Another 10 Ways Search Engines May Rerank Search Results

When a search engine shows you results for a search, the pages shown are likely in order based upon a mix of relevance and importance.

But a search engine doesn’t usually stop there. It may look at other things to filter and reorder search results.

In 2006, I wrote 20 Ways Search Engines May Rerank Search Results, which described a number of ways that search engines may rerank pages. I followed that up in 2007 with 20 More Ways that Search Engines May Rerank Search Results.

I decided that it was time for a sequel or two in this series. I came up with another 25 reranking methods, but decided to stop at 10 in this post.

Many of the following are described in patents, and some of those patents were originally filed years ago – prehistoric times in Web years. The search engines may have incorporated ideas from those patents into what they are doing now, adopted those methods and since moved on to something new, or put them in a filing cabinet somewhere and forgot about them (I’d like the key to that filing cabinet).

More important then knowing whether or not Google or Yahoo or Bing might be using something from within a patent is understanding reasons a search engines might have considered one approach or another.

Understanding that can help give you an idea why a search engine rerank search results, provide you with a starting point for doing research on what the writers of the patents and whitepapers included, and give you insight into some of the assumptions behind how search engines perceive search, searchers, and the Web.

Here are 10 more ways that search engines may rerank search results:

1. Blended and Universal Search

For many years, most search results at the major search engines were limited to lists of links to web pages. Sometimes you would see news results or images, but the most common sets of search results pages tended to be a list of “ten blue links.” Now, you’ll often see maps, pictures, tweets, blog posts from social network connections, recent news, and sometimes even actual links to web pages.

When Google launched Universal Search in 2007, one main idea behind it was to present a wider choice of results from Google’s other search respositories involving maps, pictures, news, books, videos, and others to provide a “truly comprehensive search experience.” Or, to put it another way, Google was getting plenty of Web searches, but nobody was clicking on the tabs above the search box to visit pictures or news or other specialized searches.

Before the Universal Search announcement, Google had been experimenting with providing some non-webpages at the tops of its web search results. This was sometimes called “vertical creep” into organic results or “blended” results. A Google patent on the Universal Search interface was filed back in 2003, and an Official Google blogpost by Marissa Mayer places the start of Universal Search back to a 2001 brainstorming session.

There’s a lot more than ten blue links in Google’s search results these days, as the following results on a search for “elephant” shows:

Example of Google Universal search showing images, a review, news, videos and web pages on a search for elephant.

Another Google patent filing published in 2008 described how these vertical results were interleavened (Google’s word, not mine) into the main web results. Google was more likely to include non-web results if it could place those in places on a page other than just at the top of the page, as described in the Official Google Blog post Behind the scenes with universal search.

While this process of interleavening non web page results into search results doesn’t reorder the web pages in those results, it can push web results down on a page, or onto a next page. Yahoo and Microsoft also blend non web results into what you see on their web search.

2. Phrase Based Indexing

Imagine a search engine looking at the content on one of your pages and identifying which strings of words there fit together into meaningful “good” or meaningless “bad” phrases. It might create an index of good phrases appearing upon pages on the Web, and when someone performs a search, it might look to see which phrases appear upon a certain number of top results for your query. It might then rerank the results by considering how many common “good” phrases co-occur within that set of search results, and give more weight to pages with more of those phrases.

An image from Google's patent on Automatic taxonomy generation in search results using phrases showing illustrating the process of collecting good phrases

That’s one aspect of a phrase-based indexing system that could change the order of search results, as described in a number of Google patent filings. Some previous posts on other aspects of a phrase-based indexing system from Google:

Google isn’t the only search engine looking at reranking approaches using phrase-based indexing. Here’s a post about a Yahoo patent filing on the process:

3. Time-Based Data and Query Log Statistics

When we search, the search engines collect information about our searches, to try to glean the intent behind them. A recent Yahoo patent filing tells us how the search engine may look through query logs to see if there might be a time based aspect to our searches. If there are many previous queries related to ours where there might be a time, such as a year, associated with the query. For example, someone searching for “world cup” this year might see a number of search results showing information about “world cup 2010″ at the top of the results.

Top Yahoo search results on a search for world cup show a lot of 2010 results, and few results from earlier competitions.

A flow chart from the patent filing gives a quick glimpse at the reranking algorithm:
An image from Yahoo's patent on Identifying and Expanding implicitly Temporally Qualified Queries showing how the search engine might attempt to identify queries that have a time-based element to them.

This process could attempt to see if a query has a time-based aspect to it by seeing if a good percentage of queries in its log files indicate a year or some other period of time, or by looking at query sessions of searchers to see if they refined their queries to include a time-based term, or both.

If that analysis indicates a time based element such as a year, it might rerank search results to boost results to include a temporal term, such as a results for “world cup 2010″ ranking higher in a search for “world cup”.

4. Navigational Queries

Some searches tend to be “navigational” in nature. The query is really just a shortcut to get to a specific page. For example, I type “ESPN” into my toolbar search box (Google toolbar, Yahoo Search bar, Bing bar), so that I can visit the pages of ESPN quickly (I always forget that “go” between the ESPN and the .com in the URL).

A Google search for ESPN shows me the homepage to the site; useful since I can never remember the URL.

The search engines have identified a number of pages that are very good matches for these types of navigational queries, and those pages tend to be listed at the top of searches for those terms.

An image from Microsofts's patent on Presenting Search Queries Related to Navigational Search Queries showing how the search engine might try to find the best page for a navigational query.

Which pages tend to be the best results for a navigational query? Here are a few posts I’ve written on how a search engine may decide:

In a white paper written by researchers from both Yahoo and Google (not sure why the tag-team), Expected Reciprocal Rank for Graded Relevance (pdf), describing how to evaluate pages listed within search results, we’re told that “a perfect grade is typically only given to the destination page of a navigational query.”

So, search results may be re-ordered to place a specific page at the top of search results when there is an ideal destination page (or perfect page) for a specific query when that query is perceived as a navigational, like my ESPN shortcut.

5. Patterns in Click and Query Logs

A search engine looking through its query logs might find patterns related to the query terms used in query sessions and in choices of links people click upon. The abstract to the Google patent Rank-adjusted content items tells us that:

Click logs and query logs are processed to identify statistical search patterns. A search session is compared to the statistical search patterns. Content items responsive to a query of the search session are identified, and a ranking of the content items is adjusted based on the comparison.

A Google patent image from Rank-adjusted content items shows the relationship between click and query logs and search sessions in the adjusting of search results.

Imagine a number of people search for “Chevrolet carborator,” then for “Chevrolet Carborator Rebuild Kit,” and followup with a search for “classic Chevy carburetor kits.” They then frequently choose “”. Someone else coming along searches for the same or very similar queries during a query session. The page at “” may be boosted and rank higher in search results for that searcher.

6. Google Trustrank and Yahoo Dual Trustrank

In 2004, a Yahoo whitepaper described how the search engine might identify web spam by looking at links between pages. That paper was mistakenly credited to Google by a large number of people, most likely because Google was trying to trademark the term “trustrank” around the same time, but for different reasons.

An old photo of a man being arrested by a police officer, with a handwritten annotation on the picture.

Surprisingly, Google was granted a patent on something it referred to as Trust Rank in 2009, though it’s a Trust Rank that is very different from Yahoo’s. Instead of looking at the ways sites linked to each other, Google’s Trust Rank looks at how well they trust people who have labeled web pages in annotations of those pages, somewhat like the label scrawled on the old photo above.

A Google patent image from Search result ranking based on trust showing how the reputation of people applying labels might influence how those annotations are used to rerank search results.

Google allows people who create custom search engines to apply “labels” to pages, as well as annotations in other places, such as Google’s Sidewiki. Not surprisingly (good ideas seem to follow a reuse/recycle practice in search engine circles), Yahoo added a social aspect to their Trustrank as well, mixing “trust” in annotations and user-behavior signals associated with pages and Trustrank scores to come up with something they called Dual Trustrank (double the trust, double the fun?).

A Yahoo patent image from Using community annotations as anchortext showing a dashboard of connections which includes a reputation score for each.

Both Google’s Trustrank and Yahoo’s Dual Trustrank could be used in the reranking of web pages in search results.

Do we see hints of an approach like this in Google’s Social Search?

7. Customization based upon previous related queries

The term you just searched for may influence what you see in your next search if they are determined to be related. At least, according to the Google patent Methods and systems for improving a search ranking using related queries

Google sometimes shows an announcement at the top of search results telling you that Google has customized them based upon your location or because of your previous queries. I wrote about this patent in How Searchers’ Queries Might Influence Customized Google Search Results.

A Google patent image from Methods and systems for improving a search ranking using related queries showing how related previous queries might be identified and weighted based upon click data.

Why might Google might consider some queries to be related to others? Some possibilities:

  • Others having used the same sequence of query terms previously (whether once or multiple times),
  • Queries input by a user within a defined time range (e.g., 30 minutes),
  • A misspelling relationship,
  • A numerical relationship,
  • A mathematical relationship,
  • A translation relationship,
  • A synonym, antonym, or acronym relationship, or other human-conceived or human-designated association, and;
  • Any computer or algorithm determined relationship.

8. Being Linked to by Blogs

Microsoft’s Ranking Method using Hyperlinks in Blogs describes how more “PageRank” might be distributed to pages that are linked to by blogs. The patent has a heavy focus on describing how they might distinquish blogs and non-blogs.

I wrote about their approach in Do Search Engines Love Blogs? Microsoft Explores an Algorithm to Increase PageRank for Pages Linked to by Blogs. Why blogs? The patent’s inventors tell us that blogs tend to be:

…frequently updated, more informational rather than personal, and free of spam.

The approach was tested with a large number of pages, but they tell us that it might lose some value as more “spam blogs” become prevalent on the Web.

This particular method may have lost some value over the past couple of years with the proliferation of splogs, and may be a good example of a reranking method that might potentially change over time.

9. By Ages of Linking Domains

While the age or “maturity” of a domain might be something that helps a web page rank higher in search results, a Microsoft patent filing, Ranking Domains Using Domain Maturity, looks instead at the ages of domains linking to another domain.

A page might rank higher if it is linked to by sites that have some age to them, and have been around the block a few times, instead of ones newly on the Web. The patent filing does tell us that the “maturity” of a domain may be reset if the domain expires or changes hands.

10. Diversification of Search Results

The New York Times surprised us in 2007 with Google Keeps Tweaking Its Search Engine. The article introduced the concept of “Query Deserves Freshness” (QDF), which attempts to decide when searchers might want search results with fresher pages or with older pages. It’s a concept worth including in this set of reranking approaches, but it leads to the question “Do the search engines also try to follow a “Query Deserves Diversity” algorithm, to provide searchers with diverse results when queries might have more than one meaning?”

Chances are that they do. When someone searches for a term like “java,” the intent might be to learn more about Java programming, or the island Java, or the beverage Java. A search engine could just show the most relevant and important pages that come up in a search for “java,” (probably the programming language everywhere in the world but the island of Java). But some searchers might be more interested in the coffee or the island. Some diversity in search results may be a good idea.

A Microsoft patent, Diversifying Search Results for Improved Search and Personalization, tells us what they might look at when deciding to diversify query results.

I covered those in Reranking Search Results Based Upon Personalization and Diversification, and it’s possible that Google and Yahoo both look at similar factors in deciding when to diversify the search results that they show.


Author: Bill Slawski

Share This Post On