User Needs, Features and the Science behind Bing

This morning in the wee hours, we launched the Bing preview! I wanted to take this opportunity to provide some highlights and background on how Bing came to be.

What got us here?

Everything in Bing was driven by a focus on truly understanding searchers and their needs. When we looked at how people were using search today, we saw an opportunity to focus on three key trends: 

  • Information overload: With the explosion of digital content on the Web, it is harder than ever to find what you are looking for.
  • Search failure: Search engines are pretty good at lots of things, but the “Back” button is one of the most popular clicks in search today.
  • Complex tasks and decision making: People are spending more and more time in search trying to accomplish tasks and make decisions.

Information Overload:

To give you a sense of what we mean when we say “information overload”: In 1997 there were around 26M pages (URLs) on the Web. Today we estimate there are more than 1 trillion pages of content. In 1997 the Web was mostly text. Today it combines video, images, music, with new data formats emerging every day. The amount of available data has grown exponentially. An average person would need six hundred thousand decades of nonstop reading to read through the information. This has led people to experience information overload, so it’s no surprise that they are increasingly turning to — and more importantly — depending on search to help navigate and use the breadth and depth of Web content.

 Search Failure

Analysis of search usage data across all search engines (using our toolbar logs) shows that only about half of searches are successful on the first query and click.  Twenty four percent of the time, consumers have to perform an immediate re-query or keep searching, and 15% of the time they abandon their search altogether. A click on a search result is usually just the beginning of a search experience and not the conclusion. People perform a lot of back and forth navigation. In fact, the “Back” button in the browser is the most used button in a search session. More than 50% of queries are people returning to previous tasks, yet there is no memory, and no unification between tools. Queries aren’t done in isolation. They are usually part of a session, and these sessions may be long and repetitive. More than 50% of time spent on searching is spent in sessions over half an hour long. Yet, search engines today treat each query as an independent event. 

Complex Tasks and Decision Making

Perhaps the most interesting insight is that people are turning to search engines not only for information, but to help them complete complex tasks and make decisions They no longer want to just find a web page; they want to learn, shop, be entertained, accomplish tasks and make important decisions. This requires better conceptual organization, a unified experience, deeper task specific content selection and support for longer sessions.

Best Results, Organized Experience and Simplified Tasks

 With Bing, we directly focused on addressing information overload, search failures and the shift to tasks and decisions. Our goals for the service were to:

  1. Deliver best results
  2. Organize the search experience
  3. Simplify common tasks and help inform decisions

To deliver the best results, we combined a range of technologies developed over the past five years, with new features available only on Bing. We have done a lot of work to try to anticipate searcher intent and assist consumers – by focusing on relevance, quality of results and direct access to answers. Consumers can get directly to their information, and reduce the amount of back and forth in a given session.

  • Relevance of results and ads – We’ve made significant gains in relevance of both the algorithmic Web results and the advertising.
  • Best Match – This feature is a fast way to identify the best matched site for a given query. Target is a great example – try it here.
  • Definitive card (D-card) – Building on the best match idea, we also offer additional information for some queries including hard to find information such as the customer service information telephone number – try it here for UPS.
  • Deep Links – You’ll see in both of the above examples a feature we call Deep Links. These are the set of eight links in Best Match or D-card under the site description. These offer a fast shortcut deeply into the most relevant parts of that site.
  • Answers – We continue to expand our answers service to cover “people, places and things.” We’ll blog more about the specific answers in future posts.
  • Search History – This feature is aimed at helping you get back to the searches you’ve done before and save you clicks. You can see it in the left-hand Explore Pane.

With Bing, we help make some progress against information overload by offering consumers a more organized experience and a suite of tools to quickly get them to the right information with minimum effort.

  • Page organization– You’ll see that we have made significant changes in how we present the results in Bing, both on the results page where we introduced a three column layout, but also as you navigate from Web results to images, video, maps, local or shopping results. The UI is consistent throughout the experience, meaning you find the same features in the same places no matter where you are on Bing.
  • Explore Pane – We have introduced the new left hand pane, which we call the Explore Pane. This is where we offer features that help with searching. It includes Quick Tabs, Related Searches and Search History. One can think about the Explore Pane as a permanent companion throughout the Bing experience, a bit like the new Office Menus are to the Microsoft Office Suite, ensuring a consistent and contextual experience.
  • Quick Tabs – The Quick Tabs give you easy access to the most common refinements for a given query. This categorization is aimed at making it easy to refine your query and reduce information overload.
  • Preview – With so many clicks coming back from unsatisfactory results, this feature lets you preview what is on a site from the results page, to help Bing “inform the click,” with the goal of reducing wasted trips to sites you weren’t looking for.
  • Improved captions – We have improved the quality of the document summarization to provide you an easy to read snippet to make an informed decision on whether the document is relevant to your query.

Finally we shifted our thinking a bit and moved from single queries to complex sessions, task accomplishment and decision making as organizing principles in areas of shopping, travel, health and local. Searchers’ needs have evolved and search does poorly in these areas. In fact, 66% of people told us that they increased their dependency on search engines to make decisions in the last year. We see this as an opportunity to simplify common tasks. Bing’s new interface allows us to integrate the power of specialized search experiences and tie them to your query.

  • Shopping: For shopping queries you can start at bing.com and quickly and easily move through your shopping tasks using shopping filters, our opinion ranking feature to learn what other people think of the product, read official reviews, and ultimately get cashback on product you buy with our partners.
  • Local: Often people need to find local businesses, and we simplify the task by offering simple filters, opinion ranking for local businesses, quick access to local reviews, high fidelity maps, and our unique Photosynth feature.
  • Travel: For those common travel tasks, we offer a meta-search experience that provides a great search over flights and hotels, and our exclusive Price Predictor technology for flights and Rate Indicator for hotels.
  • Health: With health queries, people are often concerned that they’re not getting reliable content – a classic issue with search overload. Bing brings together comprehensive health articles from a library of trusted resources, including Mayo Clinic, the American Cancer Society, MD Consult, Gold Standard and more.

The Science behind Bing

None of this would have been possible without a bunch of cool computer science and a team who loves to try new and innovative things.

Today we wanted to talk about five efforts that have enabled the features you see in Bing: organizing search results, intent-specific relevance and freshness of the results, enrichment of the search index, storage and computation grid and high scale infrastructure.

In Bing we took a novel approach for organizing our search results. Instead of applying simple classification techniques, we constructed user query and click graphs and used them to build true interaction models that can represent complex user tasks. This has allowed us to adapt from general to intent-specific ranking and to organize results into sets of topics that can be used to help find information, make decisions and complete tasks. We also invested in technologies and algorithms for extracting structure from unstructured data and applying organizational taxonomies.  Even though the organization is not nearly as intuitive as one done by a human editorial process, we are able to achieve it in a fully automated way allowing for fast scaling and reacting to changes.

We continue our focus on relevance, with additional spotlights on real time indexing and ranking, comprehensiveness and quality of content in the index.  We added a number of new relevance features to RankNet (our ranking system) to improve search quality. We also have scaled our experimentation infrastructure to run tens of thousands of experiments in order to improve our results and related searches. We produce specialized nets that can be applied to different query types. We rely on a large number of implicit and explicit metrics when evaluating each experiment to gain a holistic understanding on how the new changes will affect the efficiency and satisfaction of our users.

We have also enriched our index by developing technologies in HTML parsing, core Natural Language Processing, entity extraction, and document classification. We continuously increase the size of our index, while ensuring high quality in the documents we add.  In Bing, we improved our freshness detection and updated algorithms.  We understand documents which change frequently and developed what we call “Superfresh” techniques to ensure that the index always has the most real-time content.

In the last 6 months, we’ve also made large investments in our storage and computation grid. This has significantly increased our capability for complex data mining and development of new algorithms. Our data-parallel execution environment and computer grids have allowed us to do arbitrary computations and analysis on petabyte sized chucks of data in minutes.   

Finally, to make all this possible we expanded our infrastructure to hundreds of thousands of machines. This has been made possible through full automation of deployment, monitoring and self repair of each server. Our internal systems not only deploy new machines across geographically distinct datacenters, but can dramatically reduce or fully eliminate the need for human deployments or monitoring in our core infrastructure.

More to Come…

Today I’ve given you a quick overview of many of the features and new thinking that’s gone into making Bing possible. In the coming weeks, we will go deeper and you can expect to see a whole series of blog posts that go behind the scenes of the product. We’ll have some of our best minds drill down on the technical areas and give you a peek at some of the juicy computer science problems we are working to deliver and improve the features you see in Bing today.

But until then, now it’s time to BING IT ON!

Satya Nadella, Senior Vice President, Research and Development, Online Services Division