Product search engines allow searchers to find products offered by online merchants. And while product search engines have a specific mission, the challenges they face are similar to those faced by web search engines.
For instance, those who design product engines must return the best set of results they possibly can for a search query, ordered by relevance. The quality of a search engine is evaluated by using HRS (Human Relevance System) and user-engagement metrics like click-through rate, last-click rate, first-click rate, abandonment rate etc.
The same metrics are used by other vertical engines, like job search engines, image search engines, local search engines etc. In this write up, I want to share how building a product search engine is unique.
From a consumer standpoint, the important metrics around product search engines are similar to the important metrics for most vertical search engines and web search engines:
- HRS based metrics like NDCG, DCG, Win-Loss Ratio.
- Click-through rate (clicks/visit), conversion rate (views/visit).
From a merchant perspective, we consider different metrics — like revenue per visit (RPV), revenue per search session (RPS), margin per visit (MPV) and also search participation rate (searches/visit). The more revenue the user brings in and the more engaged the user is, the better for the merchant.
As we know most search engines use multiple signals, which they eventually combine with a machine-learning/hand-tuned model. While the models are important, the set of features the models use to rank results is equally as important. But what are some of the features that are unique to an e-commerce search engine?
Product and Attribute Understanding
Most web search engines use syntactical features/matching algorithms like BM25F /TF-IDF and other N-gram matching features. Product search is an area where we need to understand products and attributes in depth.
Wilson 4-Drawer Filing Cabinet Black
Product Type: File Cabinet
The understanding of these features helps us do matching better. This understanding also helps us to get into semantic understanding of queries and matching. Having a strong attribute-extraction algorithm helps here.
This is another area, which goes hand-in-hand with the product understanding. Annotating queries into understandable attributes helps us substantially in semantic matching of products.
“wilson 4-drawer filing cabinet”
Product Type: File Cabinet
Query understanding also involves understanding synonyms — in this case understanding that a “file cabinet” is the same as a “filing cabinet.” Similar to web search engines, stemming, autocorrect, related searches are areas to understand as well.
Autosuggest or Guided Search
Users often need help on what they need to find — so guiding users with autosuggest can help with search participation rate. From a product search perspective, we can funnel the users to the queries which bring in the most revenue or are most likely to increase the merchant’s revenue.
Inventory and availability are very important signals we need to understand in a product search engine. A product which has sold out or is not available in most popular sizes is not something which want to show to a customer.
Product Performance Data
Unlike web search engines — which use the number of clicks, last clicks, first clicks, click-through rate etc., — we can use the revenue contributed by the product and the propensity of a user to interact with the product, including cart additions, to assess product performance.
As users, we often want to buy the products that have the fastest ship times/best supplier. (For example, on Amazon Prime users often go for Prime products because they have faster and more reliable shipping.)
Facets are a very important part of product search engines. They help users narrow down the set of choices they can make. When we have thousands of choices, the right set of facets and the right ordering of facets makes a big difference.
Shopping search is definitely one of the verticals where personalization has a huge impact. For example, deep personalization based on gender preferences is something that can dramatically influence results. And when it comes to B-to-B merchants, personalization is key, given how widely the buying patterns of different accounts vary.
This is by no means an exhaustive list of features that we can optimize to build a better product search engine. But taken together they provide a good flavor of some of the unique considerations for building a high-quality product search engine.
Ramkumar Rajendran is a Director of Engineering at Bloomreach