned once every three months and it takes 90 hours of node operation. On average, it will be 30 hours per month:
- 30 hours of node operation = 30 * $2.5 = $75
The number of recommendations does not exceed the lower limit, so the settlement will be as follows:
- 10,000,000 recommendations = 10,000,000 / 1,000 * $0.27 = $2,700
The total monthly costs for serving personalized recommendations to portal users will be USD 3,150 per month.
