A search experience lives and dies by the quality of its ranking, and this is a key differentiator for Constructor: we provide advanced capabilities that help businesses deliver personalized results for users while optimizing for key business metrics.

In this article, we walk you through how Constructor does ranking differently and what factors affect item rankings in Search and Browse.

Why ranking matters

• Improve user experience.
  • Showing the most attractive items at the top of search or browse results allows users to find what they are looking for more easily and quickly, leading to a better overall experience.
• Optimize for key business metrics.
  • Constructor’s algorithm can optimize for key business metrics by strategically ranking items that are more likely to increase revenue per visitor, average order value, profits, conversions, or an ecommerce KPI of a business’s choosing.
• Deliver personalized results.
  • Considering user history, preferences, and context (such as device type, location, time of day, etc.) allows Constructor’s ranking algorithm to offer personalized results that are more likely to meet an individual user's needs and preferences.
• Balance business goals.
  • Constructor’s ranking algorithm can be tuned to find the optimal balance between different goals, such as maximizing conversions, while also considering inventory optimization, such as ensuring a variety of items are seen or promoting new items.

How Constructor does ranking differently

Ranking is at the heart of Constructor’s discovery experience. Our ML-powered item ranking system helps businesses present users with the most relevant and attractive items for their queries, achieving key business goals.

To deliver on this promise, we approach ranking differently.

1. We incorporate behavioral signals

Behavioral signals, or clickstream data, play a critical role in results ranking. Each interaction a user has with your website—from the time they arrive to the time they leave—is recorded and stored anonymously in the clickstream.

Constructor uses this data to learn essential information about a business’s users, their preferences, and how those change over time. Every user interaction—which pages they visit, what items they engage with or scroll past, what items they add to or remove from their cart, which items they purchase (and which items they ultimately return)—tells Constructor something important about a business’s user base and the individuals within it.

These learnings are then applied to retrieved items to ensure the most attractive items are ranked towards the top and personalized for each user and each context. For example, say a user near the coast in California searches “summer fun.” In this context, for this user, a surfboard might be an appealing item and therefore rank higher than an item with fewer purchases in the same context.

2. We deliver attractive results, not just relevant ones

Constructor’s algorithm is designed to return results that a user will find attractive—not just items that match the search query.

For example, a general search engine might take a search query like "jeans" and return pairs that are no longer in style, or are only available in a single size. These pairs may look relevant to the naked eye, but are not likely to lead to many purchases. Constructor goes beyond passing the naked eye relevancy test to return results that are actually attractive to each user and therefore likely to lead to a conversion. By delivering attractive results, businesses can increase the likelihood of things they actually care about like conversions and revenue. This allows them to achieve their KPIs while simultaneously providing a positive customer experience.

To determine attractiveness, Constructor’s ranking algorithm runs several calculations—often simultaneously—to determine an item’s attractiveness score. This score will always be tailored to the user performing the search. The higher the score, the higher the item appears in ranked results.

When determining an item score, the system uses clickstream data for both the user performing the search and users who have performed the same search before. It looks at the historical performance of particular items or groups of items for the query and also folds in a user’s personal preference.

For example, say a user searches "jeans." The algorithm sees that in the jeans category, the group "high-waisted jeans" is particularly popular at the moment. Looking at the item level, "dark wash" high-waisted jeans are what other users are looking at, carting, and purchasing. Also noted is that the specific user performing the search often views "tall" bottoms. Therefore, the algorithm will rank tall, dark wash, and high-waisted jeans as the highest for the user.

Whether users are searching for something specific or just browsing around, Constructor’s ranking service works tirelessly behind the scenes—collecting data, learning from it, and applying that knowledge instantly—to connect users with items and content they'll love.

We walk through that process more in-depth next.

ℹ️

What about keyword matching?

While Constructor can use traditional keyword matching techniques to recall items that send a strong intent signal, or if there is not enough clickstream data available, it is not the primary matching mechanism used. Rather, Constructor’s ranking algorithm prioritizes behavioral data—such as user clicks, add-to-carts, and conversions—to determine item relevance and ranking.

How ranking works at Constructor

Constructor’s ranking system is a real-time ranking service powered by machine learning models. It has two main phases: offline and online.

Phase 1: Offline

The offline phase focuses on learning connections between items, affinities, and users, and making a vast web of these connections so that the system can infer patterns later. Its goal is to determine which types of items and content will appeal to which kinds of users and under what circumstances so that this information can be used to provide personalized discovery experiences. It happens behind the scenes before a query is even made.

During this phase, the ranking system gathers data from past user interactions (e.g., search queries, items returned, items clicked, items purchased, etc.) and item feature information, and stores all this data in something like a database.

The system then uses this data to train models on what makes an item attractive or relevant to specific users. It allows the models to identify patterns, such as which items are popular during certain times of the year or with a particular user group, and make predictions based on those patterns. The system then saves and uses the trained models when a user initiates a search, browsing, recommendations, or other discovery experience.

Constructor runs this process several times a day, ensuring the models are trained using the most recent historical data.

Phase 2: Online

The online phase focuses on retrieving and ranking items related to a user’s search query or browsing experience. It uses collected data and the trained machine learning models to return instant, personalized results for the user.

The online phase includes three ranking stages: initial item retrieval and foundation ranking, second-stage advanced ranking, and searchandising. We discuss each of these stages more in depth below.

Initial item retrieval and foundational ranking

As items are retrieved for a query or browsing experience, they are ranked using four separate factors: base score, group attractiveness, item attractiveness, and basic personalization. These factors are calculated simultaneously.

Base score

The algorithm calculates an item’s base score based on its overall popularity (i.e., how users interact with it) and its relevance to the search query.

Group attractiveness

With group attractiveness, the algorithm uses historical and clickstream data to surface items within a category, facet, or affinity group that will most appeal to the searcher.

For example, say a user searches "milk." The ranking algorithm knows milk in the "dairy" category is clicked on and purchased more often than milk items in the "cooking" category, like coconut milk, so it will rank that group of milk items higher. Ranking a group of items can help individual items within that category gain visibility even if the specific items lack clickstream data (e.g., a new organic milk brand).

Item attractiveness

With item attractiveness, the algorithm uses historical and clickstream data to rank items based on the likelihood that the item will be clicked on, added to cart, and/or purchased for the specific query or browse category. In short, it is how well an item has performed and is likely to perform for the specific query entered.

Say a user searches "tennis shoes." The algorithm might rank specific tennis shoes higher because they are often carted and/or purchased by other users who have searched “tennis shoes” recently.

Basic personalization

With basic personalization (not to be confused with advanced personalization that happens in the advanced ranking stage), the algorithm uses user-specific data, such as purchases and browsing history on a business’s website, to adjust the ranking of items. The goal is to prioritize items based on the searcher's understood interests. Basic personalization happens in real-time.

For example, say a user adds a pair of tennis shoes to their cart. They then navigate to the Socks browse page. Because personalization happens in real-time, the algorithm will ensure that active socks or tennis socks appear higher on the page.

Second-stage advanced ranking

Once the initial item retrieval and ranking process is done, the ranking algorithm reprioritizes or "re-ranks" items using additional advanced attractiveness signals. Reranking allows Constructor to focus on reprioritizing a smaller set of top, previously identified items, rather than running time-consuming advanced calculations on all items.

These signals, listed below, consist of several features that incorporate data from multiple sources, including user behavior, item attributes, query information, and past interactions. Constructor's advanced ranking model calculates a weight for each signal, which contributes to an item’s final score.

These signals will vary by company, meaning not every signal will be part of every company’s ranking algorithm. This is because signals are optimized to deliver on a specific business’s key outcomes. Signals can even vary within a company, with each unique brand or business using different signals. Signals can also include custom signals identified and provided by a business via their metadata. Constructor runs A/B tests on these factors, ensuring what performs best for a given company remains part of the algorithm.

The goal during this advanced ranking phase is to deliver results that achieve the best outcome for the KPIs and business outcomes set by the business.

Searchandising

Searchandising—or, more commonly, merchandising—is the last phase of the ranking algorithm. Here, the system applies merchant-defined searchandising rules to the ranking of items on search and browse pages.

Searchandising rules allow merchants to influence the final ranking of items by boosting or burying items based on criteria (e.g., brand, category, price, etc.), manually slotting specific items into certain rank positions, and similar actions.

Until now, the ranking algorithm has optimized results for overall or long-term business goals defined by the business. However, with searchandising rules, merchants can influence the ranking to help achieve short-term goals, such as promoting specific items or brands, by bumping or hiding specific items regardless of their original ranking.