What happens when a customer arrives at your site, types in the household brand name of an item, and nothing comes up? Off he goes to another website to fulfill his goal. A product search with no results is a deathtrap for retailers. Here is how semantic vector search solves the “no results†problem, writes Justin Sears, VP of marketing, Lucidworks.
 There is an old folk song called “Big Rock Candy Mountain†that has surprisingly great analogies for modern-day ecommerce. It is the first-person ballad of a hobo, wandering Western railroad tracks, on a journey of product discovery, looking for items that would fulfill his wildest dreams.
Today, that mythical hobo would probably go online and search for the “lemonade springs†or “lake of stew†before traipsing halfway across the country to find it. But such a searcher would undoubtedly encounter a “product not found†message due to his colorful language and bizarre tastes.
Spoiler Alert: Nobody Has Ever Found That Lake of Stew.
 Ecommerce merchandisers should not only accommodate colorful search language, but they should also encourage it as a strong indicator of the shopper’s intent. Ecommerce teams that combine the power of AI-powered search solutions with a steady stream of user signals can bring consumers to their personal “candy mountains†more reliably.
We expect this predictive intelligence from search engines. A natural language search engine can understand shopper intent and then offer the best product to match it.
Learn More: Search Advertising Monopolies Are Putting Brands at Risk
Suppose I search for the shoes I am wearing right now: “Cariuma mineral blue knit IBI sneakers size 9â€. If the site carries the product, those kicks go in the cart, I check out, pay, and everyone is happy. But what if that product is not in inventory? What if I spell the product name wrong? What if I use a synonym? What if I cannot remember what the shoes are called? Then I get the dreaded “product not found†null results message, and off I go to another website to find what I want.
Warning: A Product Search With Null Results Feels Like a Straight-Arm to the Face
The real relationship killer is when a retailer has what we want but makes us guess how they prefer to describe it. If I search for “Band-aids†and you do not carry the Band-Aid brand, just show me adhesive bandages. I am bleeding here.
So how can you fix those dead-end queries? Here are three suggestions.
1:Track Trending Language With Head/Tail Analysis
The next two suggestions will cover how to prevent poor null-results customer experiences. This first suggestion makes the case that, before you eliminate them, null results queries can be beautiful diamonds in the rough.
Retailers must keep up with emerging trends, and that means understanding emerging language. Of course, it is valuable to know the most common search terms associated with the products one sells. What is more valuable than that? Knowing the uncommon search terms.
In the marketplaces of SEO and SEM keywords, this is referred to as head/tail analysis, and it is a great way for merchandisers to “buy low and sell highâ€. Let other retailers pay top dollar for the common keywords while you scoop up the uncommon (and often more descriptive) keywords for pennies on the dollar. After all, wouldn’t you rather someone search your site for a “lake of stew†than just “soupâ€?
Here is why null-result queries contain valuable signals about three potential scenarios:
- Tastes are changing, and you should prepare for that trend,
- Your language is out of tune to describe what you already offer, or
- You need to expand your catalog.
Ecommerce teams need an easy interface to watch for the hundreds or thousands of near misses that happen on sites every day. With that information, merchandisers can continuously take discrete data-driven actions that improve conversion rates and average order values.
Let me repeat that last part because it is vital: data-driven actions that improve conversion rates and average order values.
2: Understand Shopper Context With Signal Capture
Good in-store clerks can understand a shopper’s context and infer their intent. When a shopper walks into the store, she is a complete mystery, but she does give off signals that can drive insight.
How old is she? What is she wearing? Does she seem to be in a hurry? Does she have bags from other retailers slung over her arm? These are rich contextual signals that can be used to shape great commerce experiences, but that single salesperson cannot size up hundreds of people walking through the door at the same time. Also, when employees find other jobs, their learned intelligence walks out the door.
Even though the contextual signals might not be as vivid online, retailers can more than make it up on volume. The sites for some of Lucidworks’ biggest customers capture thousands of user behavior signals (e.g., search queries, page views, add-to-carts) per second. This paints an increasingly vivid picture of each customer’s journey as it is happening, which can be used to automatically personalize the experience. Machine learning (ML) watches all the signals all the time and can do a lot more learning on your site than your best sales associate can do in the store. And ML never forgets the patterns it learns.
3: Predict Shopper Intent With Semantic Vector Search
Now that you are tracking trending language and understanding behavioral context, how can you predict shopper intent and match that intent to search results? The answer is semantic vector search.
Semantic what? Let me explain. No, there is too much. Let me sum up.
First, some definitions:
- Lexical search: matching search results to the exact words in the query
- Semantic search: deriving meaning from the query string, then matching results to that derived meaning
- Vector search: encoding language into vector representations, then calculating the distance between the meaning of two concepts as a measure of similarity
- Semantic vector search: representing concepts (e.g., ideas within documents, pictures, and videos) as semantic vectors within a mathematical space, deriving intent from a search query and then minimizing the distance between query and concept to return the most relevant results
Because lexical search only looks at the words without guessing at meaning, it has a hard time with two sources of search ambiguity: synonyms and polynyms.
Synonyms are multiple words for the same thing. McClintock seeks the Big Rock Candy Mountains. Would he settle for the Big Rock Candy Bluffs? Or the Big Rock Candy Peaks? Buttes? Cliffs? Precipices? Probably all yes, as long as they were sweet.
Polynyms are multiple things for the same word. For example, consider McClintock’s dream of a place where all the “railway bulls are blindâ€. A natural language search engine without context might understand the word “bull†as the angry animal with sharp horns, slang for what that animal periodically drops on the ground, or how a politician might talk. In the context of the song, a railway bull is a railroad police officer trying to throw old Harry out of an empty boxcar.
Lexical search has trouble with synonyms and polynyms because it requires an infinite list of rules: if the shopper writes “butteâ€, he means “mountainâ€, and if he writes “bullâ€, he means “police officerâ€. But language changes, as do tastes and preferences.
Many merchandisers spend all day rewriting thesaurus entries just so that shoppers can find the products they offer. It is necessary work, but it wastes their brainpower and is boring.
Learn More: Better Together: How To Pair Programmatic and Paid Search
Semantic Search Is Better Than Lexical Search When Searches Are Ambiguous
Semantic vector search uses deep learning to predict the meaning of the query. Machine learning models learn from user signals to encode products and queries as vectors (which quantify the distance between words and concepts). Then those models build knowledge graphs that understand, in a fairly human way, that “lemonade springs†refers to an unlimited amount of cool, sweet, citrusy beverage for anyone with his or her own cup. Time to offer that shopper a soda fountainOpens a new window or a refreshment stationOpens a new window .
To identify whether you have a null results problem that could be solved with semantic vector search, ask yourself these questions:
| 3 QUESTIONS TO ASK YOURSELF ABOUT NULL RESULTS | |
| Do more than 10% of my site search queries fail to match directly to anything in the catalog? | Yes or No |
| Do I have a high bounce rate from pages where I show search results? | Yes or No |
| Do shoppers tend to submit multiple queries before they click on a product detail page? | Yes or No |
If you answered yes to any or all of the above, you could probably improve conversion quickly by using vector search to surface a related item rather than just saying, “Nope.â€
 Say Goodbye To Null Results and Missed Opportunities
“No results†searches, or “null results†searches, or “zero results†searches can crush your shoppers’ hopes and send them into the arms of your competitors. Be sure to implement these three things to save shoppers from those dreaded dead-ends:
- head/tail analysis,
- signal capture, and
- semantic vector search.
That way, we can all build a commerce experience that feels like Harry McClintock’s nirvana, crooned about almost one hundred years ago, where the bluebird sings to the lemonade springs in the Big Rock Candy Mountains.