Rockstar Consortium US LP et al v. Google Inc

Filing 18

MOTION to Change Venue by Google Inc. (Attachments: # 1 Text of Proposed Order Google Inc's Motion to Transfer Venue, # 2 Index, # 3 Declaration of Abeer Dubey, # 4 Declaration of Sam Stake, # 5 Exhibit 1, # 6 Exhibit 2, # 7 Exhibit 3, # 8 Exhibit 4, # 9 Exhibit 5, # 10 Exhibit 6, # 11 Exhibit 7, # 12 Exhibit 8, # 13 Exhibit 9, # 14 Exhibit 10, # 15 Exhibit 11, # 16 Exhibit 12, # 17 Exhibit 13, # 18 Exhibit 14, # 19 Exhibit 15, # 20 Exhibit 16, # 21 Exhibit 17, # 22 Exhibit 18, # 23 Exhibit 19, # 24 Exhibit 20, # 25 Exhibit 21, # 26 Exhibit 22, # 27 Exhibit 23, # 28 Exhibit 24)(Mann, James)

EXHIBIT 2

A Framework for Targeting Banner Advertising On the Internet
Katherine Gallagher and Jeffrey Parsons
Faculty of Business Administration, Memorial University of Newfoundland
St. John’s, NF, Canada A1B 3X5
{kgallagh, jeffreyp}@morgan.ucs.mun.ca

Abstract
Constraints that limit accurate targeting of advertising in
traditional media may not hold in cyberspace. This paper
presents a model for effectively and efficiently targeting
hypermedia-based banner advertisements in an online
information service. The model takes advantage of
information technology to micro-target banner
advertisements based on individual characteristics of users.
A simple version of the model, which has the virtue of ease
of development, is presented. Enhancements are also
proposed. These require more effort to develop, but may
lead to even more precise targeting of advertisements.
Implementation of this framework may benefit both online
advertisers and online consumers.

1. Introduction
Cyberspace is a rapidly growing new medium for
commerce. To date, a great deal of industry attention has
focused on electronic transactions over the Internet.
Although rapid growth is predicted over the next few years
[10, 17, 21], actual sales thus far have been only moderate:
users appear to regard the Internet primarily as a source of
product information--when it comes time to pay, they prefer
to buy offline by more conventional means [12, 14].
Responding to consumers’ desire for information,
businesses in large numbers have developed sites on the
World Wide Web (WWW or Web). Most commercial Web
sites describe the firm and its products and/or services, and
many offer opportunities for visitors to the Web site to
provide feedback and ask for specific information. As well,
some Web sites collect information from visitors in order to
improve future offerings. Some sites also support ordering
and payment. The interactive potential of Web sites is
particularly exciting, as it facilitates relationship marketing
and customer support, eliminating the obstacles of
geography and time [14, 22]. Not surprisingly, then,
industry and scholarly research has recently focused on
making Web sites more appealing and useful to visitors [13].
However, a Web site can only be effective if current and
prospective customers visit it. Attracting this audience is
currently a major challenge.

In this paper, we address the challenge of attracting a
defined target audience to a Web site via banner advertising.
We propose a framework for effectively targeting banner
advertising in an electronic marketplace in a manner that
benefits both advertisers and consumers. It allows
advertisers to reach consumers who are more likely to be
interested in the products and/or services offered by the
company, and exposes consumers to information about
products and services that they are likely to be interested in
purchasing. Although the framework is discussed in terms
of the Internet, we believe it will be relevant to whatever
form the "information superhighway" eventually assumes.
The framework takes advantage of the capabilities afforded
by information technology for collecting and processing data
about users. The next section examines trends in the
electronic marketplace. Subsequently, the current state of
advertising in this medium is discussed. Thereafter, a
framework for targeting banner advertising, supported by
appropriate information technologies, is proposed. Finally,
opportunities for further research are discussed.

2. Marketing and Advertising in an Evolving
Electronic Marketplace
The Internet began in the early 1970s as a US
government research project designed primarily for the
needs of the military. It expanded in the 1980s to serve the
international academic and research communities [19, 23].
In the 1990s, businesses began to appear on the Internet.
Although accurate estimates are obsolete as soon as they are
made, it is clear that today tens of millions of people have
access to the Internet [16] through over 100,000 computer
networks in 150 countries--and the numbers continue to
increase [14]. Two types of developments are particularly
noteworthy with regard to this growth.
First, a large and ever expanding number of affluent,
educated consumers are using the Internet [11]. This
concentration of very desirable consumers has led to a surge
in commercial interest. Prior to 1990, nodes on the Internet
were predominantly academic institutions. In 1990, about
1,000 businesses had Internet connections. By June 1995,
over 21,000 businesses were online, and the growth in
commercial connectivity shows no sign of slowing [8].
Second, the emergence of the hypermedia-based WWW,

Proceedings of The Thirtieth Annual Hawwaii International Conference
on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE

together with point-and-click multimedia interfaces such as
Netscape, have greatly increased usability of the Internet for
persons without extensive computer training.
The
development of "applet" technology, such as Java, which
allows programs to run on a variety of platforms, increases
the transparency of various Internet services. In other
words, as technology continues to evolve, it is no longer an
obstacle to, but an enabler of, electronic commerce.
In this environment, companies are seeking ways to use
the Internet effectively [1, 3, 13, 22]. One active area in
electronic commerce involves using the Internet as a
medium to communicate persuasive product and service
information via advertisements. These take various forms,
the most common of which are corporate Web sites and
banner advertising. We define a banner advertisement as:
" paid communication (via text, graphics, video and/or
audio) of information about an organization and/or its
products and services
" by an identified sponsor
" embedded within, and visually distinct from,
information provided by an online service
" with hypermedia links to the sponsor’s Web site.
We distinguish banner advertising from simple hypermedia
links (paid or not) to commercial Web sites: banner
advertising conveys a message even if the user does not
follow the link; simple links can only convey a message if
the user follows the link. Banner advertisements are also
distinct from what [14] refer to as "flat ads," single page
advertisements that do not contain hypermedia links. In this
paper, we restrict our discussion to banner advertising that
appears in the course of users’ browsing and searching
activities on information services, such as Yahoo!
(http://www.yahoo.com) and Excite
(http://www.excite.com), that provide an entry point to
Internet resources.
Appendix 1 shows a banner
advertisements by the Saturn automobile company.
Scant attention has been paid to banner advertising by
researchers. This may be because banners seem relatively
insignificant, especially when compared with the interactive
richness of Web sites. Technical specifications for banner
advertisements severely limit creative options and preclude
any consumer-firm interaction beyond the consumer’s
selection of the hypermedia link to the associated Web site
(Excite, for instance, specifies that "all banners are 468x60
pixels, gif format only, maximum file size is 7k" [9]).
Banner advertisements are, however, very important and
interesting when viewed as part of a system that converts
browsers and searchers into Web site visitors and,
ultimately, customers. In their model of this conversion
process, Berthon, Pitt and Watson [3] identify a sequence of
tasks. First, users must be made aware of the Web site, then
they must be attracted to and locate the site. Once users
have found the Web site, the task is to turn that hit into a

visit, ensuring there is some meaningful contact between
the firm and the consumer; then to convert the visit into a
purchase. The final task is to get purchasers to return to the
Web site and repurchase. Each task in the sequence is
dependent on the successful execution of the previous task.
Our view of the role of banner advertising in this system
is as a mechanism to make target audience members aware
of a firm’s Web site and to attract those users to the site. We
define two concepts critical to understanding this role.
Attraction effectiveness is the number of target audience
members who reach a company’s Web site via a banner
advertisement hypermedia link divided by the number of
target audience members who use the information service on
which the advertisement appears. Attraction efficiency is the
advertising cost per target audience member attracted to a
company’s Web site via a banner advertisement.
There is some evidence that the attraction efficiency of
banner advertising is low. A recent estimate indicates that
only 1-2% of banner advertisements lead viewers to seek
additional information (e.g., by selecting a hypermedia link
to the company’s Web site) [5]. Since information services
charge advertisers based on number of exposures (e.g., [9,
24]), the cost of attracting a single target audience member
to a Web site is at least 50 to 100 times what it would be if
all users who were exposed to the advertisement selected the
hypermedia link. (The cost is even higher if some users
selecting the link are not target audience members.)
Increasing attraction efficiency by reducing wasted
exposures should therefore be a priority. (An additional
motivation for improving performance of banner advertising
in converting searchers and browsers into Web site visitors
arises from recent events such as the agreement between
Procter & Gamble and Yahoo! which provides for payment
based on the number of people who actually seek additional
information (by selecting a link from a banner
advertisement) rather than those who are merely exposed to
the advertisement [20]. Such arrangements are expected to
pressure online services to eliminate wasted exposures [5].)
The estimate cited above does not provide evidence on
the attraction effectiveness of banner advertising. The fact
that only 1-2% of exposed users select a link to the
advertiser’s site is irrelevant to effectiveness if all target
audience members using the information service are among
this group. However, since banner advertisements on online
information services are shown selectively to users, there
will generally be the possibility that some target audience
members who use the information service will not be
exposed to the advertisement and, hence, will be unable to
link to a company’s Web site via it. Depending on the
strategy used to select advertisements for users, a large
number of target audience members may be missed.
We contend that both the attraction effectiveness and
efficiency of banner advertising can be improved by

Proceedings of The Thirtieth Annual Hawwaii International Conference
on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE

precisely targeting advertisements based on characteristics
and behavior of individual users of information services.
Moreover, such targeting can be more precise than the
targeting possible in traditional media. For example, visitors
to a "Travel" page on an information service may be good
targets for an advertisement for discount airfares, as would
readers of the Travel section of a newspaper. But the fact
that the online visitors have made a series of decisions and
taken a series of actions (i.e., selecting only a subset of
highlighted links within a hierarchical menu of categories)
to reach the Travel page, rather than some other page (e.g.,
the Home Decorating page) suggests they may have a
greater interest in travel than, say, readers who
unintentionally come upon the Travel section of a
newspaper and decide to read it. Since these exposures are
more likely to be target audience members, attraction
effectiveness can be improved. Targeting individual users
strategy should also lead to fewer wasted exposures, since
the advertisement would not be shown to users who have not
reached the Travel page, thereby improving attraction
effectiveness. (See Appendix 2 for a similar example.)
At present, targeting of banner advertising does not
always occur. For example, Appendix 3 shows an
advertisement for Honda that appeared when Organic
Gardening was selected from a hierarchical menu of
categories. People interested in organic gardening may not
be the best prospects for automobiles, as they are likely to be
more environmentally sensitive than the general population
and may feel that cars unnecessarily harm the environment.
Nevertheless, online information services do currently
provide some targeting capability. As of August 1996, both
Yahoo! [24] and Excite [9] offered advertisers three options:
general rotation, geographic or content targeting, and
keyword-based targeting. With "general rotation," banner
advertisements rotate randomly through user searches and
browsing on the site. The Honda advertisement that
appeared on the Organic Gardening page in Appendix 3 was
probably in general rotation. Restricted rotations permit
advertisers to purchase space in specified content areas or by
geographic region. For example, financial institutions can
limit the exposure of their banner advertisements to users
searching or browsing Business categories, and Canadian
advertisers can choose to have their banner advertisements
shown only to users who are searching or browsing in the
Yahoo! Canada site. These two options are analogous to the
targeting offered by traditional media such as newspapers,
magazines, television, and radio [4].
The third option, keyword-based targeting, makes greater
use of the targeting potential of information services. A
company can buy keywords so that whenever a user enters
one of those keywords during a search, s/he will be exposed
to the company’s banner advertisement. This ensures that
the banner advertisement is presented only to people with a

demonstrated interest in the area. For instance, a marketer
of golf equipment might buy the keyword "golf." Every
time a user enters "golf" in a search, a banner advertisement
for the equipment would appear. This is analogous to the
more precise targeting provided by magazines.
While these are useful strategies, they fail to take full
advantage of the targeting potential of banner advertising.
Current technology provides the capability to develop
sophisticated and detailed profiles of individual users of
information services based on individual characteristics and
past patterns of behavior in using the information service.
The next section proposes and describes informally two
versions of a model for targeting banner advertising by using
the information technology on which an online information
service is built.

3. A Model for Targeted Advertising
In traditional media, the quality of the information
available constrains an advertiser’s ability to target
advertising effectively and efficiently. For example, many
media buying decisions are based on data provided by
research bureaus such as the Audit Bureau of Circulations
(ABC), Business Publication Audit of Circulation (BPA),
Arbitron, and A.C. Nielsen, which collect data on the
demographics and media habits of consumers, and
sometimes on product usage and brands [4]. These survey
data are cross-tabulated to develop a profile of the audience
of each media vehicle. The audience profile is then
compared to the target audience profile identified by the
advertiser to determine where there is a good match. For
instance, an automobile manufacturer might identify the
target audience for an advertisement for a particular model
of car as middle-income females, 18 to 34, with busy
lifestyles. Based on research bureau data, as well as the
experience and judgement of the media planner, media
vehicles with good reach in that demographic group would
be chosen. Realistically, though, this type of targeting is
usually very approximate. For instance, no matter how well
the media vehicle audience profile matches the target
audience profile, it is likely that only a portion of the
audience would be in the market for a new car.
Online banner advertising may be able to overcome this
problem. It is possible to target users very precisely because
data can remain associated with individuals, so advertisers
can select exactly the users to whom they wish their
advertising to be exposed. It may be possible, for example,
to identify which users will be in the market for a new car in
a particular year. The remainder of this section describes
two versions of a model for targeting banner advertising by
taking advantage of the technological capabilities of the
online environment. The model is designed to be
appropriate for use by information services which sell

Proceedings of The Thirtieth Annual Hawwaii International Conference
on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE

advertising space.

3.1. Basic version
The basic version of the model requires that users be
assigned unique identifiers (e.g., user accounts) when they
first connect to the information service. Subsequently, they
provide these identifiers each time they connect. Users also
complete an online questionnaire the first time they use the
information service.
(Incentives to complete the
questionnaire may be provided by informing users that the
information will be used to filter out advertising for products
in which they are likely not to be interested.) The
questionnaire allows data to be collected on several
dimensions, including: (1) demographic attributes such as
geographic location, income, family lifecycle stage,
occupation, and sex; (2) psychographic attributes such as
travel patterns and hobbies; and (3) product and brand usage
attributes. This element of the basic model permits a banner
advertisement to be directed to users (and only those users)
who fit certain criteria, assuming data were collected on
relevant attributes. For instance, a banner advertisement for
baby strollers could reach parents of children under five
years old--and only individuals in that group.
In contrast, research bureau data uses demographic
correlates (e.g., males and females, 18 to 34) to identify
media vehicles that attract a relatively large proportion of
the people in the identified demographic group [4]. The
media vehicles thus chosen may miss members of the target
group (e.g., older parents) and reach consumers not in the
target group (e.g., people who are between 18 and 34 but do
not have young children). Even audience data based on
cross-tabulations, while they supply information on more
variables, still cannot isolate individuals who are in the
target audience. (For example, research bureau data may
allow an advertiser to identify a magazine whose audience
includes a large number of people between 18 and 34 who
have young children, but there will still be some readers who
are not in the target market.)
The second element of the basic model involves eliciting
the target audience profile from advertisers. An advertiser
can specify a target audience using any number of attributes
about which data have been collected. These can be
expressed conjunctively and/or disjunctively. For example,
a specification may indicate that an advertisement is to be
presented to all users who (1) have household incomes over
$50,000, and (2) either work in a job that involves travel at
least four times per year or have travelled on vacation in at
least four of the past five years.
In this version of our model, the questionnaire
determines the data collected about each user. The content
of the questionnaire will vary depending on the nature of the
information service, expected users, and expected

advertisers. However, it is imperative to design the
instrument carefully, in consultation with advertisers based
on anticipated relevant target audience attributes.
The final element of the model consists of a mechanism
to select banner advertisements to display to users. The
target audience profiles supplied by advertisers provide a
screening mechanism over users. Each time a user connects,
his/her profile is compared to all target audience profiles
from all advertisers. The user’s profile will actually match
some subset of those profiles. If the number of matches is
small (and the session is long), it will be feasible to display
all banner advertisements associated with the matched
profiles during the user’s session. However, if the number
of matches is larger (or the session is short), presenting all
advertisements associated with the matched profiles may
overwhelm the user. In such a case, it will be necessary to
present only a selection of the identified target
advertisements. A rationing system would be needed so that
users are not deluged with banner advertisements while
advertisers are assured of access to users who match the
target audience profile.
In summary, the basic model has three elements:
individual user profiles, individual advertisement target
audience profiles, and a selection mechanism for presenting
advertisements to specific users who match the target
audience profile. This framework potentially eliminates
wasted exposures and provides the capability to reach every
single user who matches the target audience profile (this
may not be realized if a rationing system is used). Users
also benefit, since they will see advertisements only for
products likely to be of interest to them.

3.2. Enhanced Version
The basic version of the model relies on users
completing a questionnaire when they initially use an
information service. This is a straightforward mechanism to
collect data about user characteristics for the purpose of
targeting advertisements. A similar approach has been
incorporated in a commercial product for use with online
catalogs to direct shoppers to products in which they are
interested [15]. However, the advantage of simplicity is
offset by several potential limitations.
First, such
information may become outdated, sometimes quickly, as
user preferences and characteristics change. To some
extent, information can be kept up-to-date by either
readministering the questionnaire periodically or giving the
user the opportunity to update her/his information (e.g., by
a menu option or hypertext link) each time s/he connects to
the information service. However, each of these strategies
is intrusive and may impose an unwarranted burden on users
in order to maintain currency of information.
A second, and perhaps more serious limitation of the

Proceedings of The Thirtieth Annual Hawwaii International Conference
on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE

questionnaire strategy is that it is subject to two potential
types of bias. First, the questionnaire designer will want to
identify as many user attributes relevant to potential
advertisers as possible. As the number of attributes
increases, so does the length of the questionnaire, creating
the possibility of higher mortality in completing the
questionnaire (especially since it may be more difficult to
induce users to complete it because they are both physically
and psychologically remote), thereby increasing the
potential nonresponse bias [7]. Second, the questionnaire
method is plagued with well-known problems, such as errors
due to inaccurate recall, telescoping, social desirability
concerns, and cognitive biases, as well as ambiguity,
intimidation, confusion, and incomprehensibility [2].
In view of these potential problems, it is appropriate to
enhance the model so that it does not rely on user selfreports, can accommodate changing user characteristics and
preferences, and is less constrained by the choice of
questions. Fortunately, information technology may provide
assistance in each of these areas.
Current technology allows a considerable amount of data
about user search activities (both deliberate search and
browsing) to be collected unobtrusively and analyzed to
determine patterns. (We are dealing here only with the
capabilities of the technology, not with the ethical issues
such capabilities raise. However, we recognize that ethical
issues must be considered explicitly in the design of systems
based on our model. For instance, we believe users should
be aware that such information may be collected, and how
it may be used, and consent to this activity before using an
information service.) In the enhanced model, we propose
that patterns of search and browsing behavior exhibited by
users while using an information service determine which
advertisements are shown to that user during current or
future sessions. In the remainder of this section, we provide
a general overview of this approach.
As before, this model relies on assigning a unique
identifier to each user for recording her/his searching and
browsing activities while using the information service.
Each session constitutes a "record", consisting of data such
as: sites visited in order; pattern of navigation through a
hierarchical category structure (as in Yahoo!); choice of
search terms in keyword-based searches; and reaction to
previously exposed targeted banner advertisements (e.g.,
which linked Web sites are selected and visited by the user
and which ones ignored). The aggregate of such records for
each user provides a profile from which preferences can be
implicitly generated. As a simple example, if a user has
made several searches using keywords such as "Atlantic
salmon" and "fly fishing", and has visited the site of the
Angling Club Lax-a of Iceland
(http://www.ismennt.is/fyr_stofn/lax-a/uk/angl_uk.html),
s/he may be targeted for a banner advertisement for a fishing

lodge in Alaska. However, if a user has previously been
exposed to the same or similar banner advertisements but
has not visited linked Web sites when there was an
opportunity to do so, s/he may not be shown these banner
advertisements in future.
This version of the model has the advantage of
transparency. A user simply visits a service for whatever
purpose s/he has in mind. Data are collected unobtrusively
in the course of the visit. Moreover, the data reflect actual
user behavior, rather than attitudes, intentions, or reported
behavior captured through a questionnaire. Hence, the
quality of data derived from user behavior should be
superior to that of questionnaire data, for purposes of
targeting advertisements.
A disadvantage of this model is the preparatory work
involved on two fronts. First, it is not clear how to structure
the data collected during visits so that useful information can
easily be coded for storage and later extraction. Research is
needed to develop useful and efficient coding mechanisms
for storing such data as sequences of visits and search terms
used. We expect this can be handled using conventional
database structures such as relations (tables); however, the
design of a relational database for this purpose is itself a
distinct research issue. Second, the ability to store the
required data does not necessarily mean useful information
can be extracted from it. Further research is required to
determine the types of analyses that yield insights into user
characteristics and preferences hidden in the data.
The enhanced model should be used in conjunction with
the basic model. A questionnaire may be very effective for
identifying various demographic data relevant to advertisers
but impossible to ascertain simply from users’ online search
and browsing behavior. However, since demographic data
has limitations for effectively targeting consumers of most
products, the enhanced model of data collection may yield
complementary data on preferences from patterns of online
search and browsing behavior.
The next section describes an implementation
architecture for the basic version of the model. Extensions
that support the enhanced version of the model remain as
future research.

4. An Implementation Architecture
The architecture required to implement the basic version
of the model consists of two parts: data structure to
represent user profiles and target audience profiles, and an
algorithm to select banner advertisements to display to a
user. This section describes these components.

Proceedings of The Thirtieth Annual Hawwaii International Conference
on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE

4.1. Data Structure
To target banner advertisements, two types of profiles
are needed: profiles describing users of the information
service; and profiles describing the target audience for
advertisements, as defined by advertisers. Each profile can
be modeled as a set of attributes.
We assume there is a finite "universe" of attributes, A
= <a 1,...,aN>, that may potentially characterize users or target
audience members.
4.1.1. User Profile. Each user, ui, of the service can be
described by a record consisting of values of the universe of
attributes, Ri = <a1 (ui ),...,aN(ui )>, where an(u i) (n=1,...,N)
denotes the value of attribute an for user ui . This may be
implemented in a relational database in which a table is
defined whose primary key is a user identifier, and
remaining attributes are those in A. Each row in the table
contains the profile of one user. (A more elaborate data
structure is needed to support the enhanced model, since
data must also be kept about the pattern of behavior of a
user over one or more sessions.) All attributes need not be
applicable or relevant to a particular user; hence, null values
are permitted.
A simple example serves to illustrate this structure.
Consider a universe consisting of three attributes: age,
income, and number of dependents. Suppose there are two
users of a service. When those users have completed a
profile questionnaire, the resulting data may be stored in a
relational table as:
USER
user_id age
u1
26
u2
45

income
34000
54000

dependents
0
2

4.1.2. Target Audience Profile. A target audience profile
is associated with each banner advertisement. A profile may
be expressed as:
(1) A characterization of an "ideal" target audience
member.
Such an ideal can be described by a record consisting of
values of the universe of attributes, Ti = <t1,...,t N>, where
tn (n=1,...,N) is a specific value of attribute a n . Some
values may be null, indicating that any values of those
attributes are permitted for the ideal; and/or a
(2) A characterization of the "acceptable" target audience.
Generally, an advertiser is interested in reaching those
within specified ranges of the attributes of interest.
Given N attributes of interest, acceptability can be
thought of as a region in N-dimensional space. This
region can be defined by specifying ranges of acceptable
values for various attributes in the universe.

Combinations of attributes may be expressed:
conjunctively, indicating that users in the target region
must satisfy all the conditions or restrictions;
disjunctively, indicating that acceptable users must
satisfy one of a set of conditions; or using a combination
of disjunctions and conjunctions.
Note that "distance" from the ideal point may become
relevant if an advertiser has to choose a subset of users
whose profiles fall within the acceptable region.
Operationally, profiles for ideal or acceptable users can
be maintained in a relational database structure. In the case
of ideal profiles, a table can be defined in which each row
describes the ideal target audience member for each
advertisement. The primary key for this table consists of an
identifier for the advertisement, while the remaining
attributes are those of the universe of attributes of interest.
Since not all attributes may be relevant in specifying an
ideal, null values are permitted.
To illustrate, consider a simple example in which there
are two advertisements, each with a different target audience
profile, designated T 1 and T2 . The ideal target profile for T1
is users aged 35 with incomes of $50,000 (no restrictions on
number of dependents), while that for T 2 is users aged 25
with incomes of $25,000 and no dependents. These profiles
are shown in the following relational table.
TARGET
ad_id
age
T1
35
T2
25

income
50000
25000

dependents
0

In this example, the space of profiles is threedimensional. Since the "dependents" attribute is not
relevant in describing the target audience of the first
advertisement, the ideal profile can be depicted as a point in
two-dimensional space, as shown in Figure 1. In the case of
target audience profiles based on attribute values within a
range, the necessary data structure can be provided by a
table whose rows describe the acceptable ranges of specified
attributes for each profile. Each row in this table provides
a lower and an upper bound on a specified attribute for a
specified profile. The primary key of this table consists of
the identifier of the profile plus the name of the attribute.
To illustrate, consider a variation of the example above.
Suppose the profile T1 is no longer age = 35, income =
$50,000, but is relaxed: the advertiser will accept any user
between 20 and 50 with an income between $40,000 and
$60,000. Similarly, suppose the profile T2 is relaxed to
encompass ages from 20 to 30 and incomes from $20,000 to
$30,000. Such profiles might be stored as follows:

Proceedings of The Thirtieth Annual Hawwaii International Conference
on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE

RANGE
ad_id
attribute
T1 age
20
T1 income
40000
T2 age
20
20000
T2 income

lower upper
50
60000
30
30000

In this case, the acceptable profile can be depicted as a
region in two-dimensional space. Figure 2 shows the profile
for T 1.
It is possible that both ideal and acceptable profiles could
be generated by the same advertiser. By overlaying Figures
1 and 2, shown in Figure 3, we note that the ideal point need
not lie at the geometric center of the acceptable region.
Figure 1
Ideal Profile
60000
50000
40000
Income
30000
20000
10000
Age
10

20

30

40

50

Figure 2
Acceptable Profile
60000
50000
40000
Income
30000
20000
10000
Age
10

20

30

40

50

Figure 3
Combined Profiles
60000
50000
40000
Income
30000
20000
10000
Age
10

20

30

40

50

To handle measures of "distance" from an ideal, ranges of
values on relevant attributes can be replaced with advertiserspecified information about the acceptable distributions of
values over attributes. For instance, an advertiser may
specify a mean (ideal) and standard deviation for an attribute
if "acceptability" is normally distributed about a central
value. Other measures of central tendency and dispersion
may be appropriate for attributes in which the range of
acceptability is quantified differently. The data structure of
the RANGE table can be modified to accommodate this
additional complexity.

4.2. Selecting Advertisements for Users
The primary challenge in effectively and efficiently
targeting banner advertising is matching user profiles with
target audience profiles. Figure 4 uses a data flow diagram
to depict the matching process described below.
When a registered user visits an information service,
his/her profile is retrieved. This profile is then compared
with the target audience profiles of banner advertisements
currently being run by the information service. Each target
audience profile is associated with a banner advertisement.
For each target audience profile, if there is no match with
the user profile, the associated advertisement is dropped
from further consideration.
After the comparisons are completed, a set of matched
target audience profiles from a variety of advertisers
remains. If this set is small, it may be feasible to show all
the associated banner advertisements to the user during the
session. In general, though, it will be necessary to select
some subset of advertisements from the matched set to
display to the user. We envision that the advertisers whose
advertisements are in the matched set will compete for the
opportunity to have their banner advertisements displayed
to the user.
The concept of acceptable regions in target audience
profiles provides a basis for competition. Profiles
accommodate the possibility that some users within the
region of acceptability may be more desirable to an
advertiser than others. Hence, a distance metric capturing
the relative desirability of a user with respect to an ideal
profile is possible. It is not the purpose of this paper to
propose or evaluate metrics. However, recognizing a notion
of distance allows the possibility for advertisers to "bid" for
the opportunity to display an advertisement to a user. Such
bids would be determined by the advertiser, based on
variables such as the user profile (to determine the distance
from the ideal target audience profile) and advertising
budget. It may be feasible to automate this by having
software agents associated with each advertisement that
would calculate the distance measure for the user and
formulate a bid based on this, in addition to other

Proceedings of The Thirtieth Annual Hawwaii International Conference
on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE

Figure 4
Selecting Advertisements for Users

Users

User ID

Retrieve
User
Profile

User Profile

Match
User and
TA
Profiles

User Profile

Target Audience
(TA) Profile File

Request
Bids

Compile
Bids

TA Profile

User File

Matches

Questionnaire
Data

Ranked Bids

User Profile

Bid

Ad File
User Profile

Banner
Ad

Select
Ad(s) for
Display

TA Profile

Update
User
Profile

Advertisers

Banner
Advertisement

information such as whether the user had seen this
advertisement, or other advertisements for the same or
similar products, in previous sessions (information which
could be carried as part of the user profile).
When bids are received, they can be ranked. The banner
advertisement corresponding to the winning bid is displayed
to the user. Other advertisements may be displayed
according to their ranking if there is an opportunity to
display additional advertisements (e.g., if the user engages
in several search or browse activities during a session).
This architecture provides guidance for implementing
the basic version of the model. We present next a simple
example showing how the architecture operates.

4.3. Example
Consider the relational database tables USER,
TARGET, and RANGE presented earlier. Suppose first that
the user with profile u1 connects to the information service.
This user’s profile, consisting of the database record <u1, 26,
34000, 0> is retrieved from the USER table. Next, target
audience profiles T1 and T 2 are retrieved from the TARGET
table. These identifiers determine the attributes whose
profile ranges have to be selected from the RANGE table.
Next, the age range for T 1, namely (20,30), is retrieved from
RANGE. Since the age value of u1 is 26, there is a match on

this criterion. So, the salary range for T1 , (40000,60000) is
retrieved. Since the user u1 does not match this criterion of
the target audience profile (salary is 34000), the
advertisement corresponding to the profile T1 will not be
shown to u1 . Applying the same operations to the target
audience profile T2 , u 1 would not be exposed to the
advertisement corresponding to T2 since the income of u1
(34000) is greater than the upper bound of 30000 specified
in the target audience profile T2. Thus the user with profile
u1 would not be exposed to any banner advertisements when
s/he used the information service. This is efficient, since
showing either banner advertisement to the user with profile
u1 would entail a cost and constitute a wasted exposure.
Suppose now that the user with profile u2 connects to the
information service. This user’s profile, <u2 , 45, 54000, 2>,
is first retrieved from the user profile file. The target
audience profiles T1 and T2 are then retrieved. Applying the
matching algorithm, a match will be found between u2 and
the profile T 1. (Note that the target audience profiles in our
example do not specify restrictions on number of
dependents; hence, any values are permitted on this
attribute.) However, there is no match between u2 and T2
since the income of u2 (54000) is beyond the upper bound
of income for T2 (30000). Hence, the user with profile u2
will be exposed to the advertisement corresponding to the
target audience profile T 2.

Proceedings of The Thirtieth Annual Hawwaii International Conference
on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE

This example does not show the full scope of the model,
since there is no case where there are two or more target
audience profiles that match a particular user profile. To
illustrate this, consider an additional target audience profile
having only the condition that user income must be at least
50000. This profile would require adding the following
record to the RANGE table: <P3 , income, 50000, - >. Now
if the user with profile u2 connects to the information
service, a match with both T 2 and T3 will be found. In this
case, the advertisers (or software agents) responsible for T2
and T3 will be contacted and provided with the profile u2 .
Each advertiser (agent) will prepare a bid indicating how
much it is willing to pay to have the banner advertisement
corresponding to its profile exposed to the user with profile
u2. These bids are compiled and returned to the information
service, where they are ranked. If we allow that the user
will be shown only one advertisement, the one which placed
the highest bid will be chosen.
In summary, this model makes use of rich, multiattribute
data at the individual user level in determining whether each
one will be exposed to a particular banner advertisement.
This leads to more effective and efficient targeting than is
possible using strategies such as general rotation, which
does not use data at the individual level, and restricted
rotation or keyword search, which rely on only a single data
item about an individual in determining which banner
advertisement(s) to present to the user. It is also more
effective and efficient than targeting in the traditional media,
which does not use any data at the individual level.

5. Future Research
This paper has presented a framework for leveraging
information technology to target online banner advertising
more effectively to benefit both users (who would be
exposed only to advertising that is very probably of interest
to them) and advertisers (whose advertisements would reach
only those users who fit the target audience profile). This
framework is, however, merely a starting point. Additional
research on several fronts is needed before its potential can
be realized. Several specific research concerns have already
been noted. In addition, there are more general issues.
First, a system supporting the basic version of the model,
based on the implementation architecture presented in this
paper, should be implemented.
Second, an implementation supporting the enhanced
version of the model is needed. This will require research to
develop a more sophisticated database structure that can
preserve users’ searching and browsing behavior over time.
In addition, techniques for detecting patterns of behavior are
needed.
Third, both theoretical and empirical research is needed
to explore agent bidding in the context of the framework

proposed in this paper.
Fourth, empirical work needs to be done to evaluate the
relative effectiveness and efficiency of this framework. A
priority should be to compare the (1) the basic version of the
model, (2) the enhanced version of the model, and (3)
existing approaches to targeting advertisements. For
instance, it would be interesting to test whether placing a
banner advertisement on a relevant page (e.g., an
advertisement for a new movie on the Entertainment page of
an information service) would be more or less effective than
directing the same advertisement to individual users selected
on the basis of their answers to a questionnaire (i.e., the
simple version of the model) or their search and browsing
behavior (i.e., the enhanced version of the model).
Finally, the utility of this framework in other online
contexts should be investigated. For instance, this approach
could be used in developing Web sites that are more useful
to visitors.
Visitors with different profiles could
automatically be shown different pages more likely to be of
interest to them, eliminating the need for them to search the
Web site for the information they desire.

6. Conclusion
Cyberspace is a new medium for advertising. In 1994,
Edwin Artzt, chairman of Procter & Gamble, the largest
advertiser in the United States, warned advertising agencies
to "get their interactive act together" [6, p. 75]. As the
advertisements in Appendices 1 and 3 show, even major
advertisers and their agencies may not be taking full
advantage of the opportunity to target their online banner
advertising. The information technology that makes the
WWW possible also allows the unobtrusive collection of
detailed information about user interests based on their
online searching and browsing. Advertisers should not
assume that the same constraints that make media planning
in traditional media a very inexact science also apply to
online advertising.

7. References
1. Armstrong, Arthur and John Hagel III (1996), "The Real Value
of On-line Communities," Harvard Business Review 74(3),
134-141.
2. Barnes, James G. (1991), Research for Marketing Decision
Making. Toronto, ON: McGraw-Hill Ryerson.
3. Berthon, Pierre, Leyland F. Pitt, and Richard T. Watson
(1996), "The World Wide Web as an Advertising Medium:
Toward an Understanding of Conversion Efficiency," Journal
of Advertising Research 36(1), 43-54.
4. Bovée, Courtland L., John V. Thill, George P. Dovel, and
Marian Burke Wood (1995), Advertising Excellence. New
York, NY: McGraw-Hill.
5. Business Week, June 3, 1996, 44.
6. Castro, Janice (1995), "Just Click to Buy," Time, Spring, 74-

Proceedings of The Thirtieth Annual Hawwaii International Conference
on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE

Proceedings of The Thirtieth Annual Hawwaii International Conference
on System Sciences ISBN 0-8186-7862-3/97 $17.00 © 1997 IEEE

1060-3425/97 $10.00 (c) 1997 IEEE