Introduction
Vertex AI Recommendations is one of the first publicly available Google algorithms that allows retailers to use the same intent-based search and prediction algorithms pioneered by Google’s search, ads, and YouTube engines. Many legacy ecommerce recommendations solutions simply recommend products to shoppers based on popularity or crowdsourcing purchase data. While this can be effective in some use cases, Google’s algorithms can take your website a massive leap forward by learning from individual customer browsing activity across devices, pages, and channels. Then, after learning from this activity, the AI models can prioritize recommendations that are personalized to every single shopper at all of your touch points.
When compared to baseline recommendation systems, retailers can expect up to a 90% increase in clickthrough rates, 50% increase in revenue, and 40% increase in conversions driven by recommendations. On many stores, this can result in a 5% total lift in revenue per session.
Google provides several AI models to choose from, and this article explains the best practices to look out for to benefit from the largest lift in revenue.
A Note on Serving Configurations
Each Vertex AI Recommendations model can be connected to one serving configuration.
The Three Steps to Creating AI Models
1. Select Recommendation Type
Vertex AI Recommendations includes several pre-built models that have been optimized through years of machine learning research at Google. Each model has been created to understand your customers’ journeys and their shopping intent, and each model should be placed on a different part of your website depending on your goals.
The built-in AI models are:
Recently Viewed Items
Recommended for you
Others you may like
Frequently bought together
Similar items
2. Identify a Business Objective
Each AI model type allows you to select different business objectives based on the required input data for model training. Generally, the business objectives to choose from are:
Click-through rate
Conversion Rate
Revenue per session
This business objective should be based on your wider business goals. For example, if your organization prioritizes higher average order values, select revenue per session. These are simple to change and you can create several models with different goals. However, each AI model can only be associated with a single serving configuration.
3. Add Custom Rules (Optional)
Each serving configuration associated to an AI model can be configured to filter or change results based on business or catalog requirements. For example, a rule can be added to filter out-of-stock items, diversify or personalize results, or re-rank prices.
For best results, we recommend starting your models without custom rules unless they’re required. Traditional merchandising has required a lot of human input to provide effective recommendations, however, Google’s AI models operate very well without human intervention. If the recommendations provided by Discovery AI still require adjustments after they have been deployed for a few weeks in a production environment, then add custom rules as guardrails to improve results.
A Note on Model Training Frequency
Training a model unnecessarily can create additional Google Cloud costs. We recommend using the recommended default settings for each model and pausing training if your catalog data or storefront traffic becomes stagnant.
Where to Use AI Models
There are several combinations of Recommendations AI models and website locations that benefit adding personalized recommendations. You can either experiment with locations or attempt to define them prior to implementation. Regardless, a good A/B testing procedure is paramount to measure success.
Any Page
Two models, Recommended for You and Buy it Again can be placed on any page of your website. Both of these models are tailored to each individual user.
Home Page
The Recommended for You model is designed primarily for home pages. Outside of your ecommerce platform, it can be used in custom email campaigns, input for personalized ad campaigns, and social media.
Product Pages
The Frequently Bought Together and Others You May Like models use the data from your product pages to categorize items.
Add-to-Cart Pages
If you have a separate add-to-cart page or modal, Frequently Bought Together will be the best choice.
Checkout Pages
Similar to add-to-cart pages, the Frequently Bought Together model will use the product IDs in a customer’s cart to recommend the best products for them.
Email Marketing
If your organization uses email personalization tools, you can include Recommended for You and Others You May Like predictions in each email based on their visitor ID.
Nimstrata Connect on Shopify
If you’re using the Nimstrata Search and Discovery Shopify App, you can use app blocks to add any of these Recommendations AI models to your online storefront themes.
Simply add the Recommendations AI app block to any of the pages recommended above and follow the instructions in this article to successfully return results in minutes. The App Block will automatically detect the products and page types to return the most relevant results.
