Query Understanding: When Search Has to Guess What the Shopper Means

A shopper who types running shoes might want shoes for running, shoes for walking, or shoes that look like running shoes. The search box has to guess, and the guess is the whole job.

Query understanding is the work search does before ranking: deciding what the shopper means by the words they typed. It covers synonyms, intent, and context, and it decides which products are even eligible to rank. Ranking can only order what understanding lets through.

Matching is the job, ranking is the sorting

Most of the attention in search goes to ranking: which product should be first, which second, which buried on page three. Ranking matters, but it is not the job. The job is matching: deciding which products are even eligible to appear for this query, before any ranking happens. A search engine that ranks perfectly and matches badly will show the shopper a beautifully ordered list of the wrong products.

The shopper and the catalog speak different languages, and that is the gap understanding has to close. A shopper types running shoes. The catalog says performance footwear. Without a synonym map, the two never meet, and the shopper sees nothing. Amazon's A9 system ranks on two axes, relevance and performance, and relevance is the first gate: a product that does not match the query cannot rank no matter how well it sells.

Related reading on this site: RelevanceRail and Zero-Result Searches: What Your Search Log Is Telling You cover the neighbouring problems in more depth.

Where the guess goes wrong

The guess goes wrong in three places. First, synonyms: the shopper uses one word, the catalog uses another, and the mapping does not exist. Second, intent: a shopper who types gift for dad is not looking for a product called gift, they are looking for products a dad would like, and the search has to interpret the intent instead of matching the word. Third, context: the same query means something different from a first-time visitor than from a returning buyer, and the search has to use the session to narrow it down.

The symptom of a bad guess is easy to spot in the search log. Shoppers retype the query, which means the first guess was wrong. Shoppers get zero results, which means the guess was so wrong nothing matched. Shoppers click a result and leave immediately, which means the guess was close enough to show something and wrong enough that the something was not what they wanted. Each symptom is a guess that failed, and the log is the list of failures.

The signals that fix the guess

The fix starts with the search log, not with a new algorithm. Read the reformulated queries and the zero-result queries, and the list of guesses that failed is right there. The most common misspellings and the most common synonyms of the top products are usually a handful of words, and fixing those covers most of the loss. A judged set of queries with expected results is the honest test: if the search cannot return the expected products for the queries that matter, the understanding is broken.

The second fix is to use the session. A shopper who has been browsing running shoes and then searches for shoes is almost certainly looking for running shoes, and the search should use that context. A shopper who has been browsing gifts and searches for dad is looking for a gift for a dad. The session is the cheapest context available, and most search systems ignore it. Understanding is not a model. It is a habit of reading the log and using the context.

Frequently asked questions

What is query understanding in site search?

The work search does before ranking: deciding what the shopper means by the words they typed. It covers synonyms, intent, and session context, and it decides which products are eligible to appear at all.

Why do shoppers and search engines disagree on words?

Because shoppers use the words they know, not the words in your catalog. A shopper types running shoes, your catalog says performance footwear, and without synonym mapping the two never meet.

How do you measure whether search understands queries?

Watch the reformulation rate, how often shoppers retype a query, and the zero-result rate. A high reformulation rate means the first guess was wrong. A judged set of queries with expected results is the honest test.

Is query understanding the same as personalization?

No, but they stack. Understanding decides what the query means for anyone. Personalization re-ranks the results for this shopper. Get understanding right first, because personalizing a misunderstood query just shows the wrong products faster.