Search engines have used natural language processing and semantics for a while now, but marketers have yet to harness the potential of using natural language search on their sites to boost conversions.
The incongruent experience of e-commerce search
Let’s take http://www.sears.com as an e-commerce search example. Sears is an American, multinational mid-range department store. Its website is also its e-commerce and online business branch.
We tried two search requests on its website search engine:
- tv 22inches
- television 22”
These are the results:
Looking at the screenshots you can easily see that they are different.
Another example: PC Mall, Inc. is a value-added direct marketing company that offers computer and electronic products, services, and solutions. Its e-commerce website is http://www.pcm.com.
These are the two search queries we tried:
- laptop 4gb ram 1tb disk
- laptop 4 gigabytes of RAM and 1 terabyte of disk
Below are the results at the time of our search:
Again, you can see that results are different.
In both cases with Sears and PC Mall, the searches were intended for the same product – but why are the results so different?
Although users may find the products they want, is this the search experience that companies want to offer its customers? It is concerning that relatively similar search queries, would have results that different.
Yet another example, the largest online retailer: Amazon (http://www.amazon.com) started as an online bookstore, but soon diversified, selling DVDs, MP3, software, video games, electronics, apparel, furniture, food, toys, jewelry and even musical instruments!
If you an aficionado to rock & roll or blues music, you will know what an electric bass guitar is, although the bass guitar is used in many styles of music including pop, country, reggae, gospel and jazz.
The standard design for the electric bass guitar has four strings, although five strings is also a very common option.
We tried these 2 search queries at Amazon:
- 5 string electric bass guitar above $1000
- bass guitar with five strings for more than 1,000 dollars
And these are the results we got:
￼Again, just looking at these 2 pictures, you will realize that the search results are very different, although the search query was basically the same.
It is worth noting that Amazon recognizes structures like “above $1000” or “below $1000”. For example, if you search “guitar below $1500” you will get exactly what you were looking for: guitars that cost less than 1,500 dollars. Similarly, if you search “guitar above $1500”, you will get a list of guitars that cost more than 1,500 dollars.
Interestingly enough, if you search “guitar above $1000” you will get as the first result an Apple computer! Along 17,866 more results.
Anyway, e-commerce search experience is far from being a coherent, straightforward experience.
The problem of e-commerce search
When it comes to e-commerce search, most websites focus on SKU (Stock Keeping Units) to manage their “catalog”, as this is the logical component that can be bought by the user.
However, users don’t look for SKU’s: users search for products and services that they need, describing the attributes, features and characteristics that they need like:
- screwdrivers 6 inches
- car radio with bluetooth compatible with iphone
- guns and roses concert in san francisco bay area
Search queries with 3 or more words are considered “Long Tail”. And Online Marketers invest a respectable amount of money trying to find the keywords that would deliver ROI.
However, this apparent “Long Tail” is really the external form of a much more complex structure, that we could modeled like:
“popular product name” + “features”
Where “popular product name” would be any combination of words that described the product we need, and features the set of attributes we want the product to comply (price, size, color, voltage, dates, city, ….and an endless list of attributes, all of them ambiguous, expressed in many different units, and unstructured in nature).
Needless to say, “features” and “product name” can appear in any order, making ecommerce search particularly challenging.
A model for e-commerce search engine
Instead of considering products in our e-commerce catalog as a sequence of SKU, and let a keyword-based search engine try to solve the problem of search, we should consider the catalog a bi-dimensional matrix, where different products would be rows, and features would be columns.
|Name||Category||Price||Color||Size||Thousands of features……|
|…Hundreds of thousands of products…||…|
Features, attributes or characteristics describe how every product is and behaves, and might be of very different nature:
- Can be numeric values in a particular unit (price in dollars, length in inches, voltage in volts, resolution in pixels…)
- Can be a list of possible values (colors, days of the week, cities in the country….)
- Can be a multidimensional value (resolution would be width times height, dates would consist of day, month and year, …)
- …and they could be many other things not listed above!
With this approach, searching in an e-commerce site will mean finding the right product (row) that satisfies the right attributes (columns).
The role of natural language processing
Ticketmaster Entertainment, Inc. is an American ticket sales and Distribution Company based in West Hollywood, California, USA, with operations in many countries around the world.
The following example is extracted from http://www.ticketmaster.es, the website of the company that Ticketmaster has in Barcelona, and that is powered by Inbenta Natural Language Search Engine.
In this case, the search query is “theater in barcelona next weekend” and this the search results, at the time of this writing:
Observe how “barcelona” has been interpreted by the natural language processor as a city, so in the “faceted search” shown on the left side, Barcelona is clicked.
In the other hand “theater” has been interpreted as a “category”, selecting all shows that belong to this particular kind of show.
Also “next weekend” is interpreted as a range of 2 dates (Friday 4/12/2013 and Sunday 4/14/2013), which is the weekend after the weekend after today (4/4/2013 as I am writing this).
Also, 2 FAQ are shown among the search results, “How to buy tickets” and “I can’t find an event on your Website”, that would solve the most common questions users have when searching Ticketmaster’s catalog.
That’s an example on how natural language processing is used in conjunction of e-commerce site search to refine search queries and produce relevant results.
The “search to cart” rate
“Search to cart” rate or S2C rate is computed based on how many search results end up in the shopping cart.
It is computed as: s2c = cart / search where “cart” is the total number of products shown on the search results page that where put in the shopping cart divided by the total number of search requests in the product search engine.
Using natural language processing we have measured a great increase in S2C rates, which bring the whole ROI.
Return on investment
At Inbenta we are devoted to producing a best-class NLP technology that allows our customers to:
- Increase revenue online by increasing conversion ratio and “search-to-cart” efficiency
- Decrease customer support costs by providing the best self-service application and knowledge base management
By applying these techniques on e-commerce website search, we have accomplished the following results in the first few weeks.
- Increase in “search-to-cart” conversion ratio: +2.42%
- Increase average purchase value: +11%
As we talk, new features are being implemented in our e-commerce search engine, and new features and attributes are being deployed, so we expect these ratios to be growing in the future.
natural language search is a concept that everyone will need to become familiar with in the next few months. By integrating it properly, e-commerce website will be able to provide shoppers the right answers at the right time to improve conversion rates and drive additional revenue.
Founder and CEO
Inbenta Technologies Inc.