The process of product discovery helps users to find the right product and this process involves finding both the relevant set of products and ranking them in the right order. This helps the eCommerce business to adjust the ordering of the products and hence optimize the shopping experience for higher expected revenue.
Shoppers not only look out for the relevant products but they also consider the popularity and the rating of the product. They give importance to how other shoppers perceive the product and the feedback received from them. The shoppers would always want to buy the most popular product in any eCommerce store and hence for all practical reasons, online stores are always looking to promote the best selling products.
In view of achieving this, the ranking of the products becomes important to the merchandiser as they would want to promote products with better profit margins or with better discounts to improve conversions. Ranking the relevant products in the order of their likelihood of purchase then becomes inextricably connected steps.
With the advent of artificial intelligence and machine learning, eCommerce businesses want to automate all such processes for their business. In general, most of the AI products remain a black box to the user with almost no controls being exposed to the users. But, they do still lookout for some controls to optimize the functionalities and control how the search engine ranks the relevant products. With the new Unbxd Ranking control, the merchandisers/analysts would be able to optimize the ranking of the products based on their use cases and business needs.
These ranking controls can also be overwritten through the merchandising options that Unbxd product gives to the merchandisers. The merchandiser can solve for the various custom business use cases using the merchandising section of the Unbxd console. You can read more about The New Unbxd console here.
What are the different ranking scores and its controls?
Unbxd Search Ranking generates several scores to rank the products for a query in addition to the relevancy scores. The scores are based on multiple signals like clickstream data, catalog data, user demographic data, seasonality, etc.
(Read this blog Learn to rank with eCommerce data to understand how Unbxd is deploying Learning to Rank (LeToR) AI-model to rank the relevant products higher for optimal results.)
Unbxd combines these scores based on weightage to calculate an overall ranking score for each product in the retrieved set of products. The products are then sorted in descending order of that score to show the results for a query. Merchandising functions in the Unbxd Dashboard such as boost and bury are available for you to overwrite the computed scores and change the sort order of products.
With the new release of ranking controls, users would be able to create different models of the various ranking scores and manage how the products are ranked in the search results. The users can assign weights to different signals and control their relative importance in the scoring.
User behavior score tries to rank products for a particular query based on clickstream data and catalog-based signals. The score looks at how shoppers have interacted with products for the different queries. These signals affect the ranking of the query based on how significant the data is for that query.
Queries are broadly divided into 3 buckets based on their frequency:
- Head Query – These are high-frequency queries that have significant clickstream data.
- Torso Query – These are mid-frequency queries and have considerable clickstream data. Optimizations for these queries involve partially mapping these to their closest head queries and leveraging the head queries clickstream data for optimization.
- Tail Query – These are low-frequency queries and have very low or insignificant clickstream data. These queries heavily rely on Head Queries for optimization in their rankings.
By default, UB scores are generated for queries with significant clickstream data based on the last 30 days of clickstream data. The score generated is a function of clicks, carts, and orders for different time periods.
As part of the User behavior score, users can create custom models where they can configure the following parameters:
- Number of queries – The number of head queries for which scoring gets applied
- The observation period for the clickstream data – The time period of clickstream data used to train the model
- Clickstream attributes – The user can add various clickstream attributes like click, carts, order, revenue, etc and assign relative weights to them
- Catalog attributes – The user can add catalog attributes like isSale, isClearance, isNew, AverageRating, Discount% and assign weightage to them.
Once the model configuration is set, the model training kicks in and the model would be ready within a span of a couple of minutes. Once the model is ready, you can also preview the results of the model within the console
The popularity score of the product is related to performance of every product across the site. Shoppers generally prefer to buy the best products in any store and the popularity of the product is the best metric to gauge it. Thus it makes most sense in including the popularity into the ranking score.
As part of the Popularity score, users can create custom models where they can configure the following parameters:
- Observation period for the clickstream data – The time period of clickstream data used to train the model
- Clickstream attributes – The user can add various clickstream attributes like click, carts, order, revenue, etc and assign relative weightages to them
- Catalog attributes – The user can add catalog attributes like isSale, isClearance, isNew, AverageRating, Discount% and assign weightages to them.
Personalization as a feature enables the shoppers to create their own shopping journey. It helps eCommerce businesses to present relevant products to shoppers in order to make the online shopping journey more efficient, enjoyable, and rewarding.
Unbxd uses a deep learning algorithm to provide personalized results to shoppers by using various signals like catalog data, user data, demographics, clickstream data, seasonality, attribute affinity, etc. The model learns from automated signal detection and it also facilitates automated user profiling, segmentation & user context understanding.
Users would have control of both the model training and model application. The levers that would be exposed during the model training are:
- Product attribute selection – the list of catalog attributes based on which the products would be personalized. This list of attributes would be used to measure the affinities of the user
- Clickstream attribute selection – The events based on which product would be personalized. The events can be clicks, carts, orders, or revenue
- User attributes – Signals like location, device type, etc can be used to train the model.
In addition to the above, users would be able to decide on how to apply the model and apply different strategies. They can use either of the following strategies to rank the products:
- Boosting strategy – An additive boosting score would be assigned to each product
- Re-rank strategy – the user would be able to re-rank the top ‘n’ products just based on the personalization score
Any new product added to the catalog faces the “Cold Start” problem. The problem arises because any new product never gains popularity as it would also be buried in the bottom and never have any clickstream data. The freshness feature helps new products to gain a default boost for a particular period of time and hence gain an initial popularity and clickstream data.
Users would be able to configure the following to control the score:
- Freshness period – This is the period of time for which the extra additive boost would be applied and the period of time for which the product would be considered fresh
- Decay Function – the degrading function based on which the score would reduce as time increases
How can eCommerce teams leverage the in-console controls?
While most of the search solutions in the market promise to fix the relevance issue for search queries, no one talks about the how part of it. Merchandisers and eCommerce teams have no clue what happens at the backend to fix search relevance and ranking of the relevant products. And generally, the results are not impressive. Plus, they have no control over these relevance and ranking techniques and algorithms. Everything is like a black box to them.
We have had many instances where eCommerce stores came to us with similar challenges in the hope of finding an out-of-the-box solution. Unbxd ranking algorithms diligently score the relevant products for a search query so that the right products can be shown in descending order of the popularity or likelihood of purchase. I believe the most powerful thing about our technology and the product is its capability to leave enough control in the hands of merchandisers and eCommerce teams so that they are not left in the dark and make business decisions suited to their business needs.
In case ranking and sorting of the products is one of the pressing challenges you are facing to convert your shoppers to your online store, you should try the Unbxd Site Search.