“I intend to buy” or the easiest way to evaluate the quality of grocery search
“Why do you need a screwdriver, take this wonderful scooter better” - you probably know a similar situation. Modern online stores (and even more so marketplaces) perplex a potential buyer with such a mountain of irrelevant products that sometimes you may want to turn to the good old specialty store with two options and an expert seller.
Let's try to look at the problem from the side of specialists, who by all means try to improve sales in their online store. With the growth of the range, the search becomes a key component, the stability and quality of which affects the profit directly.
In other words, the “once, once and in production” strategy can have a tangible effect on the density of the hair covering of the responsible individuals. And constant testing and, excuse me, enough coverage is a very good way to the world of good unicorns and cloudless trouble-free. But how?
Consider possible alternatives (the author allowed himself to simplify some points in order to attract the sister of talent).
We sit down together and handle the main search queries. Top1000 will be enough? Or maybe better than Top10000? And what if a significant percentage of profit is buried in the long “tail” of rare requests? Oh, and for each new increment you need to repeat everything?
Actually, it is necessary to try to answer two simple (at first glance) questions:
How many products from the selected ones are relevant? (Precision)
How many relevant products are selected? (Recall)
Minute of Bore:
Difficulty, obviously, having to know which products are relevant for each specific request. In essence, this is a “golden set” of search results, with which it is necessary to compare current results. There are a lot of such sets (for every request we are interested in), moreover, the sets are not static: the emergence of new products or the end of sales of old ones is a clear reason to rebuild the “gold sets”. At least within the same category.
In a word - not the easiest way. From the word very.
CTR must always be monitored, but in the context of the task in hand, this is the case when, if something noticed, it is still too late. Helpful? Of course. Solves the problem? Definitely not, because it's better to know about the problem before the buyer finds out about it.
Very interesting, but how? And, most importantly, by whom? Crowdsourcing with random and cross validation? Girls from HR department? The case when you really want to call the good kind robots and load them with work on the most unattractive. Robot is not hard!
Wait, stay, it's all good, but can it be the same, but easier and cheaper?
Intent-based search has long been no news, but how can intentions help simplify and, most importantly, automate the search check? Consider a simple example, someone really wants to buy a Leatherman Skeletool. These keywords uniquely describe a specific product (omit the color).
So, having a completely unambiguous desire, a person goes to an online store in the hope of placing an order as quickly as possible (who is interested in wasting time on insignificant trifles). Note once again: a search query uniquely describes a person’s desire. Moreover, the query looks quite sufficient for validating the search.
Walmart , by the way, could have done better.
Are there any other similar requests that clearly describe the intentions of the buyer? In fact, there are a lot of such requests. Summarizing, we can distinguish at least two classes of search queries:
Thus, it is not so difficult to create a fully automated tool for basic validation of a search solution. The percentage of buyers with clear intentions is large enough (depends on the specific store), moreover, it is buyers with clear intentions who quickly lose their patience without getting what they want. Another example of unopposed results is Maxxis Bighorn:
By the way, "negative options" are very important important. For example, if there are not some products in the store assortment, you should not “find” them. And even more ridiculous are the results obtained by unjustifiably simplifying the initial search query (a very old approach, by the way):
Recognizing and working correctly with the intentions of the buyer is the key to success. The next article will be devoted to automated tools and significant budget savings for testing. High-quality search without a large team of testers is quite real.
Let's try to look at the problem from the side of specialists, who by all means try to improve sales in their online store. With the growth of the range, the search becomes a key component, the stability and quality of which affects the profit directly.
In other words, the “once, once and in production” strategy can have a tangible effect on the density of the hair covering of the responsible individuals. And constant testing and, excuse me, enough coverage is a very good way to the world of good unicorns and cloudless trouble-free. But how?
Consider possible alternatives (the author allowed himself to simplify some points in order to attract the sister of talent).
AB testing
We sit down together and handle the main search queries. Top1000 will be enough? Or maybe better than Top10000? And what if a significant percentage of profit is buried in the long “tail” of rare requests? Oh, and for each new increment you need to repeat everything?
F-measure (Precision / Recall)
Actually, it is necessary to try to answer two simple (at first glance) questions:
How many products from the selected ones are relevant? (Precision)
How many relevant products are selected? (Recall)
Minute of Bore:
Difficulty, obviously, having to know which products are relevant for each specific request. In essence, this is a “golden set” of search results, with which it is necessary to compare current results. There are a lot of such sets (for every request we are interested in), moreover, the sets are not static: the emergence of new products or the end of sales of old ones is a clear reason to rebuild the “gold sets”. At least within the same category.
In a word - not the easiest way. From the word very.
Clickthrough rate (CTR)
CTR must always be monitored, but in the context of the task in hand, this is the case when, if something noticed, it is still too late. Helpful? Of course. Solves the problem? Definitely not, because it's better to know about the problem before the buyer finds out about it.
Subjective assessment of relevance?
Very interesting, but how? And, most importantly, by whom? Crowdsourcing with random and cross validation? Girls from HR department? The case when you really want to call the good kind robots and load them with work on the most unattractive. Robot is not hard!
Wait, stay, it's all good, but can it be the same, but easier and cheaper?
Clear intentions greatly simplify the assessment of search quality.
Intent-based search has long been no news, but how can intentions help simplify and, most importantly, automate the search check? Consider a simple example, someone really wants to buy a Leatherman Skeletool. These keywords uniquely describe a specific product (omit the color).
So, having a completely unambiguous desire, a person goes to an online store in the hope of placing an order as quickly as possible (who is interested in wasting time on insignificant trifles). Note once again: a search query uniquely describes a person’s desire. Moreover, the query looks quite sufficient for validating the search.
Walmart , by the way, could have done better.
Are there any other similar requests that clearly describe the intentions of the buyer? In fact, there are a lot of such requests. Summarizing, we can distinguish at least two classes of search queries:
- Requests with a unique name for a specific product
- Inquiries with a unique brand name and a specific product
Thus, it is not so difficult to create a fully automated tool for basic validation of a search solution. The percentage of buyers with clear intentions is large enough (depends on the specific store), moreover, it is buyers with clear intentions who quickly lose their patience without getting what they want. Another example of unopposed results is Maxxis Bighorn:
Bad strategies or Skelita comes to you
By the way, "negative options" are very important important. For example, if there are not some products in the store assortment, you should not “find” them. And even more ridiculous are the results obtained by unjustifiably simplifying the initial search query (a very old approach, by the way):
Recognizing and working correctly with the intentions of the buyer is the key to success. The next article will be devoted to automated tools and significant budget savings for testing. High-quality search without a large team of testers is quite real.
miscellanea
- At the time of writing, no search service was harmed (but many turned red).
- “Bad” search in Mvideo
- Funny people can take here .
- The author does not recommend buying a skeletool (fragile), and tire standards :)