Random thoughts

When Google engineers noticed that Bing unexpectedly returned the same result as Google for a misspelling of tarsorrhapy, they concluded that somehow Bing considered Google’s search results for its own ranking. Danny Sullivan ran the story about Bing cheating and copying Google search results last week. (Also read his second article on this subject, Bing: Why Google’s Wrong In Its Accusations.)

Google decided to create a trap for Bing by returning results for about 100 bogus terms, as Amit Singhal, a Google Fellow who oversees the search engine’s ranking algorithm, explains:

To be clear, the synthetic query had no relationship with the inserted result we chose—the query didn’t appear on the webpage, and there were no links to the webpage with that query phrase. In other words, there was absolutely no reason for any search engine to return that webpage for that synthetic query. You can think of the synthetic queries with inserted results as the search engine equivalent of marked bills in a bank.

Running Internet Explorer 8 with the Bing toolbar installed, and the “Suggested Sites” feature of IE8 enabled, Google engineers searched Google for these terms and clicked on the inserted results, and confirmed that a few of these results, including “delhipublicschool40 chdjob”, “hiybbprqag”, “indoswiftjobinproduction”, “jiudgefallon”, “juegosdeben1ogrande”, “mbzrxpgjys” and “ygyuuttuu hjhhiihhhu”, started appearing in Bing a few weeks later:



The experiment showed that Bing uses clickstream data to determine relevant content, a fact that Microsoft’s Harry Shum, Vice President Bing, confirmed:

We use over 1,000 different signals and features in our ranking algorithm. A small piece of that is clickstream data we get from some of our customers, who opt-in to sharing anonymous data as they navigate the web in order to help us improve the experience for all users.

These clickstream data include Google search results, more specifically the click-throughs from Google search result pages. Bing considers these for its own results and consequently may show pages which otherwise wouldn’t show in the results at all since they don’t contain the search term, or rank results differently. Relying on a single signal made Bing susceptible to spamming, and algorithms would need to be improved to weed suspicious results out, Shum acknowledged.

As an aside, Google had also experienced in the past how relying too heavily on a few signals allowed individuals to influence the ranking of particular pages for search terms such as “miserable failure”; despite improvements to the ranking algorithm we continue to see successful Google bombs. (John Dozier's book about Google bombing nicely explains how to protect yourself from online defamation.)

The experiment failed to validate if other sources are considered in the clickstream data. Outraged about the findings, Google accused Bing of stealing its data and claimed that “Bing results increasingly look like an incomplete, stale version of Google results—a cheap imitation”.

Whither clickstream data?

Privacy concerns aside—customers installing IE8 and the Bing toolbar, or most other toolbars for that matter, may not fully understand and often not care how their behavior is tracked and shared with vendors—using clickstream data to determine relevant content for search results makes sense. Search engines have long considered click-throughs on their results pages in ranking algorithms, and specialized search engines or site search functions will often expose content that a general purpose search engine crawler hasn’t found yet.

Google also collects loads of clickstream data from the Google toolbar and the popular Google Analytics service, but claims that Google does not consider Google Analytics for page ranking.

Using clickstream data from browsers and toolbars to discover additional pages and seeding the crawler with those pages is different from using the referring information to determine relevant results for search terms. Microsoft Research recently published a paper Learning Phrase-Based Spelling Error Models from Clickthrough Data about how to improve the spelling corrections by using click data from “other search engines”. While there is no evidence that the described techniques have been implemented in Bing, “targeting Google deliberately” as Matt Cutts puts it would undoubtedly go beyond fair use of clickstream data.

Google considers the use of clickstream data that contains Google Search URLs plagiarism and doesn't want another search engine to use this data. With Google dominating the search market and handling the vast majority of searches, Bing's inclusion of results from a competitor remains questionable even without targeting, and dropping that signal from the algorithm would be a wise choice.

Should all clickstream data be dropped from the ranking algorithms, or just certain sources? Will the courts decide what constitutes fair use of clickstream data and who “owns” these data, or can we come up with a technical solution?

Robots Exclusion Protocol to the rescue

The Robots Exclusion Protocol provides an effective and scalable mechanism for selecting appropriate sources for resource discovery and ranking. Clickstream data sources and crawlers results have a lot in common. Both provide information about pages for inclusion in the search index, and relevance information in the form of inbound links or referring pages, respectively.

Dimension	Crawler	Clickstream
Source	Web page	Referring page
Target	Link	Followed link
Weight	Link count and equity	Click volume

Following the Robots Exclusion Protocol, search engines only index Web pages which are not blocked in robots.txt, and not marked non-indexable with a robots meta tag. Applying the protocol to clickstream data, search engines should only consider indexable pages in the ranking algorithms, and limit the use of clickstream data to resource discovery when the referring page cannot be indexed.

Search engines will still be able to use clickstream data from sites which allow access to local search results, for example the site search on amazon.com, whereas Google search results are marked as non-indexable in http://www.google.com/robots.txt and therefore excluded.

Clear disclosure how clickstream data are used and a choice to opt-in or opt-out put Web users in control of their clickstream data. Applying the Robots Exclusion Protocol to clickstream data will further allow Web site owners to control third party use of their URL information.

Labels: bing, google, microsoft, seo, technology, webdevelopment