Introduction

A search experience lives and dies by the quality of its ranking, and this is a key differentiator for Constructor: we provide advanced capabilities to help optimize your result rankings.

The goal of this article is to help you understand the components that affect result ranking.

Machine learning & behavioral data

The most critical input to result ranking and Constructor's key differentiator is the integration of behavioral signals. This includes obvious signals like searches, clicks, and conversions, but also less obvious signals, such as refinements, imputed non-clicks, excessive pagination and others.

These behavioral signals inform an ensemble of ranking factors that together constitute the Constructor Ranking Engine.

Relevance vs. attractiveness

The other broad theme of relevance is that our algorithms are designed to positively impact key business metrics, not just perceived relevance. You can read more here, but broadly this is the difference between returning laptops for the laptops query (even if they're out of date and only include one brand on the entire first page) and returning attractive results including up-to-date high-converting laptops from a variety of manufacturers to increase diversity.

Ranking factors

Base score

The first component is the base score, which is calculated based on an item's overall popularity. The goal of this number is to tell Constructor how the item ranks in terms of sales to seed the algorithm with knowledge of the most popular items.

For example, McCormick's Organic Farm Fresh Baby Spinach sold 900 units in the last 30 days, and that is used to calculate its base score.

Calculated score

Constructor takes the suggested score as input and then applies additional ranking intelligence based on observed clicks, add to carts, purchases, and other signals over time. This composite or calculated score attempts to estimate a broad understanding of how effectively this item performs across search and browse results.

For example, McCormick's Farm Fresh Baby Spinach has seen an uptick in add-to-carts and purchases since going live on Constructor's platform, and is also frequently clicked, so its calculated score has been increased by the algorithm.

Base item-query score

When retrieving matches, Constructor applies boosts or penalties to the calculated score based on how close the item matches the query, what type of spelling differences were found in the query as compared to all products, and whether any synonyms or stemming rules were applied.

For example: our user has searched for baby spinahc. We spell correct this to the most common terms in the catalog and the most typographically likely baby spinach. At this point all keyword matches for baby spinach are returned in order of popularity.

Since baby food containing spinach is more popular than baby spinach, McCormick's Delicious Baby Food with Spinach and Bananas far outranks our McCormick's Farm Fresh Baby Spinach.

This is where most other search engines leave your users - misunderstanding intent and applying one ranking function (mainly overall popularity) to all queries.

Affinity engine

Constructor continues the ranking process: the result set is filtered through product taxonomy and user behavior understanding to:

• Increase ranking of products from categories and facets historically associated with conversion for the particular query.
• Decrease ranking of products historically less associated with conversion for the particular query.

Since Vegetables is more associated with conversion for baby spinach, all products from this category and facet are boosted. Meanwhile, since Baby Care is less associated with conversion for this query, the baby food items are penalized. This means baby spinach results are at the top and baby food towards the bottom.

Learning to rank

While the affinity engine boosts and buries products based on their broad categories and facets, Learning to Rank changes the score for specific items based on what is most and least likely to convert for the particular search term as well as hundreds of machine-learned rules capturing item relevance and result attractiveness, as well as optimization parameters for the highest-value business metrics defined in collaboration with the Constructor data science team, such as margins or gross merchandise value.

In this case, there is another item, Dole Baby Spinach and Kale Salad Mix. The Learning to Rank algorithm:

Has learned that Farm and Fresh are relevant terms for spinach
Observes that "and kale," while close in meaning to spinach, suggests that it isn't solely baby spinach.
Identifies that McCormick's Farm Fresh Baby Spinach has a higher gross margin
For all these reasons, Dole Baby Spinach and Kale Salad Mix outranks McCormick's Farm Fresh Baby Spinach.

Searchandising

Searchandising provides merchants the ability to manually slot items to particular positions as well as to promote, bury, or remove rules matching complex sets of filters and categories. This provides merchants the capability to increase the rank of own-brand products, or decrease the rank of products in specific price ranges or from particular facet values.

In this case, the merchant has added a rule to boost McCormick's products, but the two McCormick's Spinach products are already ranked highest, leaving the ranking unchanged.

Personalization

Finally, Constructor applies the last phase of ranking, personalization. This stage evaluates a user's history of behavior to gain an understanding of their preferences and increases the rank of items that most closely relate to their purchase history.

Returning one last time to McCormick's Groceries, the user hasn't purchased spinach previously, but they have purchased several bags of organic frozen vegetables, so McCormick's Farm Fresh Organic Spinach is presented first.