How Search Engines Answer Questions
The primary purpose of a search engine is to help users complete a task (and, of course, to sell advertising).
Sometimes this task may involve the acquisition of complex information. Sometimes the user just needs one answer to a question.
In this chapter, you’ll learn how search engines determine which category a query belongs to, and then how they determine the answer.
How Search Engines Qualify Query Types
Whole articles, or probably books, could be written on this one issue.
But we will try to sum it all up in a few hundred words.
Just to get rid of it, RankBrain has little to no role here.
So what is really going on?
Basically, the first step in the process is to understand what information is being requested.
That is, classify the query as a who, what, where, when, why, or how query.
This classification can take place regardless of whether these specific words are included in the query, as illustrated by:
So what we see happening here is two things:
- Google has determined that the user is looking for an answer to a question as the likely primary intent.
- Google has determined that if this is not the primary user intent, the secondary intents are likely different.
You might be wondering how search engines can tell that the user is asking a question in the second example above. It’s not built into the query, after all.
And in the first example, how do they deduce that the user is looking for weather information in their location rather than in general.
There are a number of systems that connect and provide data to create this environment. Basically, it is based on the following elements:
We tend to think of a request as a single request with a single response. This is not the case.
When a query is executed, if there is no known probable good intent, or when the engine may want to test its assumptions, one of the methods it has is to create canonical queries.
Google described the process in a patent granted in 2016 titled “Evaluating Semantic Interpretations of a Search Query” (link is to my analysis for easier reading).
In short, the problem is summarized in the following image:
A query with several possible meanings.
In the patent, they describe a process by which all possible interpretations could be used to produce a result. In short, they would produce one result set for all five queries.
They would compare the results of queries 204a, 204b, 204c, and 204d with the results of 202. The one in the 204 series that best matches the one in 202 would be considered the likely intent.
Judging by the current results, it looks like 204c won:
Which would have required two rounds of this process.
The first to select the films, the second to select which film.
And the fewer people click on a search result from that page, the more successful the result will be considered, which is described in the patent in the statement:
“Using search results to evaluate different semantic interpretations, other sources of data such as click data, user-specific data and others that are used when producing search results research are taken into account without the need for further analysis.”
From the context of the patent, this does not mean that the CTR is a direct measure. In fact, this statement is more akin to what John Mueller meant when he answered a question about Google’s use of user metrics:
“…it’s something that we look at through millions of different queries and millions of different pages, and we generally see if this algorithm is going the right way or is this algorithm going the right way meaning.”
Basically, they don’t just use it for the success of a single result, they use it to judge the success of the SERPs (including the layout) as a whole.
Google uses neural matching to essentially determine synonyms.
Basically, neural matching is an AI-driven process that allows Google (in this case) to understand synonyms at a very high level.
To use their example, this allows Google to produce results such as:
You can see that the query is for an answer to why my TV looks strange, which the system recognized as a reference to “the soap opera effect”.
The ranking page does not contain the word “strange”.
So much for keyword density.
Their AI systems look up synonyms at a very complex level to understand what information will fulfill an intent, even when it’s not specifically requested.
There are a variety of examples and areas where situational context comes into play, but fundamentally we need to think about how query intent varies based on situational conditions.
Above we mentioned a patent on systems that create canonical queries. Included in this patent is the idea of creating a model.
A pattern that could be used for other similar queries to start the process faster.
So while it took resources to determine that when someone types in a single word that tends to have broad context they probably want a definition, they can apply it more universally, producing results like :
And from there, start looking for patterns of exceptions, like food.
And speaking of food, this is a great example that supports my (and I think logical) belief that search engines are also very likely to use search volumes.
If more people search for restaurants than recipes for a term like “pizza” I think it’s safe to say that they would use that as a metric and know if a food item doesn’t follow that pattern, then the pattern may not apply.
Based on models, I think it is very likely, if not certain, that starting datasets will be used.
Scenarios where engines form systems based on a real understanding of what people want, programmed by engineers, and models are generated.
Dave sat down at the Googleplex, wanted pizza, googled [pizza]got a top 10 list, thought, “That’s silly,” and began working with the team on a model.
I actually haven’t read anything about seed sets in this context, but it makes sense and most certainly exists.
Search engines will test if their understanding of an intent is correct by placing a result in an applicable layout and seeing what users do.
In our context above, if a possible intent of the “what’s the weather like” query is that I’m looking for an answer to a question, they’ll test that hypothesis.
It seems like on a large scale, that’s an answer people want.
So what does this have to do with answering questions?
To understand how Google responds to questions, we first needed to understand how they can put together the data to understand if a query is a question.
Of course, it’s easy when it comes to a who, what, where, when, why, or how query.
But we need to think about how they know that a query like “weather” or “meme” is a query for specific information.
This is a Five Ws query without any W (or an H for that matter).
Once this is established using an interconnection of the techniques discussed above combined (and I’m sure I’ve missed a few), all that remains is to find the answer.
So, a user typed in a single word, and the engine jumped through its many hoops to establish that it’s likely a specific response request. Now they have to figure out what that answer is.
For that, I recommend you start by reading what John Mueller has to say about code snippets and work your way up based on your business.
Featured Image: Paulo Bobita
Screenshots taken by the author