Deciphering Trends In Mobile Search Maryam Kamvar and Shumeet Baluja Google

Understanding the needs of mobile search will help improve the user experience and increase the service’s usage. An analysis of search data from a large US carrier showed that cell-phone subscribers are typing longer queries in less time and clicking on more results.


ust as computer-based Web search has been a gateway to increased data consumption, mobile search will help meet the growing user demands for anytime, anywhere data access. With 76 percent of the US population, or 233 million people, subscribing to cell-phone service in 2006 (http://ctia.org/media/industry_info/index.cfm/AID/ 10323), the potential impact of wireless applications is enormous. Understanding the unique needs of mobile search will help improve the user experience and increase the service’s usage.

OVERVIEW OF MOBILE SEARCH We analyzed data from more than 1 million page-view requests randomly sampled from Google logs during a one-month period earlier this year. The requests were anonymous; we maintained no identifying information that could associate searches with users. To eliminate confounding factors between different carriers, we restricted our examination to a single US carrier. To differentiate among computers, PDAs, and cell phones, we looked at the browser’s user agent sent in the HTTP request. Unless otherwise noted, the mobile statistics we present pertain to cell phones. At the time of our study, the Google mobile interface presented users with the option of searching four information repositories: Web (standard Web searches), local (information related to particular geographies), image (keyword-based picture searches), and mobile Web (searches of sites tailored for presentation on mobile 58


phones). To allow accurate comparisons with wired searches, we concentrated our study on Web queries. We grouped the requests into sessions, which we defined as “a series of queries by a single user made within a small range of time.”1 We referred to this time range as the session delta and used a session time-out of 5 minutes—we deemed a user’s session closed if no interaction happened within 5 minutes and considered the next interaction to be the start of a separate session. A typical search session from a mobile or wired device consists of • formulating and entering the query, • browsing the provided search results, and • viewing the selected result. Figure 1 illustrates these three steps.

Mobile queries The average mobile query was 2.56 words (median, 2; maximum, 39; standard deviation, 1.7) and 16.8 characters (median, 15; maximum, 224; standard deviation, 9.2). Interestingly, this was similar to the statistics published for PDA and computer-based queries, where the average number of words per computer-based query reported was 2.351,2 and 2.6,3 and per PDA query was 2.64 (median, 2; maximum, 29; standard deviation, 1.57). Despite the drastically different input techniques used, the similarity in median and mean query terms across search mediums might suggest that the number of terms

Published by the IEEE Computer Society

0018-9162/07/$25.00 © 2007 IEEE

Data download/ network latency

Data upload/ network latency

Time to enter a query (seconds)

Data download/ network latency

per query is currently a ground truth for today’s Web search. In fact, a small study Formulate and Browse search View selected enter query results result page done on a speech interface to 4 search also found that the average length of spoken queries to Google was 2.1 terms. Users might have Figure 1. Search-session process. A state diagram illustrates the three steps in a search session. learned how to form queries for today’s search engines to get neither too many nor Google receives Google receives too few search results. front page request query It’s surprising that mobile users don’t enter shorter queries given the difficulty of query input. Mobile users 44.8 seconds have the challenge of entering the query on miniature keypads, most often consisting of a nine-key layout, rather 39.8 seconds than the conventional qwerty layout. Assuming that users input their mobile queries using the multitap technique, they need an average of 40.9 key presses per query (median, 36; maximum, 720; standard deviation, 1.8). User starts User hits User views With multitap, users access letters by repeatedly presstyping search button search results ing the key and the system cycles through the letters in the order they’re printed. Pausing for a set period of time will automatically select the current letter in the cycle, Figure 2.Timeline for query.To account for latency, we subas will pressing a different key. The amount of effort (the tracted 5 seconds from the time it took to return results. number of key presses) required to enter a word on a cell-phone keypad is more than double the effort required 100 to enter a query on a standard qwerty keyboard. 2007 Cell phone It takes users a significant amount of time—an esti80 2007 PDA mated 39.8 seconds—to enter these queries. To compute this number, we examined the amount of time between 60 when a user first requests the Google homepage and when Google receives the query request, as Figure 2 40 illustrates. This number encompasses the time to download the google.com page, input the query, and upload 20 the HTTP request to the server. The average difference between the two requests 0 (including upload and download time) was about 44.8 1-4 5-9 10-14 15-19 20-24 25-29 30-34 35-39 >=40 seconds (median, 34; standard deviation, 37.8). We subLength of query (number of characters) tracted from that 5 seconds to estimate the network latency (the upload and download time) in order to Figure 3.Time to query. Graph of the time it takes to enter a determine the time it took a user to enter a query. Going query versus the length of the query. forward, for all of our estimates of the time it took a user to perform an action, we’ve subtracted 5 seconds to conventional Web searches accounted for less than 10 peraccount for network latency. cent of all queries in 2001.3 The same study reported that As Figure 3 shows, we found that the time to query was the proportion of pornographic queries in conventional proportional to the query length. Furthermore, we found Web searches declined 50 percent from 1997 to 2001. We have two hypotheses surrounding the relatively that time to query was also proportional to ease of input. Although queries from PDAs (which often have qwerty high percentage of pornographic queries submitted in keyboards) were longer than queries from cell phones, the wireless search. First, since wireless search is a more average time to input a query decreased to 30.1 seconds. recent phenomenon than desktop search, it could be folSince users were willing to spend almost 40 seconds typ- lowing the same trend as wired searches. The high pering their query, the next analysis examined the topics they centage of pornographic queries may decline as the were willing to spend so much time querying. Table 1 lists service attracts more users. Second, we speculate that people might feel more the five most popular query categories. The most popular was the adult category, which typically consists of porno- comfortable querying adult terms on private devices. graphic queries. In comparison, pornographic queries on The screen is smaller, so it’s less likely that a passerby August 2007


Table 1. The top five categories in mobile search.


Percent of all queries

Adult Entertainment Internet/telecommunications Lifestyles/online communities Local Other

>25 >10 >4 >4 >4 >45

Cumulative frequency



2007 Cell phone 2007 PDA



After the query

0 Query rank

Figure 4. Cumulative frequency of top queries.The frequency of the top 1,000 queries made from cell phones ranked higher than the frequency of the top 1,000 queries made from PDAs.

will notice the nature of the search. Also, users often consider cell phones personal and private, perhaps even more so than their computers. Thus, there’s a perceived smaller probability of others discovering their search behavior through cached pages, autocompletion of query terms, or URL history lists. Examining the distribution of queries across a broad set of topics, as Table 1 shows, is one method to examine the diversity of search requests received. Another measure of the diversity is to examine what percentage of the total query volume the top-N unique queries account for. The larger the volume accounted for by the top-N unique queries, the less diverse the set of queries received. To analyze this, we used a random sampling of more than 50,000 queries from cell-phone and PDA searches during a month. Figure 4 illustrates the distribution of the top 1,000 queries. The top mobile query accounted for about 0.8 percent of all wireless queries, and the top 1,000 mobile queries accounted for about 17 percent of all cell-phone-based queries. PDA queries had significantly more variation; the top 1,000 PDA queries accounted for about 13.5 percent of all queries. Computer-based queries are even more diverse. A 2005 study showed that the top 1,000 queries from wired search accounted for only 6 percent of all queries.5 One hypothesis for the higher homogeneity of mobile 60


queries is related to the nascent state of the mobile Web itself. People might have adapted their queries to those that return “usable” sites. Usable sites are those that have content that will display well on the search medium (for example, adult content and ring tone sites are “usable” in mobile browsers). Accordingly, desktop browsers are the most advanced, which would lead to a more diverse set of queries.5 PDA browsers are less advanced than desktops (they can often display HTML but not JavaScript), and cell-phone browsers are the least advanced, often capable of displaying only limited XHTML content. A second hypothesis for the decrease in query diversity across wireless mediums is that there’s a smaller user base, and that the user base may share similar profiles (for example, cell-phone searchers are likely to be technologically savvy, and PDA users have a businessoriented bent). Following this hypothesis, because desktop browsers are the most readily available and reach the most users, they generate the most diverse queries.

After issuing a query, the user receives 10 search results. Most users either found what they were looking for on the first page of results or chose not to look further; only 10.4 percent of queries had requests to display more than the initial set of search results. More than 50 percent of queries led to a click on a search result. It took the average user 30 seconds to scan the search results before selecting one. Of those queries that didn’t lead to a click, it’s possible that the user found the answer in one of the Web-page summaries returned with each search result, gave up on the search entirely, or refined the search in a subsequent query. As Figure 1 illustrates, at any point in a search session, a user might choose to modify the original query. The average number of queries per mobile session is 2, (median, 1; maximum, 48; standard deviation, 1.8). Here, we looked at the query pairs that occurred in sessions that had more than one query. Two queries, query 1 and query 2, were considered to be a pair if query 1 occurred before query 2 in the same session. Some 66.3 percent of all query pairs in a session fell in the same category. Furthermore, in all query pairs, the second query was a refinement of the first 58.6 percent of the time. We considered a pair of queries to be a refinement if: • query 1 was a substring of query 2; • query 2 was a substring of query 1; or • the edit distance between query 1 and query 2 was less than half the length of query 2. From this, we inferred that the majority of wireless searchers approach queries with a specific topic in mind, and their search topics don’t often lead to general exploration.

A LOOK BACK About 18 months have passed between this study and our original study of mobile search in 2005.5 While that’s a short period, we already see a few interesting trends emerging. Table 2 summarizes the statistics.

Users type faster

Table 2. Summary of mobile search statistics in 2005 and 2007.

Mobile search statistics Words per query Characters per query Percent of queries that had at least one click Percent of queries that had at least one “more search results” request Time to enter a query* Time between receiving results and clicking on a spelling correction for a query* Time between receiving results and clicking on a search result*



2.3 15.5 <10.0 8.5 56.3

2.6 16.8 >50.0 10.4 39.8

Time to enter a query (seconds)

Although mobile queries have 15.6 15.1 slightly increased in length since 29.1 30.0 2005, the time delta from requesting the Google front page to sub* Assuming 10-second network latency in 2005 and 5-second network latency in 2007 mitting a query has decreased from 66.3 seconds in 2005 to 44.8 seconds in 2007. We suspect part of this difference is due 120 to shorter network latencies, but we estimated that only 30.8 2005 5.5 seconds of the 21.5-second speedup in query entry was seconds 2007 100 due to network improvements. We estimated the improvement in network latency by 80 comparing the 20.1 seconds it took users to accept a 28.1 seconds spelling correction in 2007 to the 25.6 seconds it took 60 users to accept a spelling correction in 2005. Since the user interface for spelling correction remained a con19.4 seconds 40 stant, and since we’ve noticed that most users are likely to accept a spelling correction without browsing the 20 results, we take the difference in these times to be indicative of the improvement in network latency. The graphs of query length versus time to enter a 0 1-4 5-9 10-14 15-19 20-24 25-29 30-34 35-39 >=40 query shown in Figure 5 provide evidence that users are Length of query (number of characters) typing faster (possibly due to better keyboards or more experience with mobile-phone typing). Note that if network latency were the only factor, we’d expect to see a Figure 5. Reduction in query-entry time. Due to faster typing, it constant decrease in time to enter a query across query took less time in 2007 to enter a query than it did in 2005. lengths. However, this isn’t the case; instead, we observed that the time saved on longer queries was improvement in query-entry speed) have encouraged greater than the time saved on shorter queries. more users to interact with the search-results page. Although we find that more users are clicking on the More users are clicking search-results page, the behavior for users who click has In 2005, users followed less than 10 percent of queries remained consistent. The average clicks on search results with at least one click on a search result. In 2007, that per query and number of “more-search-result” requests percentage rose to well over 50 percent. Additionally, per query are similar in 2005 and 2007. the percentage of queries followed by a request for “more search results” increased from 8.5 percent to 10.5 More exploration within a session percent. We attributed the increase in clicks to at least The number of queries per session has increased more two factors. than 25 percent from 2005. Although there’s low catFirst, there have been drastic improvements in the egory diversity within a session (most users stick to one transcoder technology that converts a search-results page category during their search session), we see an increase to a format the user’s cell phone can display. In 2005, the in query diversity within a session. In 2005, the percent transcoder converted HTML to WML, stripping a Web of unrelated consecutive queries was approximately page of all its images and formatting. Now, the transcoder 20-25 percent.5 Unrelated queries aren’t generated by converts HTML to XHTML and retains much more of spell-correction suggestions, and they don’t classify as the formatting and all of the images on the resulting Web query refinements (defined above). In 2007, the number page. Second, we believe that the reduction in time to get of unrelated queries in a session nearly doubled, to to the search results (the shorter network latencies and 41.4 percent. August 2007


the user clicks on a search-result link. We believe this trend will reverse, as it did with wired queries. As evidence, we look to the UK, which is often considered more advanced in mobile Web usage. The UK has a much smaller percentage of adult queries. The confounding possibility that UK users are less likely to want adult content is called into question by an anecdotal study which examines the image-search logs for both the UK and the US. The percentage of queries related to adult content remains consistent across both countries.


Cumulative frequency


2005 2007

15 10 5 0 Query rank

Figure 6. Cumulative frequency of queries.The data compares the frequency of the top 1,000 queries made from cell phones in 2005 to the frequency of the top 1,000 queries made in 2007.

One confounding factor in comparing the two statistics was that in 2005 the measure was taken on consecutive queries, where query 2 occurred directly after query 1 (with no clicks between the two queries). In 2007, the measure was made over query pairs, a less stringent filter where query 2 occurred sometime after query 1 in the same session. However, if we apply the more strict analysis to the 2007 data, we still see an increase: 38.1 percent of consecutive queries aren’t related. A partial explanation for this is that the number of identical consecutive queries decreased from 31.7 percent in 2005 to 4.5 percent in 2007.

Less homogeneous queries As expected, mobile queries are becoming less homogeneous. The top query in 2007 accounted for 0.8 percent of all queries, as opposed to 1.2 percent in 2005. When measuring the cumulative frequency of the top 1,000 queries from a random set of more than 50,000 mobile queries in 2005 and 2007, we observed a decrease from approximately 22 percent to approximately 17 percent, as Figure 6 shows. This may indicate the increasing diversity of mobile Web users and the increased diversity of mobile Web content.

More high-end devices The percentage of requests from PDAs in the search logs used to account for about 25 percent of the number of requests from cell phones (for the carrier studied). Today, the number of queries from the same carrier originating from PDA devices is about the same as the number of queries from cell phones.

More adult queries While the relative order and magnitude of query categories remains the same, the percentage of adult queries increased. We attribute the gain to the transcoder improvements, which no longer strips page images after 62



sing anonymous log data, we’ve presented a brief examination of wireless search patterns for a major US carrier. The strength of such large-scale logs analyses lies in the breadth of data we used. Google is a popular mobile-search site, and analyzing Google’s usage provides a wealth of general quantitative information about search traffic. The weaknesses of this method are that these numbers don’t tell the story behind a user’s experience—we know for what and when a user queried, but have no context for what inspired the search. We also don’t know anything about the user’s demographics. Despite these caveats, we presented a wide assortment of data on the state of wireless search to provide a useful benchmark in the nascent world of research in this area. ■

References 1. C. Silverstein et al., “Analysis of a Very Large Web Search Engine Query Log,” SIGIR Forum, vol. 33, no. 1, 1999, pp. 6-12; www.acm.org/sigs/sigir/forum/F99/Silverstein.pdf. 2. B.J. Jansen et al., “Real Life Information Retrieval: A Study of User Queries on the Web,” SIGIR Forum, vol. 32, no. 1, 1998, pp. 5-17; http://ist.psu.edu/faculty_pages/jjansen/academic/ pubs/sigirforum98/forum98.pdf. 3. A. Spink et al., “From E-Sex to E-Commerce: Web Search Changes,” Computer, Mar. 2002, pp.107-109. 4. A. Franz and B. Milch, “Searching the Web by Voice,” Proc. Conf. Computational Linguistics (COLING), Morgan Kaufmann, 2002, pp. 1213-1217. 5. M. Kamvar and S. Baluja, “A Large Scale Study of Wireless Search Behavior: Google Mobile Search,” Proc. SIGCHI Conf. Human Factors in Computing Systems, ACM Press, 2006, pp. 701-709.

Maryam Kamvar is a software engineer at Google and a PhD candidate at Columbia University. Her research interests include human-computer interaction, small devices, and search interfaces. Contact her at [email protected]. Shumeet Baluja is a senior staff research scientist at Google, where he works on machine learning, wireless application development, and user interaction measurement. Baluja received a PhD in computer science from Carnegie Mellon University. Contact him at shumeet@ google.com.

Deciphering Trends In Mobile Search - Semantic Scholar

Aug 2, 2007 - PDA and computer-based queries, where the average num- ber of words per ... ing the key and the system cycles through the letters in the order they're printed. ... tracted from that 5 seconds to estimate the network latency (the ..... Maryam Kamvar is a software engineer at Google and a. PhD candidate at ...

599KB Sizes 1 Downloads 435 Views

Recommend Documents

Mobile Computing - Semantic Scholar
May 29, 2011 - Advanced virtual security lab enables students to combat simulated cyber attacks. • Financial aid and an interest-free monthly payment plan available ile puting: ing to ... couraging the adoption of such a fusion between phone.

Distance Education Trends: Integrating new ... - Semantic Scholar
Nevertheless, it is second-generation Web tools that promise to take ... Weblogs are best used as student portfolios that keep record of an individual's progress ...

Object Instance Search in Videos via Spatio ... - Semantic Scholar
The dimension of local descriptors is reduced to 32 using PCA, and the number of Gaussian mixture components is set to 512. As in [20], a separate set of images .... 7, 4th–6th rows). 2) Parameter Sensitivity: We test the Max-Path search at dif- fe

Mobile elements reveal small population size in ... - Semantic Scholar
Feb 2, 2010 - and 5,000 bases away the nucleotide diversity is only 166% the genome ..... For all sim- ... Proc Natl Acad Sci USA 104:17614–17619. 16.

Efficient Tag Identification in Mobile RFID Systems - Semantic Scholar
model, we propose efficient solutions to identify moving RFID tags, according to the .... the amount of energy reflected by an object is dependent on the reflective area ...... an alternative scheme to estimate the number of tags. Our methodology ...

Aggregating Bandwidth for Multihomed Mobile ... - Semantic Scholar
Department of Electrical Engineering & Computer Science. University of ..... devices (e.g., cell phone, PDA, laptop, etc.) forming a .... 10 many other intriguing issues beyond the scope of this paper, including policing malicious MC2 community.

Budget Optimization in Search-Based Advertising ... - Semantic Scholar
ABSTRACT. Internet search companies sell advertisement slots based on users' search queries via an auction. While there has been previous work on the auction process and its game-theoretic aspects, most of it focuses on the Internet company. In this

Scalable search-based image annotation - Semantic Scholar
for image dataset with unlimited lexicon, e.g. personal image sets. The probabilistic ... more, instead of mining annotations with SRC, we consider this process as a ... proposed framework, an online image annotation service has been deployed. ... ni

Jul 5, 2013 - eBay is the largest consumer auction platform in the world, and the primary ... posted-prices under standard assumptions and common production costs (e.g., the Peters and Severinov 2006 model for auctions ..... (e.g., prices) between ve

Scalable search-based image annotation - Semantic Scholar
query by example (QBE), the example image is often absent. 123 ... (CMRM) [15], the Continuous Relevance Model (CRM) [16, ...... bal document analysis.

gender discrimination estimation in a search ... - Semantic Scholar
and occupation, has long been considered a possible indication of prejudice ...... The availability of administrative data has made the construction of data sets ..... “An Equilibrium Model of Health Insurance Provision and Wage Determination,”.

Differences in search engine evaluations between ... - Semantic Scholar
Feb 8, 2013 - The query-document relevance judgments used in web search ... Evaluation; experiment design; search engines; user queries ... not made or distributed for profit or commercial advantage and that ... best satisfy the query.

Aggregating Bandwidth for Multihomed Mobile ... - Semantic Scholar
Department of Electrical Engineering & Computer Science. University of ..... devices (e.g., cell phone, PDA, laptop, etc.) forming a ... In fact, the best a node i can do is to offer all of its WWAN bandwidth Oi and compete for ∑ j=i Oj. As a resul

Our Mobile Planet: Ireland - Semantic Scholar
Implication: Having a mobile-optimised site is critical and a cross-channel strategy is needed to engage .... videos of TV shows, TV movies online, etc.) on your ...

In Search of Usable Security: Five Lessons from ... - Semantic Scholar
Such a model ties digital security to physical security. At PARC, for example, users must pre- sent their badge to a system administrator before that ad- ministrator will unlock the enrollment room. Usability studies demonstrate that this approach is

Deploying Google Search by Voice in Cantonese - Semantic Scholar
Aug 31, 2011 - web scores for both Hong Kong and Guangzhou data. Can- ... The efficient collection of high quality data thus became a cru- cial issue in ...

in chickpea - Semantic Scholar
Email :[email protected] exploitation of ... 1990) are simple and fast and have been employed widely for ... template DNA (10 ng/ l). Touchdown PCR.

in chickpea - Semantic Scholar
(USDA-ARS ,Washington state university,. Pullman ... products from ×California,USA,Sequi-GenGT) .... Table 1. List of polymorphic microsatellite markers. S.No.

A Tweet Consumers' Look At Twitter Trends - Semantic Scholar
forms named entity extraction via Natural Language Processing (NLP) using a ... pivoting named entities by Analytics data, like users' geographic locations.

Trends and Lessons from Three Years Fighting ... - Semantic Scholar
a number of key challenges facing app marketplaces that ..... App IDs, and Amazon Affiliate IDs. ..... Table 5: Top 10 regions impacted by malicious extensions.

Global Development Trends, Prospects and Issues ... - Semantic Scholar
Options to match demand growth ... Cost Evaluation and Optimization available? • Cost of nuclear .... Optimization of the grace period (i.e. operator coping time).

A Tweet Consumers' Look At Twitter Trends - Semantic Scholar
Twitter Trends allows for a global or local view on “what's happening in my ... pivoting named entities by Analytics data, like users' geographic locations.

Networks in Finance - Semantic Scholar
Mar 10, 2008 - two questions arise: how resilient financial networks are to ... which the various patterns of connections can be described and analyzed in a meaningful ... literature in finance that uses network theory and suggests a number of areas