How Amazon A9 Algorithm Works?

The Amazon search engine typically operates in two stages: matching to identify the products that match the search query, and ranking to sort them.

Here is the workflow when a customer enters a search query: 

• A user enters a query in the search bar
• The query passes through a matching engine, which retrieves all products containing words in the query called a match set
• The match set passes through ranking stages, where the top results are sorted through a machine-learned rank function.
• The A9 search engine then displays the most relevant results for the user.

Matching

The matching engine creates a list of products in three ways.

Optimizing for A9’s top 3 factors — sales velocity, keywords, and price — moved listings an average of 5.7 positions higher within 60 days.

Direct Matching

Products that contain the exact query are the primary candidates. If a search query is ‘wireless gaming mouse’, the results will primarily show products that have the exact search query in the product listing. However, pure lexical matching has problems due to a lack of semantics. 

A lexical search looks for exact matches in a database, while a semantic search examines the meaning of words and phrases.

Semantics Matching

While many products have different keywords in their product listings, semantics resolves many search experience problems. The following are some semantic understandings of the Amazon A9 algorithm:

• Hypernyms: Umbrella words such as footwear including shoes, sandals, boots, etc.
• Synonyms: Different words with the same meaning
• Antonymous: Words with opposite meanings
• Stemming: Roots words such as joggers stem from jog.
• Misspellings: 10-15% of queries are misspelled on Amazon

Semantics significantly improves customers’ search experience. However, there is another subset of the matching engine which takes customer feedback to obtain a list of qualified products.

Behavior Data

How customers react to a product listing is a huge indicator of matching the right products with the right search query. If the algorithm sees a pattern in which customers search for ‘type C adapter’, or similar terms, and finds this one product getting the most conversions, that is the ultimate confirmation of matching. 

Generally, there are four signals of purchase behavior:

• Clicks on a product based on a search query are a signal
• Added a product to their shopping cart is a stronger signal
• Purchasing a product is the strongest signal
• Viewing and not engaging with the product is a negative signal.

Even with behavioral data, there are issues that the algorithm is trying to solve. For example, a product on page one has a higher chance of being clicked than a product on page 3. If a product is not visible to a customer, it won’t get clicks. Similarly, if it isn’t getting clicks, it won’t be visible.

Therefore, to counter the chicken-egg issue and sort a huge match set, Amazon is working on several models to improve its ranking algorithm.

Ranking

After getting a match for a keyword, algorithms resolve a much more complex problem i.e. sorting the products in order. The algorithm re-runs the behavioral data at the ranking stage through multiple models to give a better customer experience.

User intent signals

The Amazon A9 algorithm improves search results by understanding latent user intent. For instance, a customer won’t necessarily tell the intent of the product they are searching for, therefore, the algorithm trains to interpret the intent of the customer through activity and audience. 

Activity defines how the product will be used, while the audience defines the people who will use the product. By appealing to these two elements, you ensure that your products are more likely to rank for relevant searches. If a customer searches for ‘men’s running shoes’, it will find products that have keywords related to the “activity of running, CrossFit, or outdoor running as well as the audience like “men,” and “athletes” for example. Therefore, implementing activity and user-related keywords in your product listing increases your likelihood of higher visibility.  

Moreover, Amazon learns about these attributes by looking through Amazon reviews that are not indexed. For example, if the product listing doesn’t have keywords such as CrossFit, and several product reviews mention CrossFit, there is a higher chance of getting visibility on CrossFit related searches. Therefore, using keywords from your customer reviews can give you a stronger ranking.

Products visibility problem

Most new products struggle with minimal visibility in searches and recommendations. To deal with such cold launches, the Amazon A9 algorithm runs Learning-to-Rank (LTR) models. These models take signals from two sets of data, prior values and posterior values. 

A prior value represents initial predictions about the product before it has any data such as search history. Posterior values are ones with some search data. The Amazon A9 algorithm improves product search by making predictions about user preferences based on their search history. 

The algorithm uses prior values to represent assumptions about user preferences before any search data is observed. It is then updated based on the search data to obtain posterior values, which are used to make predictions about future search queries. The document shows that Amazon experimented with millions of queries and found that LTR models do bring in impressions and positive engagement to new products.  

While LTR models may take time to get visibility to your products, you can expedite the process by automating advertising and promotions your newly launched products.

Read more: Top 10 Amazon PPC Campaign Tips

Context-aware search

To improve searches, the Amazon A9 algorithm leverages feedback through personalized history such as user behavior and brand preferences. While not all customers will have an extensive history, it uses a combination of long-term and short-term interests to create a context-aware model.

Data sets of long-term interests are created through the efficacy of purchase history in personalized product searches. For short-term interests, the algorithm leverages click within a query session throughout the results pages leading to finding its ideal product.

Imagine a high-end luxury bag that gets more clicks because customers want to see more images. It would have a low conversion rate against other lower-priced bags because it is relatively expensive and customers are only window shopping. If the algorithm understands the context that a high-end luxury bag will naturally have a low conversion rate against other bags, it would take the user behavior on the product into consideration and may not take a low conversion rate as a strong negative signal. 

Another use case is if a customer has brand preferences through search or purchase history, it would be more likely to show more results of that particular brand. The paper shows that using such context models gives better search results to users, especially for short-term datasets. 

As a seller, make sure you identify the context of your product and use keywords that resonate with your target audience. If you sell high-end luxury bags, understand the context of your product against other products on the shelf. Using keywords such as elegant, modern, quality, and exclusive will get you the right audience.