Search As Conversation: Surf Canyon’s Mark Cramer
In my last Just Behave post, I talked to Mark Cramer about his search plug-in, Surf Canyon. The plug in dynamically reorders your results based on the results you click on, taking the click as a signal of intent. This time, I’d like to present the part of the interview where Mark and I explored […]
In my last Just Behave post, I talked to Mark Cramer about his search plug-in, Surf Canyon. The plug in dynamically reorders your results based on the results you click on, taking the click as a signal of intent.
This time, I’d like to present the part of the interview where Mark and I explored the main theme for this series: where search is going. The interview was done before the release of Google Instant, so it is interesting that we jumped almost immediately to the concept of a search experience as a conversation, moving away from the “stateless” form of searching we’re all familiar with.
Of course, Google Instant adds a much more dynamic dimension to the search experience. Mark emailed me right after the release saying Google Instant was an impressive move forward. I’ll share more of Mark’s comments about the benefits and limitations of Google Instant in a bit, but first, let’s explore the idea of searching as a conversation.
Cramer: I was at the SMX Conference in Santa Clara only a few months ago, Chris Sherman put the question (where is search going?) to Peter Norvig, who’s Director of Research at Google. He had an interesting answer, one that I’ve never heard before. He said he saw was search moving away from statelessness towards something more resembling a conversation.
From there I have been speculating on what he might have meant by that, but it’s the first time I’ve ever heard anyone actually, outside of our little company, use the word “state” when referring to search, because search has historically been stateless. For the past 40 years since the inception of search, the way it’s worked is this: user enters a query, that query is matched to an index of documents, tremendous activity is deployed to try to determine the relevancies of those different documents, and a search result set is produced. But that search result set is static. There’s an order to those results. They go from 1 to 10 and then 11 to 50 million, and that order does not change. It’s stateless. What we have been doing for the past few years is essentially applying state to the search page in order to make the results dynamic. And if you consider “dynamic” to be something resembling a conversation in the sense that the search result page is actually responding to every input from the user to alter the content on the fly, then I think that’s an interesting way of looking at searching.
Ironically, this comment from Mark led me to ask how increasingly dynamic search results would be accommodated by Surf Canyon. The more dynamic the results, the more complex the task of reordering becomes. Of course, at the time, neither of us knew Google Instant was just around the corner.
Cramer: We have been seeing elements on the search page becoming more dynamic. If you look at any one particular element, like be it a real-time OneBox on Google with scrolling Twitter or a Twitter information or a map or something that you can expand and collapse—each one of those is an individual element and those individual elements are dynamic. The result set itself is still static. Result number 5 might be a result which has dynamic content in it—previews or things that expand and collapse or content that scrolls through it, but the position of that result is static. And that’s the statelessness that has existed in search for decades.
As the different elements become dynamic, it does add some complexity to what we’re doing… if the user makes a selection on a dynamic result, inferring intent becomes a little more challenging because we actually need to know what the content was in that dynamic result that elicited the selection from the user. However, that being said, actually treating each dynamic result as a unique element in the re-ranking scheme is still something that we’re able to do. So we can move around results, essentially re-rank results whether they be dynamic or not, and essentially change the relevancies of the different results based on the real-time actions of the user.
That was then. What about now? Does Google Instant throw a wrench in the works of Surf Canyon? Does it render it irrelevant? In a follow-up note, Cramer indicated that while he thinks Instant is a step in the right direction, he still feels Surf Canyon serves a critical user need.
Cramer: In a nutshell, it’s very impressive. Not only is it a dazzling user-facing feature, but it’s sure to deliver considerable value to users. It’s a great step forward, but it doesn’t fundamentally affect the value proposition of real-time personalized search.
Speed is an essential element of search, but the future still need to address the issue of “too many results.” Even with Google Instant, short two- and three-word queries continue to chase an exploding quantity of information (hundreds of billions of documents). It does little good to be presented with 50 million results in response to a query when the searcher will only bother looking at the first page because of the cost of digging is so high. Reformulation will help, but unfortunately there’ll continue to be an enormous amount of relevant, valuable information that remains buried beneath page one.
Google has made the search page update in real time as the user types. The next logical step is to make the order of those results update in real time as the user clicks. That, of course, is the thing we’ve been doing for the past 4+ years.
It’s this disconnect between the amount of information available and the amount of information that the user really needs that Cramer identified as the biggest unmet need of search.
Cramer: At a very high level, I think that there has been a divergence between the quantity of information that’s available for searching and the user input required that indicate information need. Cuil, which you’re probably familiar with, is saying that their index is 120-130 billion different items—and it’s growing. You throw in the real-time information and all the real-time tweets and it becomes an enormous, incredible amount of information. (Editor’s note: Cuil ceased operation on September 20, 2010).
On the user side, the number of keywords that are being entered for a query—some say it’s relatively flat, some say that the number of keywords has been increasing. I think in the best case scenarios, we might say that it’s been inching up, but in general you still have two- and three-word queries chasing tens of billions of documents. And so it becomes increasingly difficult to put all the relevant stuff on page one based on the limited keyword input from the user.
There are a number of solutions people have been working on: personalization—looking at the entire history of that user or other users—which results did they click on, what trends are emerging. All that can be very useful. But if we’re looking out 2 or 3 years, the thing that search needs is a search engine for the search results. When you do a search and you get 20 million results, you need a search engine to help you search those search results. Most people can’t dig through 20 million results, nor would they be able to. I think this is something that the industry will be focusing on moving forward.
If there’s too much information available, and the user only needs a fraction of that data, perhaps the answer is to group the information into logical categories and provide search tools more relevant to that category—in other words—vertical search experiences. Other interviews I’ve done in this series, notably John Battelle’s, indicated that a fragmentation of search experiences is one possible scenario. Cramer offered a cautious counterpoint to that view. The more search destinations we have to choose from, the harder we have to work.
Cramer: Fragmentation of search into lots of different vertical search engines or vertical apps does present cognitive load for the user, which basically means that users have to decide “Where am I going to search?” If they’re interested in doing a medical search, they might say, “I need to go to WebMD or PubMed,” if they need to travel, they need to go to Expedia. Cognitive load is always a challenge. Whenever you apply cognitive load to the user, you’re asking them to do work and this can be tough for everyone, including myself—I don’t like cognitive load.
What we’ve also been seeing in search is sort of these vertical components being integrated into the general horizontal search. So if you go to Bing, for example, and you type in “Continental SFO LaGuardia,” you can actually get the flight information, which is very vertical in nature, on the top of the search page. There are ways to navigate to the vertical search site, which is the travel site, and then go ahead and actually book that flight or to get arrival information or whatever you’re looking for. My guess is there’s going to be increasing integration of that into the horizontal search. And if there are specialty-type search engines that have particular functions like features and functionality and even data, people will probably navigate to those through the horizontal search. You’ll always start your search at the horizontal place, the big search engine, and then go to the vertical specialty.
I offered another perspective—the meta-search app. I believe the next thing we may see is applications that act as “middleware” for the various search destinations. These applications could have the benefit of maintaining a personal profile of our likes, interests, histories and personal information, making the heavy lifting of fragmented searching much lighter. Couldn’t, I wondered, an app like Siri also be the future of search? And, if that was the case, couldn’t you use real time signals of the sort Surf Canyon collects, using clicks and user interactions as indicators of intent, allowing the app to continually get smarter and smarter in interpreting that intent? In this scenario, search could truly become a conversation—quite literally if the app is voice enabled as in the case of Siri.
Cramer: I think Siri is the perfect example of that—Not to geek out too hard here, but when Captain Kirk is speaking to the Enterprise, he doesn’t get the list of top 10 results, right? He says, “Computer, tell me what is this alien life force,” and then they have a conversation and they get the answer. We’re obviously not there yet, but…
Before we spun too far into unexplored galaxies, I brought us back to earth. Isn’t the whole point of search to make our lives easier? It has to help us find what we’re looking for as quickly and with as little effort as possible.
Cramer: We’ll have to be in violent agreement on this point. Search has to be able to remove work from the user—however that might be defined and whatever the form that might take. Siri is one example of something that’s able to understand the needs of the user without it having to be explicitly input yet still produce some relevant results. That’s fantastic. I’ve seen Q&A-type interfaces where you type in something and the search results are presented along with the suggestion “Did you mean this other thing?”
So, is the “state” of future search a conversation? I suspect that’s one road search will go. But search requirements come in many different forms. Some are best met with an iterative, interactive approach, as in a conversation. Some are better met by intelligent agents that seek out information, aggregate it and even act on it automatically. Even if search stays consolidated in a few destinations, as Cramer predicts, I suspect the functionality of search will become less of a “one size fits all” solution.
Opinions expressed in this article are those of the guest author and not necessarily Search Engine Land. Staff authors are listed here.