 |
Internet Marketing · eMarketing · Internet Advertising · Online Branding |
|
Internet Marketing · eMarketing · Internet Advertising · Online Branding |
|
|
|
||
| ADVERTISING BY ICONOCAST |
||
Internet Advertising: Is Anybody Watching?
Xavier Drèze
University of Southern California
François-Xavier Hussherr *
Ecole Nationale Supérieure des Télécommunications
August 1999
Internet Advertising: Is Anybody Watching?
Xavier Drèze
University of Southern California
François-Xavier Hussherr *
Ecole Nationale Supérieure des Télécommunications
August 1999
This is a Draft, do not circulate or quote without prior consent from authors.
*Xavier Drèze is an Assistant Professor at the Marshall School of Business, University of
Southern California, Los Angeles, CA 90089 (xdm@sbaxdm.usc.edu). François-Xavier
Hussherr is a researcher at the Centre Nationale d’Etude des Télécomunications (CNET) and at
the Ecole National Supérieure de Télécommunication (ENST), Paris, France
(francois.hussherr@francetelecom.fr).
The authors would like to thank Voilà and France Telecom for funding this research and Philippe
Taupin for his help with the Eye-Tracker device.
2
Internet Advertising: Is Anybody Watching?
Abstract
Click-through rates have emerged as the de facto measure of Internet advertising
effectiveness. Unfortunately, click-through rates are plummeting. This decline prompts four
critical questions: (1) why do banner ads seem to be ineffective; (2) what can advertisers do to
improve their effectiveness; (3) does an immediate measure such as click-through rate undervalue
online advertising; and, (4) are memory-based measures such as recall or awareness more
appropriate? To address these questions, we utilized an eye-tracking device to investigate online
surfers’ attention to online advertising. Then we conducted a large-scale survey of Internet
users’ recall, recognition, and awareness of banner advertising.
Our research suggests that the reason why click-through rates are low is that surfers
actually avoid looking at banner ads during their online activities. This suggests that the larger
part of a surfer’s processing of banners will be done at the pre-attentive level. If such is the case,
click-through rate is an ineffective measure of banner ad performance. Our research also shows
that banner ads do have an impact on traditional memory-based measure of effectiveness. Thus,
we claim that advertisers should rely more on traditional brand equity measures such as brand
awareness and advertising recall.
Finally, our study shows that although repetition leads to lower click-through rates, it has
a beneficial impact on brand awareness and advertising recall.
Keywords: Internet, Advertising, Awareness, Brand Equity, Eye-tracking, Click-through.
3
Introduction
As the Internet matures into a viable commercial medium, many web sites (e.g., Lycos,
Go Network, Yahoo!) rely on advertising to finance their operations. The lure of advertising is
such that some companies provide users with free Internet access (e.g., NetZero.com, FreeI.com)
and even free computers (e.g., Free-PC.com) in exchange for their eyeballs (Berst 1999). This
should not come as a surprise as advertisers have long used every conceivable vehicle to display
their messages in front of the gazing eyes of potential customers, be it magazines, television, or
racecars.
As the Internet becomes more mainstream, many companies are budgeting significant
dollar amounts for online advertising. The Internet Advertising Bureau (1999) reports 1998
online advertising expenditure of $1.92 billion, more than double 1997 revenues. The bulk of
this expenditure is allocated to banner ads. Banner ads typically consist of rectangular images
displayed at the top of web pages and contain the message that the advertiser wants to send to
web surfers.
The most widely used measure of online advertising effectiveness is the percentage of the
total number of ad exposures that induce the surfer to actually click on the banner in response to
the advertised message. This measure is known as the click-through rate (Novak and Hoffman
1997). Click-through rate has become such a dominant measure that in 1996 Procter and
Gamble made a deal with Yahoo! in which P&G would pay only for click-throughs and not for
exposures (Associated Press 1996). The ability of a site to generate click-throughs also affects
the advertising rates it can command (Hamilton 1998).
Click-through rates started in 1996 at around seven percent. However, they have
declined steadily to around 0.6% in 1999 (Nielsen//Netratings 1999). This is problematic
4
because advertisers typically do not knowingly allocate budgets to media that are not effective.
Does this trend suggest that the online advertising community is going to fritter along with the
decline in click-through rates? Should one sell his or her stocks in Yahoo!? Not necessarily.
Many authors (Ambler 1998, Batra, Lehmann, Burke, and Pae 1995) argue that good advertising
affects long-term brand equity, not necessarily short-term sales. They contend that equity
variables such as brand or ad awareness are better gauges of advertising effectiveness. In this
spirit, Briggs and Hollis (1997) have shown using Milward Brown’s Brand Dynamics™ system
(Dyson, Farr, and Hollis, 1996), that banner ads can have an impact on consumers’ attitudes
toward a brand independent of click-through.
Briggs and Hollis’ study, combined with the decline in click-through rate, begs the
questions (1) why do banner ads seem to be ineffective; (2) what can advertisers do to improve
their effectiveness; (3) does an immediate measure such as click-through rate under-value online
advertising; and, (4) are more traditional measures such as recall or awareness more appropriate?
The purpose of this paper is to answer these four questions. We intend to show that because
banner ads operate mostly at the pre-attentive processing level (Shapiro, MacInnis, and Hoyer
1997), traditional effectiveness measures are more appropriate than click-through rates. We will
then use these measures to study some of the factors that might impact banner ad effectiveness.
The remainder of the paper is organized as follows. The first section discusses the results
of a study which utilized an eye-tracking device to determine whether web surfers see banner ads
and which factors increase or decrease the probability that a banner ad is seen. We use the
results from this first study to generate hypotheses about the characteristics of banner ads that
might increase or decrease viewers’ attention. The following section relates the results of the
follow-up study that tested the hypotheses generated in the first study on a broader sample of
5
web surfers (807 respondents). The study also explored the effects of Internet advertising on
recall, recognition, and awareness. We then consider the results of both studies and discuss their
managerial relevance. Finally, we close with concluding remarks, a discussion of the limitations
of our methodology and results, and directions for future research.
Study 1: Eye-tracking
The Internet differs from traditional media in at least one significant way. When an
advertiser uses Television or Radio to deliver his messages, he preempts the program being
broadcast (e.g., a sitcom or song) and uses all the bandwidth of the medium to transmit his
message. This means that by default, the viewer or listener is paying attention to the advertisers,
and the message is only interrupted if the listener zaps away. Zapping, however, is quite
infrequent. Siddarth (1999) reports zapping rates for commercials of less than 3%. By contrast,
online banner ads share their bandwidth with other elements of the pages in which they are being
displayed. A banner ad typically occupies less than 10% of the area of a web page on a standard
VGA computer screen (640x480 pixels). Therefore, the attention of the web surfer is generally
focused on other elements of the page. The task of the banner ad is to first grab a surfer’s
attention and second to induce the surfer to click on the ad. If surfers never look at a banner,
they cannot click on it!
Shared bandwidth might explain why click-through rates are low, but not why they are
declining. There is some evidence that some online surfers dislike banner ads (Bass 1999). This
dislike is widespread enough that various software exist that actually prevent browsers from
downloading ads (AdsOff!, @Guard, JunkBuster...). One can thus hypothesize that, as surfers
gain more familiarity with the medium, they learn to differentiate informational content from
advertising. Ultimately, this would give them the ability to disregard banner ads.
6
Given this possible learning and avoidance behavior, we start our investigation by
measuring the extent to which surfers pay attention to banner ads. We begin by formulating the
following two hypotheses:
H1: Internet users avoid looking at banner ads.
H2: The more time users have spent on the Internet, the less they pay attention to
banner ads.
To test these hypotheses, we asked a group of subjects to look at various web pages while
hooked up to an eye-tracking device that records their eye movements and fixations. Eyetracking
studies are not new. Javal (1878) used eye-tracking to study reading patterns more than
100 years ago. Although reading studies are still being conducted through eye-tracking (Hyönä
1995), a growing number of eye-tracking studies have recently addressed marketing problems.
For instance, Russo and Leclerc (1994) studied in-store brand choice, Fischer et al (1989) studied
warning labels on tobacco ads, Janiszewski (1998) looked at exploratory search behavior with
catalogs, Kroeber-Riel (1979) investigated the effect of arousal on advertising copy processing,
and Lohse (1997) studied Yellow Page advertising.
Most of the eye-tracking studies, including ours, use Pupil Center/Corneal Reflection
(PCCR) monitoring devices to track the eye movement of their subjects. These devices
illuminate a subject’s eyes with a near-infrared LED while a video camera collects images of the
eyes. From these images, a computer calculates the position of the center of the pupil and the
specular highlight of the LED (corneal reflection). From the relative position of the pupil and
the reflection, the computer can recover the location of subjects’ fixation within 1.5 degrees
(Young and Sheena 1975).
7
Study design
Our study was conducted using information portals as a background. The cover study
was an ergonomic research on the design for one of the largest French portals: Voilà
(www.voila.fr). The subjects were asked to perform five searches using three portals: Voilà, an
alternate layout for Voilà (henceforth called Voilà Bis), and Voilà’s largest competitor. Three of
the searches related to general topics (e.g., find information about ‘Le Louvre’), the other two
related to individuals (e.g., find the phone number of ‘Jean Dupont’). Each of the three generaltopic
searches was made using a different portal. The two other searches were made with Voilà
and Voilà Bis.
To accomplish the task they were assigned, the subjects would click on a link or enter a
search string. For the three general-topic searches, this would prompt the display of an answer
page (see Figure 1 for the answer page to Voilà) containing the information they were asked to
look for. They were then asked to indicate with the mouse where the information they were
looking for was located. The answer pages were designed to match the look and feel of the
question page. In other words, the Voilà search page led to a Voilà answer page of similar
design; the Voilà Bis search page led to a Voilà Bis answer page, and likewise for the competing
portal.
=====================
Insert Figure 1 about here
=====================
The order in which the pages were shown was rotated across subjects with the only
restriction being that the three search/answer page pairs were kept unbroken. Eight banner ads
were integrated within the design of the eight web pages (see Figure 1 for the ‘Club Internet’
8
banner of Voilà Answer). At no time before or during the search task on these eight pages was
any reference to banners ads made.
We collected our data in two steps. First, we used an eye-tracking device to collect eye
movements and fixations1 during the experiment. Second we asked our subjects to fill out a
short survey after completing their assigned task on the eight web pages. The survey asked
questions about their Internet savvy, the experimental process (e.g., did they encounter any
stress), their preferences regarding the various pages to which they were exposed, and a series of
questions regarding the banner ads they saw (e.g., do you remember seeing any banner ads).
=====================
Insert Figure 2 about here
=====================
To simplify the analysis of the eye-tracking data, each page was dissected in a series of
mutually exclusive rectangular zones. One zone was created for each paragraph of text, banner
ad, or graphical element of the page (see Figure 2 for the zone definition of Voilà Answer). The
eye-fixation data were then coded by zone. Hence, for every subject we have a list of each zone
that they fixated on during the experiment and for how long. Of the 60 subjects that were
recruited for the experiment, 11 had to be eliminated because they suffered from heavy
nystagmus or because the calibration of the eye-tracking device could not be performed
satisfactorily on them. This left us with 49 usable subjects. The experiment was conducted at
the CNET from December 16th to the 18th of 1998. The subjects were selected through a street
intercept in the center of Paris and paid 100FrF (@$16) to participate in the study.
1 A fixation was recorded whenever a subject stared at an element of a page for at least 100ms.
9
Analysis
The primary goal of this first experiment was to measure the extent to which surfers
actually look at the banner ads that are embedded within the web pages. Each of our 49 subjects
was exposed to 8 banner ads (one per page). Looking at the zones that were focused on by our
subjects, we find that every subject looked at one or more banner ads (i.e., nobody managed to
avoid every ad). On average they looked at 3.96 of the eight banners during the experiment,
which yields a probability of being seen of 0.49 for each individual banner (see Figure 3 for a
frequency distribution of the number of banners looked at). This probability is low relative to
other media such as television (>90%, per Siddarth 1999) or Yellow Page ads (89 percent for
small display ads, 93 percent for large display ads, per Lohse 1997).
=====================
Insert Figure 3 about here
=====================
To test Hypotheses 1 and 2, we created a data file that lists all of the zones that a subject
might focus on (i.e., 111 zones x 49 subjects). Each zone is associated with variables describing
their location, shape, and content (see Appendix 1 for a description of the zone description
variables). We then ran a logit regression using as the dependent variable an indicator that was
set to 1 if the subject fixated on the zone, and 0 otherwise. As test variables we used an Ad
dummy (1 if the zone is a banner ad, 0 otherwise), an Expert dummy (1 if the subjects have been
on the Internet at least 25 times, 0 otherwise), and an interaction term. To control for page
layout as well as possible differences across gender or age, we use the variables described in
Appendix 1 as control variables along with two demographic control variables (Age, and
Gender).
10
=====================
Insert Table 1 about here
=====================
The results shown in Table 4 are very revealing. As one could have expected, a zone’s
location and size are important. The positive coefficient on Area shows that the bigger a zone is,
the more likely it is to capture subjects’ attention. The significant interaction between the page
and the zone interaction shows that the page layout is important. Similarly, the zone’s content is
important as evidenced by the significant content dummies.
The negative coefficient on the Ad dummy provides support for Hypothesis 1. It
indicates that viewers avoid looking at ads. It also indicates that they are able to recognize that
an item is an ad without having to look at it directly. Although the Expert dummy is marginally
significant (p=0.11), the interaction term between expertise and banner is not significant
(p=0.46). Hence, we do not find support for Hypothesis 2.
Although we did not find support for Hypothesis 2, it is still interesting to contrast the
behavior of our various demographic groups. During the experiment, as well as throughout our
analysis, we found significant differences in behavior between novices and experts as well as
between young and older surfers (see Figure 4 for an example of eye movements for both an
expert and a novice). To illustrate these differences, we ran a series of regressions on the
number of fixes, number of zones looked at, and time spent during fixes across these groups. For
each page looked at by each respondent, we regressed the three dependent variables against an
expertise dummy, an age dummy, and a gender dummy, as well as seven control variables to
account for the differences across pages. As the results in Table 2 show, experts tend to process
each page by making fewer fixes, looking at fewer zones, and spending less time then novices.
11
Older people look at the same number of zones as young people, but it takes them longer and
they fixate on a larger number of positions. Finally, males and females seem to behave similarly.
=============================
Insert Table 2 and Figure 4 about here
=============================
As part of the debriefing questionnaire, we asked our subjects if they remembered seeing
any banner ads. Only 46.94 percent of the subjects indicated they did. After asking them if they
remembered seeing any ad, we showed our subject four banners and asked them if they recalled
seeing the ads during the test. Two of the ads were fake ads that had not been part of the test; the
other two were real. We did not find significant differences in recognition level between the
fake and the real ads (m=23.5% vs. 18.4, respectively, p=0.38). The number of false positives
we encountered is similar to those reported in other studies. Janiszewski (1990) reported 21
percent of false positives (119 subjects, 5 different ads). A forthcoming study jointly conducted
by the French division of the Internet Advertising Bureau and SOFRES (1999) reports false
positive levels of 17 percent (6,872 subjects, 14 ads).
Take away from Study 1
Our first study was very revealing. It provides us with an answer to the first question
motivating this study (Why are banner ads not effective?) and a hint of an answer to the second
and third questions (What can advertisers do to improve banner effectiveness? Does clickthrough
rate under-value online advertising?). The study shows that one of the problems
hindering banner ad effectiveness is that half of the banner exposures are not attended to. The
problem is not only that surfers do not look at the banners, but they also seem to purposefully
avoid looking at them (Hypothesis 1).
12
There are at least two possible explanations for this apparently clairvoyant behavior.
First, site designers have traditionally located banner ads at the top of their web pages. This
might lead web surfers to treat as a potential ad every item that is located at the top of the screen.
Second, as has been noted by Janiszewski (1998), peripheral vision allows subjects to recognize
objects that are located outside their focal point of attention (as measured by the eye-tracking
device). This ability, coupled with the fact that most banner ads have the same shape (468x60
pixels) provides web surfers with the ability to train themselves into recognizing banner ads for
what they are without having to actually focus on them. Both of these explanations assume that
surfers learn over time and develop strategies to avoid devoting attention to advertising.
Unfortunately, we did not find support for Hypothesis 2. This probably means that it takes less
than 25 Internet sessions to learn to avoid banners.
That only half the banner ads are looked at is highly detrimental to click-through rates.
One cannot click on something one does not look at! It does not mean, however, that half of the
banner exposures are wasted. Research has shown that consumers do not need to fully process a
message in order to be influenced by it. Janiszewski (1990a, 1990b, and 1993) has researched
the topic extensively. Among other things, his research shows that incidental exposure to
advertising can enhance a consumer’s liking for the ad and brand advertised despite the
consumer’s inability to recognize having previously seen the ad and brand (a situation similar to
ours). Other researchers (e.g., Shapiro, MacInnis, and Heckler 1997) have reached similar
conclusions. This means that a large part, if not most, of a consumer’s processing of banner ads
will be done at a pre-attentive level rather then at a full attention level. Further, it implies that
click-through rates will not capture the full extent of an ad’s effectiveness since pre-attentive
processing does not lead to immediate action.
13
Our study also reveals that experts are more efficient at processing web pages than
novices and that young surfers are more efficient than older ones. This does not, however,
translate into fewer banners seen by experts or young surfers.
As to what factors might help improve banner effectiveness, we found that location, size,
and zone content are important factors when trying to predict whether a zone is attended. We
will further investigate these factors along with the relevance of traditional advertising
effectiveness measures in Study 2.
Study 2: Impact of Online Advertising
Study 1 provides us with somewhat discouraging results about Internet advertising.
Although everybody fixated on at least one ad, only half of the subjects remembered seeing a
banner ad during the experiment. Further, the reported recognition rate for ads they have been
exposed to is similar to the recognition rate for ads they have never seen before.
The purpose of study 2 was to further explore these findings using a different
methodology. Indeed, eye-tracking studies have two major drawbacks. First, they are artificial
environments; second their cost and infrastructure prevent their use on large samples. Hence, we
sought to validate the results of the first study in a much broader experiment (~1,000
respondents). Study 2 first attempts to measure the effects of online advertising using traditional
memory-based measures (i.e., other than click-through), and then establish how the factors
uncovered in study 1 (i.e., size, shape, content) affect advertising effectiveness.
Our study of the effects of Internet advertising uses the same benchmarks for online
advertising as those that have been used for advertising in traditional media. For television or
print ads, advertising agencies use various measures to evaluate an ad campaign (Tellis 1998).
They are interested in such constructs as unaided advertising recall, aided advertising recall,
14
brand recognition, and brand awareness. Consequently, we formulate the following four
hypotheses regarding banner ads:
H3a: Banner ads will have a positive impact on aided advertising recall.
H3b: Banner ads will have a positive impact on brand recognition.
H3c: Banner ads will have a positive impact on unaided advertising recall.
H3d: Banner ads will have a positive impact on brand awareness.
These four measures are sorted by increasing level of advertising effectiveness. Brand
awareness is the most desired effect; aided advertising recall is the least preferred. Should our
analysis reveal that banner ads do indeed have a significant impact beyond immediate clickthrough,
we will be able to use these four measures to analyze the different factors that may
ultimately affect banner effectiveness.
Previous studies (e.g., Chatterjee, Hoffman, and Novak 1998) have shown that repetition
has a negative impact on click-though rates. Click-through rates are maximum after the first
exposure. This is at odds with studies of television advertising (Pechman and Stewart 1989) that
shows how a low level of repetition is beneficial. In light of Study 1’s result that indicate only
50 percent of the exposures are attended to, we believe that the total number of surfers who pay
attention to a banner ad will increase with repeated exposure to the banner. With this increase in
audience, we should see an increase in advertising effectiveness. To resolve the issue of the
positive or negative impact of repetition we phrase hypothesis two as:
H4: Advertising effectiveness increases with frequency of exposure.
Our logit analysis performed in Study 1 also gives us some insights as to what attributes
of a banner ad will affect its ability to attract attention. The factors included in the logit analysis
were size, shape, content, and location. We will not study the impact of location as it is an
15
attribute over which advertisers have little control. As we have mentioned before, site designers
typically use the prime locations to display the informational content of their site, and relegate
banner ads to the top of their pages.
As far as size is concerned, our analysis shows that bigger is definitely better, although
the S-shaped nature of the logit link function implies that at some level one will see decreasing
marginal returns from increases in size. This leads to our next hypothesis:
H5a: Larger banner ads will be more effective than smaller banner ads.
Study 1 also revealed that orientation matters. A vertical zone is more likely to be
attended to then a horizontal one. Consequently, we will have the following corollary to
Hypothesis 5a:
H5b: Banner ads that are laid-out vertically will be more effective than banner ads that
are laid-out horizontally.
Besides location and shape, Study 1 showed that a zone’s content impacts its ability to
attract attention. The next series of hypotheses will deal with this issue. The size of the
CONTRAST coefficient in Table 4 shows that it is an important factor. Hence, we will
formulate our next hypothesis as:
H6a: Banner ads that contrast with their environment will be more effective then ads
that do not.
One characteristic that sets banner ads apart from other elements on a web page is
animation. There is some anecdotal evidence that users are put off by blinking banner ads
(Hamilton, 1998). However, animation provides advertisers with the ability to put more content
in the same space by spacing the content temporally rather than spatially. Animation also allows
16
advertisers to build dramatic pauses by delivering the message sequentially. This increased
flexibility should translate in more effective ads. Thus, we hypothesize that:
H6b: Animated banners will be more effective than static ones.
As they have gained experience with the online medium, advertisers also have become
more and more creative with their use of the limited space available on a banner. They have
gone from simple images to fully interactive games, from static words to audible messages. The
tools they have at their disposal are now richer. This should, hopefully, translate into more
effective ads. This presupposes, of course, that advertisers use these tools effectively and that an
ad’s execution matters.
While it is hard to evaluate the ability of advertisers, it is relatively easy to figure out
whether their ability matters by testing the impact of different executions of the same concept.
We will thus have:
H6c: The effectiveness of a banner ad depends on its message.
In summary, we have generated ten hypotheses that fall into four different types. H3a
through H3d deal with the use of traditional measures of advertising effectiveness to evaluate
online advertising. H4 relates to the effect of frequency on online advertising effectiveness. H5a
and H5b relate to the physical characteristics of the banner box. Finally, H6a through H6c relate to
the content of the banner box.
Study Design
The second study consisted of a two part self-administrated web-based questionnaire.
Respondents were recruited through a mass e-mail sent to 6,000 individuals. The e-mail asked
them to participate in a study about Voilà, the new search engine of France Telecom. The stated
purpose of the study was to better understand how people use the Internet. The e-mail contained
17
an hyperlink that would lead the respondent directly to the web pages containing the survey. As
an incentive to participate, 2 digital cameras and 25 CD-Roms were to be awarded to participants
through a random drawing.
The questionnaire began with three questions addressing a respondent’s current Internet
usage and their brand awareness. Once these questions were answered, respondents were led
through nine web pages in which they had a specific task to accomplish or a question to answer
(e.g., “Enter the word ‘Voyage’ in the SEARCH string input box”). Seven of the nine pages
contained a banner ad. The ads related to five different products. Four of the products were
online services offered by France Telecom: @près l’école (a subscription based homework help
web site), Le Mél (a free web e-mail service similar to Hotmail), Nouba (a web site listing the
time and places of films, plays, and concerts), and Tout En Ville (a city guide about France’s
major cities). The fifth advertised product was Cortal, a large French financial institution. Each
respondent saw one banner ad for each of the five services plus a second ad for both Cortal and
@près l’école, for a total of seven exposures.
Each of the banner ads was executed with different creative specification purposefully
designed to test the hypotheses derived in the previous section. For instance, the ad for Le Mél
had five different executions. The five banners were identical except for their color (blue,
yellow, orange, red, or green). In contrast, Cortal was executed in two different sizes: 230x33
pixels and 468x60 pixels. A total of 18 different banners were used. The details of each
execution will be discussed along with the experimental results in the next section. The order in
which the banners were presented as well as which execution of each banner was used was
counter balanced across respondents.
18
Twenty-four hours after they filled out the first survey and went through the nine test web
pages, the respondents who completed the first task were sent a second e-mail asking them to
fill-out a follow-up survey. This survey asked them several questions about their advertising
recall, and to rate the banner ads to which they were exposed. Of the 6,000 e-mails that were
originally sent, 211 came back as undeliverable. 2,297 of the remaining 5,789 prospects
successfully completed the first part of the experiment. After being contacted the next day, 807
respondent filled out the second survey, for an overall response rate of 13.5 percent. The data
was collected between January 28 and February 4, 1999.
Analysis
Our first order of business was to determine what effects, if any, banner ads have on
recall, awareness, and recognition. We will not report any analysis of click-through rate since
only three click-throughs were recorded during the experiment. Out of 807 respondents, 460 (57
percent) recalled having seen an ad during their task the previous day. Table 3 reports the brands
they recall seeing during the experiments.
=====================
Insert Table 3 about here
=====================
The five brands with the highest recall (11.4 percent average recall) are those that were
advertised during the test. The other brands do not seem to be random brands mentioned just to
satisfy the questionnaire. Voilà is the search engine of France Telecom, the sponsor of the
research. The Voilà logo was prominently displayed at the top of each page. It is not a banner
ad per se, but it could be mistaken for one. Pages Jaunes and Pages Blanches (Yellow and
White pages) are telephone number search services offered by Voilà. Pages Zoom is the former
19
name of these services. Honda and 123 Achats are brands that were heavily advertised on the
web (and on Voilà) at the time of the experiment. Finally, Internet Explorer, Netscape, and
Linux were a set of three buttons that were located at the bottom of each page as they are on
many web sites. Again, they do not constitute banner ads per se, but could be construed to be
advertisements.
To test whether these differences in unaided advertising recall are significant, we ran a
logit regression with 14 observations per subject (1 for each brand mentioned except for Voilà2).
Our dependent variable was set to 1 if the brand was mentioned by the subject, and 0 otherwise.
We then created three dummy variables. EXP1 was set to 1 for each of the five banner ads,
EXP2 was set to 1 for the two banner ads that had received two exposures (Cortal and @près
l’école), finally BUTTON was set to one for Internet Explorer, Netscape, and Linux. The
estimates for the logit regression are shown in Table 4.
=====================
Insert Table 4 about here
=====================
As one can see, the coefficients for both one and two exposures are significant (note that
the total effect for two exposures is the sum of EXP1 and EXP2). This supports H3c regarding the
positive impact of banner ads on unaided advertising recall. BUTTON is not significant;
indicating that the recall of Netscape, Internet Explorer, or Linux cannot be attributed to our
experiment.
=====================
Insert Table 5 about here
=====================
2 We chose to ignore Voilà since some subjects might consider its logo to be a banner ad while others might not.
20
To investigate hypotheses H3a and H3b, we removed the brand name from each of the five
banner ads used in the experiment and showed them to our subjects. We then asked them
whether they recalled seeing the ads during the experiment and, if so, could they name the brands
that are advertised. Table 5 provides aided advertising recall, brand recognition (conditional on
recalling the ad), as well as unconditional brand recognition. We can compare the recall levels to
the level obtained in Study 1 for the ads that had never been seen (Recall = 23.45 percent). We
find that the overall mean recall of 30.1percent is significantly larger then 23.45 percent
(p=0.0001), while the individual ads themselves are significantly larger then 23.45 percent for
Tout en Ville (p=0.0001), Le Mél (p=0.0001), and @près l’école (p=0.0001). The differences
are, however, only marginally significant for Cortal (0.1174), and not significantly larger for
Nouba (0.4635). If we compare these recognition levels to the one obtained by IAB
France/SOFRES (17 percent) on its much larger sample, we find that they are all significantly
larger at the p=0.0001 level. This gives us some support for H3a; the hypothesis that banner ads
have a positive impact on aided advertising recall.
=====================
Insert Table 6 about here
=====================
In terms of brand recognition, we can turn to Table 6, which shows the brands associated
with each ad. As one can see, the overwhelming majority of people make the correct
association. The few people that make incorrect brand recognition tend to mistake Voilà,
Wanado or France Telecom for the correct brand. A c2 test rejects independence at the
p=0.0001. Hence the data supports the hypothesis that banner ads have a positive impact on
brand recognition (H3b).
21
Our last measure of advertising effectiveness is brand awareness. In the screening
questionnaire subjects indicated whether they knew each of ten brands. Five of these brands
were the brands advertised later during the experiment; the remainder were Internet-related
brands (e.g., Yahoo!, Netscape). The same questions were asked the next day in the follow up
survey. Pre- and Post-awareness measures and the number of exposures for each brand are
reported in Table 7. As one can see, the non-advertised brands suffered a slight loss in
awareness whereas the advertised brands enjoyed an increase in awareness. For the purpose of
this particular analysis, we will consider that the effect of the Voilà logo at the top of each page
is similar to that of a banner ad for Voilà, and include the data in the analysis.
Various models of brand awareness have been proposed over time (see Mahajan, Muller,
and Sharma 1984 for an exhaustive review, see also Little 1979). Those models, whether they be
TRACKER (Blattberg and Golanty 1978), NEWS (Pringle, Wilson, and Brody 1982), or
LITMUS (Blackburn and Clancy 1982), assume that exposure leads to increased awareness.
They also assume that people forget about brands in the absence of advertising, and that there is
some saturation level to awareness (typically 100 percent). They are usually estimated using
OLS on time series data. In our case, we possess data for only two time periods. Hence, we
propose an alternative model that takes advantage of the data we have at hand, and still adheres
to the spirit of the previous models in that it takes into account buildup due to exposure, decay
due to non-exposure, and saturation effects.
If one has individual level data for two consecutive periods (as we do) and one considers
that the awareness level of a brand is nothing else than the probability that one remembers the
brand, then one can set up a logit regression to represent the awareness building and decay
phenomenon. We can say that:
22
A
e
e it
t t Exposure
t t Exposure
i i
i i
=
+
- +
- +
a b b
a b b
. . .
. . .
1 2
1 1 2
Where: Ait is the awareness level of brand i at time t;
t is a time index taking the values 0 or 1;
Exposuresi is the number of exposures to brand i between time 0 and time 1;
ai, b1, and b2 are parameters to be estimated.
In this setting, the ais are indicators of each brand’s pre-test awareness (the bigger ai, the
bigger brand i’s awareness). b1 represents the decay occurring in the absence of advertising, and
b2 is an indication of the benefits to be gained from advertising exposure. In addition, this
functional form accounts for saturation effects when the awareness levels become large.
=====================
Insert Table 7 about here
=====================
Following the above awareness model, we ran a logit regression on a data set containing
20 observations per subjects (10 pre-awareness, 10 post-awareness). The data set contained 9
dummy variables for the ten brands, one dummy variable (TEST) to indicate the pre-post test,
and one variable (EXP) indicating the number of exposures to the brand. The exposure variable
is set to 0 for all pre-test data points, and, depending on the brand, 0, 1, 2, or 9 for the post-test
data points. In this setup, the brand dummies control for difference in pre-awareness across
brands, the test dummy indicates the change in awareness occurring after zero exposure, and the
exposure (EXP) variable tells us by how much awareness changes as the number of exposure
increases. The parameter estimates and their p-values are shown in Table 8.
23
=====================
Insert Table 8 about here
=====================
The large variation in brand dummies accurately reflects the differences in brand
awareness. The negative sign on the test dummy indicates that there is a slight decrease in
awareness after 24 hours in the absence of ad exposure. However, it is not statistically
significant in our case. Finally, the exposure variable is positive and significant. This provides
support for H3d, the hypothesis that banner exposure has a positive impact on brand awareness.
One must remember, however, that the logit link function is inherently S-shaped and thus, as the
number of exposure increases, the benefits will decrease. To better illustrate the post-awareness
levels as a function of exposure, one can look at Figure 5 that plots the post-awareness level as a
function of the pre- level (0, 1, 2, and 9 exposures). One can also look at Figure 6 that plots the
post-awareness as a function of the number of exposure for different pre-awareness levels (25
percent, 50 percent, and 75 percent). As one can see, there is a big gain to be earned by having
multiple exposures. Further, the higher the current awareness, the less there is to gain from
additional exposures.
=========================
Insert Figures 5 and 6 about here
=========================
So far we have found support for H3a through H3d. This gives us a benchmark with which
we can test Hypotheses 4 through 6. We, however, already have a partial answer for Hypothesis
4. In our discussion of H3c and H3d we have highlighted the fact that repetition increases both
unaided advertising recall and brand awareness. We now only need to look at its effect on aided
24
advertising recall and brand recognition. To do so, we ran two logit regressions. The first one
had five data points for each respondent (one for each of the banner ads) and used advertising
recall as dependent variable. As an independent variable, we used a dummy variable that took
the value of 0 if the banner was seen once, and 1 if the banner was seen twice. We proceeded in
the same fashion for the brand recognition variable except that we limited ourselves to the
observations where the recall indicator is set to 1. For each of these regressions, a positive and
significant dummy variable indicates that repetition is beneficial.
=====================
Insert Table 9 about here
=====================
The parameter estimates for each regression as well as their significance level are shown
in Table 9. Repetition improves brand recognition, but it does not improve aided advertising
recall. That is, repetition improves unaided advertising recall, brand recall, and brand awareness,
but does not improve aided advertising recall. This data suggests support for Hypothesis 4 that
repeated exposures increase online advertising effectiveness.
The procedure to test the remaining five hypotheses was the same for each hypothesis. In
order to be as concise as possible, we will describe it only once, leaving it to the reader to make
the necessary adjustments to fit each case. We test each hypothesis by fitting logit regression
models on our four advertising effectiveness constructs: unaided advertising recall, aided
advertising recall, brand recognition, and brand awareness. In each case, we set up a series of
dummy variables that reflect the experimental design used, and we test whether these dummies
are significantly different from zero. Rather than report detailed numbers about each of the
regression, we report only the significance level of the dummies as a whole.
25
When conducting these tests, we consider that an ad is more memorable than another if,
at the minimum, it leads to higher brand recognition. A banner that leads to higher aided
advertising recall, but does not improve brand recognition, unaided advertising recall, or brand
awareness will not be considered effective as surfers can remember the banner, but not the brand
that was advertised.
H5a and H5b relate to the size and orientation of banners. We tested three different banner
sizes. We used the traditional 468x60 pixel banner as our null case, and pitted it against both a
quarter (230x333) and a double (468x120) size banner. To test vertical versus horizontal
orientations, we pitted a 144x240 pixel banner against the traditional 468*60. The significance
levels for the logit regressions are shown in Table 10.
=====================
Insert Table 10 about here
=====================
Overall, we do not find any support for H5a. Small banner ads perform just as well as
large ones. This contradicts the findings of Study 1, which showed that bigger is better. The
lack of size effect might be due to the decreasing returns from larger size. The small banner ads
might be big enough to attract attention and the benefits gained from the larger ads might be too
small to be measured accurately given our sample size. We find very weak support for H5b. It
seems that banner orientation affects aided advertising recall, but not unaided recall or brand
recall. It might also have a small effect on brand awareness, although the effect is only
marginally significant (p=0.17). The fact that only aided advertising recall is significant might
indicate that our subjects remember the shape of the banner more than its content.
3 230*33 is slightly larger than a quarter of 468*60, but it is a widely accepted banner size.
26
H6a to H6c relate to the graphical content of banner ads. To test H6a, we used five banner
ads that differ only in their background color (Red, Blue, Yellow, Orange, or Green). Given the
background color of the pages and the color of the elements surrounding the ads, the Yellow and
Orange ads were coded as not contrasting while the other ads were contrasting. To test H6b, we
used a still and an animated banner ad, both displaying the same message. Finally, to test H6c,
we compared four different ads for the same product. Our base case ad is a traditional 468x60
pixel banner ad with the brand name of the sponsor. The second ad is an offer to participate in a
game associated with the sponsor. The third execution has a drop down menu embedded in the
banner ad. A sound file that enunciates the name of the brand being advertised as the banner ad
is displayed accompanies the last one.
=====================
Insert Table 11 about here
=====================
As reported in Table 11, we find a weak contrast effect (H6a). However, the effect goes
in the opposite direction as predicted. High contrast ads seem to perform worse than noncontrasting
one. We do not find any animation effect (H6b), but we find support for H6c. The
fact that brand recognition is significant for H6c is important because it shows that the different
messages used affect not only the recognition of the visual component of the ads as is the case
for H6a and H6b but also the comprehension of the message. Hence, ad content is important.
Further, the message is more important than how the message is sent.
Summary of results
We started with an eye-tracking study conducted on 49 subjects. Study 1 revealed that a
banner ad typically has a probability of about 50 percent of being seen by a surfer looking at a
27
page in which the banner is embedded. This number is dramatically lower than the 97 percent
reported by Siddarth (1999) for television ads or 93 percent reported by Lohse (1997) for Yellow
Pages ads indicating that pre-attentive ad processing might be much more important on the
Internet than in more traditional media. We also found that different levels of expertise, age, or
gender lead to differences in page processing behavior. However, these differences did not seem
to translate into different probabilities of seeing banner ads. A logit analysis of the eye-tracking
data revealed that some of the factors that would most likely influence the probability of content
being looked at on a web page were a zone’s size, shape, content, and location.
We used the results of the eye-tracking experiment to design a large-scale experiment
where 807 subjects were exposed to nine pages containing six banner ads. The banner ads were
designed to test a series of six hypotheses relating to size, shape, color, animation, message, and
repetition. The study was also used to test whether traditional memory-based measures of
advertising effectiveness could be applied to the Internet.
We found that traditional memory-based effectiveness measures provide valuable insight
into the effects of Internet advertising. These measures outperform the immediate effects
measured by the click-through rate. On average, for 100 surfers exposed to a banner ad, 11
recall seeing the ad and can mention the brand name on the ad without any aid 24 hours later (see
Table 12 for summary results). Thirty respondents remember seeing the banner when they are
shown the same banner but without brand name. Of those 30 surfers, 18.5 (62 percent) can name
the banner’s brand. In addition, three of the 100 surfers become aware of the brand. We can
compare these numbers with the average click-through rate of 0.6 percent (Nielsen//Netratings
1999). The effect on unaided brand awareness is 4.5 times larger than the click-through rate; the
effect on unaided brand recall is 19 times larger.
28
=====================
Insert Table 12 about here
=====================
In terms of what factors influence online advertising effectiveness we found that
frequency is important. Repetition affects unaided advertising recall, brand recognition, and
brand awareness. However, execution has little effect. Banner size seems to be unimportant.
Contrast, animation content, and shape of the banner influence aided advertising recall but no
other dependent measure. Nevertheless, a banner’s message influences both aided advertising
recall and brand recognition. This indicates that what an ad says is more important than how it
says it.
Our finding that repetition increases advertising effectiveness might seem to contradict
the results found by Chatterjee et al (1998) in their study of the negative effect of repetition on
click-through. We do not believe they do. If we use the framework of Lavidge and Steiner
(1961), awareness measures cognition while click-through measures action. Click-through
requires the user to see the banner, process it, be convinced by its message, and take action. It is
very unlikely that a user would click on a banner ad after two exposures if he or she actually
processed and rejected the message during the first exposure. Hence the only users who might
click on a banner after multiple exposures are those who have not processed the ad already. As
the number of exposures increases, the proportion of ‘virgin’ users decreases, and thus the clickthrough
rate will decrease too. This might lead Web advertisers to pursue the logic that if shortterm
results are desired, avoid repetition, but if long-term results are sought, repetition should be
encouraged!
29
Conclusion
We started this research to answer four basic questions regarding online advertising.
Given the steady decline in click-through rate over the past three years, we asked (1) why are
banner ads not effective; and, (2) what can advertisers do to improve their effectiveness?
Further, given the focus of the Internet community of click-through rates as a measure of banner
ad effectiveness, we asked (3) does an immediate measure such as click-through rate actually
under-value online advertising; and, (4) are more traditional measures such as recall or
awareness more appropriate?
Our research shows that traditional memory-based measures of advertising effectiveness
should be applied to banner ads as it is to television or newspaper advertising. By focusing on an
immediate measure such as click-through rate, one fails to capture the positive effect that
banners have on the equity of the brand being advertised. To be sure, click-through rates are not
irrelevant. They provide a measure of advertising effectiveness that, along with other measures,
will help managers better allocate their advertising budgets. But click-through rates alone are
insufficient. Managers who rely solely on them are short-selling their advertising campaigns.
Managers must also be aware that click-through rates are still declining and thus inter-temporal
comparisons should be treated cautiously.
As to why banner ads seem to be ineffective, we believe that they are actually effective.
Banners lead to brand awareness. Click-through rate might be low, but in the long run,
awareness is more important than click-through. The low click-through rates are primarily due
to the fact that surfers fixate on less than 50 percent of the banners to which they are exposed.
Not only do they not see banners, but they actually avoid looking at them.
30
Finally, advertisers can increase advertising effectiveness by concentrating on the
message they send. What is said and how often it is said is more important than how it is said.
Bigger ads or animated ads will not compensate for ineffective content.
Putting what we have learned from this study in perspective, we would be inclined to say
that the medium Internet advertising resembles most is outdoor billboards. As with banner ads,
drivers encounter billboards while engaging in other activities. Billboards occupy only a small
portion of their field of vision and typically consist of a simple message and visual. As Donthu,
Cherian, and Bhargava (1993) have shown, billboards influence awareness and recall even if
they only rarely prompt consumers to take immediate action. Thinking in these terms might help
design future research and yield better insight in the mechanisms underlying Internet advertising.
One should also be aware of some of the shortcomings of our research. First and
foremost, all of our subjects were engaged in a goal-directed search. As Janiszewski (1998) has
shown, goal-directed searches are very different then exploratory searches. Second, we
examined memory measures but not behavior measures (e.g., sales or ‘store visits’).
31
References
Ambler, T. (1998), “Myths about the mind: time to end some popular beliefs about how
advertising works,” International Journal of Advertising, Nov, 501-509.
Associated Press (1996), “Procter & Gamble World Wide Web ad strategy raises online ire,” San
Francisco, Apr 28.
Bass, S. (1999), “Unclog the Net: Eliminate Web Ads,” PC World Online, January.
Batra, R., D. R. Lehmann, J. Burke, and J. Pae (1995), “When Does Advertising Have an
Impact? A Study of Tracking Data,” Journal of Advertising Research,
September/October, 19-32.
Berst, J. (1999), “Free PCs, ISPs and Software: Are They Worth the Price?,” ZDNet, February
10.
Blackburn, J. D. and K. J. Clancy (1982), “LITMUS: A New Product Planning Model,” Studies
in the Management Sciences-Marketing Planning Models, Andy Zoltners (Ed.), 18, 43-
62.
Blattberg, R. and J. Golanty (1978), “Tracker: An Early Test Market Forecasting amd Diagnostic
Model for New Product Planning,” Journal of Marketing Research, 15 (May), 192-202.
Briggs, R. and N. Hollis (1997), “Advertising on the Web: Is There Response before Click-
Through?,” Journal of Advertising Research, 37 (2), 33-45.
Chatterjee, P, D. L. Hoffman, and T. P. Novak (1998), “Modeling the Clickstream: Implications
for Web-Based Advertising Efforts,” Project 2000, Vanderbilt University.
Donthu, N., J. Cherian, and M. Bhargava (1993), “Factors Influencing Recall of Outdoor
Advertising,” Journal of Advertising Research, 33 (3), 64-72.
32
Dyson, P., A. Farr, and N. Hollis (1996), “Understanding, Measuring, and Using Brand Equity,”
Journal of Advertising Research, 36 (6), 9-21.
Fischer, P. M., J. W. Richards, E. J. Berman, D. M. Krugman (1989), “Recall and Eye Tracking
Study of Adolescents Viewing Tobacco Advertisements,” Journal of the American
Medical Association, 261 (1), 84-89.
Hamilton, A. (1998), “Secrets of Super-High Web Ad Click-Through,” ZDNet, August 19.
Hyönä, J (1995), “Do Irregular Letter Combinations Attract Readers’ Attention? Evidence From
Fixation Locations in Words,” Journal of Experimental Psychology, 21 (1), 68-81
Internet Advertising Bureau (1999), “Internet Advertising Revenues More Than Double in
1998,” New York, May 3.
Internet Advertising Bureau France/SOFRES (1999), “Net.Imp@ct,” May.
Janiszewski, C. (1990a), “The Influence of Nonattended Material on the Processing of
Advertising Claims,” Journal of Marketing Research, August, 263-278.
Janiszewski, C. (1990b), “The Influence of Print Advertisement Organization on Affect toward a
Brand Name,” Journal of Consumer Research, 17 (June), 53-65.
Janiszewski, C. (1993), “Preattentive Mere Exposure Effects,” Journal of Consumer Research,
20 (December), 376-392.
Janiszewski, C. (1998), “The Influence of Display Characteristics on Visual Exploratory Search
Behavior,” Journal of Consumer Research, 25 (3), 290-301.
Javal, E. (1878), “Essai sur la physiologie de la lecture,” Anales d’oculistique, 79, 97-117.
Kroeber-Riel, W. (1979), “Activation Research: Psychobiological Approaches in Consumer
Research,” Journal of Consumer Research, 5 (March), 240-250.
33
Lavidge, R. J. and G. A. Steiner (1961), “A Model for Predictive Measurements of Advertising
Effectiveness,” Journal of Marketing, October, 59-62.
Little, J. D. C. (1979), “Aggregate Advertising Models: The State of the Art,” Operation
Research, 27 (July-August), 629-667.
Lohse, G. L. (1997), “Consumer Eye movement Patterns on Yellow Pages Advertising,” Journal
of Advertising, 26 (1), 61-73.
Mahajan, V., E. Muller, and S. Sharma (1984), “An Empirical Comparison of Awareness
Forecasting Models of New Product Introduction,” Marketing Science, 3 (3), 179-197.
Nielsen//Netratings (1999), “The Nielsen//Netratings Reporter,” www.nielsen-netratings.com,
June 17.
Novak, T. P. and D. L. Hoffman (1997), "New Metrics for New Media: Toward the
Development of Web Measurement Standards," World Wide Web Journal, 2 (1), 213-
246.
Pechmann, Cornelia and David W. Stewart (1989), "Advertising Repetition: A Critical Review
of Wearin and Wearout," Current Issues and Research in Advertising, James H. Leigh
and Claude R. Martin Jr., Eds. Ann Arbor, MI: University of Michigan, 285-329.
Pringle, L. G., R. D. Wilson, and E. I. Brody (1982), “NEWS: A Decision Oriented Model for
New Product Analysis and Forecasting,” Marketing Science, 1 (Winter), 1-30.
Russo, J. E. and F. Leclerc (1994), “An eye-fixation analysis of choice processes for consumer
nondurables,” Journal of Consumer Research, 21 (2), 274-290.
Shapiro, S., D. J. MacInnis, and S. E. Heckler (1997), “The Effects of Incidental Ad Exposure on
the Formation of Consideration Sets,” Journal of Consumer Research, 24 (June), 94-104
34
Siddarth, S. (1999), “Describing the Dynamics of Attention to TV Commercials: A Proportional
Hazards Model of the Time to Zap an Ad,” UCLA Working paper.
Steinman, R. M. and J. Z. Levinson (1990), “The Role of Eye Movement in the Detection of
Contrast and Spatial Details,” in Eye movements and Their Role in Visual and Cognitive
Processes, ed. Elizabeth Kowler, North-Holland, Elsevier, 115-160.
Tellis, G. J. (1998), “Advertising and sales promotion strategy,” Addison-Wesley Educational
Publishers, Inc.
Young, L. R. and D. Sheena (1975), “Eye-Movement Measurement Techniques,” American
Psychologist, 30 (March), 315-330.
35
Appendix 1: Zone description variables
Building on past research (e.g., Janiszewski 1998, Steinman and Levinson 1990) we
hypothesize that a zone’s ability to attract a subject’s gaze will be a function of its location on the
page, its size and shape, and its pictorial or textual content. Location was captured by the x and
y coordinates of the center of each zone, as well as by the distance to the nearest zone seen by
the subject. The top-left corner of the screen was defined as the origin point (0,0), the bottomright
corner was set to (255,255). We also created squared terms for the x and y coordinates to
allow for the fact that the center of the pages might be preferred.
Shape was described in terms of the zone’s area as well as its orientation4 (horizontal vs.
vertical) and the general orientation of the page. An interaction dummy was also created to
indicate whether the zone and the page follow the same orientation.
The content of each zone was coded using the following dummy variables:
- TEXT (set to 1 if the zone contains text, 0 otherwise),
- IMAGE (1 if the zone contains non-photographic graphical elements),
- PICTURE (1 for photographs),
- INPUT (1 if the zone contains input elements such as buttons),
- CONTRAST (1 if the zone contrasts with its surrounding), and
- AD (1 if the zone contains an advertisement).
Our last control variable is a variable that indicates the number of zones present on each
page.
4 Remember that all zones are rectangular.
36
Table 1: Logit regression (Zone Attractiveness)
Variable Coefficient Wald c2 Pr>c2
Intercept -3.1947 30.22 0.0001
AD -0.1888 5.10 0.0240
Expertise -0.0633 2.50 0.1136
Test
AD x Expertise -0.0554 0.54 0.4639
X 0.0203 8.24 0.0041
Y 0.0211 32.73 0.0001
X2 -0.0001 14.52 0.0001
Y2 -0.0001 59.24 0.0001
Location
Distance -0.0380 326.27 0.0001
Area 0.0005 348.96 0.0001
Zone Orientation 0.0482 0.37 0.5437
Page Orientation 0.0597 1.00 0.3190
Shape
Orientation Int. 0.1135 3.68 0.0550
IMAGE 1.2067 91.52 0.0001
PICTURE -3.9477 236.82 0.0001
CONTRAST 0.2813 12.06 0.0005
TEXT 1.8351 150.20 0.0001
Content
INPUT 0.1751 2.23 0.1353
# of Zones/Page 0.0083 0.15 0.7029
Age 0.0595 2.89 0.0892
Control
Gender 0.0531 1.60 0.2013
Regression 1234.03 0.0001
Table 2: Differences among groups of surfers
Number of Fixes Number of Zones Fix Time
Estimate p-value Estimate p-value Estimate p-value
Intercept 26.59 0.0001 6.99 0.0001 12.01 0.0001
Expert -6.03 0.0006 -0.67 0.0162 -2.52 0.0002
Old 5.36 0.0018 0.21 0.4536 2.65 0.0001
Female -2.32 0.257 -0.59 0.0753 -0.86 0.2822
Voilà -0.91 0.7742 0.28 0.5851 -0.23 0.8507
Voilà Answer 4.73 0.1363 1.59 0.0019 2.22 0.0742
Yahoo! 4.62 0.1413 -0.49 0.3309 0.63 0.6078
Yahoo! Answer 12.18 0.0001 -2.27 0.0001 4.39 0.0004
Voilà Bis 4.13 0.1912 0.43 0.3872 1.65 0.1806
Voilà Bis Answer 10.2 0.0014 2.26 0.0001 4.2 0.0008
Voilà Annuaire 6.48 0.0436 2.78 0.0001 0.95 0.4481
N 378 378 378
R2 0.11 0.28 0.13
37
Table 3: Unaided Advertising Recall (24 hours after surfing)
Brand Brand Recall*
Cortal 14.3%
@près l’école 12.6%
Le Mél 12.6%
Tout En Ville 10.2%
Nouba 6.9%
Voilà 4.3%
Netscape 3.3%
Honda 3.0%
France Telecom 2.8%
Internet Explorer 2.8%
123 Achats 2.2%
Pages zoom 2.0%
Linux 1.7%
Pages Jaunes 1.5%
Pages blanches 1.1%
Other 4.6%
Cannot Recall Brand 54.1%
*Numbers do not add to 100% due to multiple answers
Table 4: Logit Regression (Unaided Advertising Recall)
Variable Coefficient Wald c2 Pr>c2
INTERCEPT -3.8413 837.8418 0.0001
EXP1 1.6360 104.0824 0.0001
EXP2 0.3460 6.8699 0.0088
BUTTON 0.2496 1.3723 0.2414
Regression 245.96 0.0001
38
Table 5: Aided Recall
Banner
Aided Ad
Recall
Brand Recognition
(Conditional)
Brand Recognition
(Unconditional*)
Tout en Ville 35.4% 64.1% 22.7%
Le Mél 36.3% 66.7% 24.2%
Nouba 25.3% 40.8% 10.3%
@près l’école 31.0% 72.1% 22.4%
Cortal 23.6% 60.6% 14.3%
Overall** 30.1% 61.6% 18.5%
*Unconditional Brand Recognition = Ad Recall * Brand Recognition
**Overall numbers are weighted averages of individual ad’s numbers
Table 6: Brand Recognition
Le Mél Tout-en-Ville Nouba Cortal @près l'école
Le Mél 180 1 0 0 1
Tout-en-Ville 1 134 8 1 1
Nouba 0 4 78 0 1
Cortal 0 0 0 106 1
@près l'école 0 0 1 0 168
Voilà 6 2 15 0 1
Similar brands 4 3 2 2 0
Wanado 1 0 0 0 7
France Telecom 4 2 0 0 1
Honda 0 0 1 0 1
Other 3 1 3 3 2
Don't Remember 71 62 83 63 49
Total 270 209 191 175 233
39
Table 7: Aided Brand Awareness
Brand Pre Post D D%
# of
Exposures
Alta Vista 95.8% 95.2% -0.6% -0.6%
Nomade 85.0% 83.9% -1.1% -1.3%
Lycos 81.5% 78.9% -2.6% -3.2%
Yahoo! 98.5% 96.7% -1.8% -1.8%
0
Le Mél 47.6% 50.4% +2.8% +5.8%
Tout-en-Ville 22.3% 25.5% +3.2% +14.3%
Nouba 11.3% 14.4% +3.1% +27.4%
1
Cortal 17.2% 18.6% +1.4% +8.1%
@près l'école 21.4% 25.0% +3.6% +16.8%
2
Voilà 78.8% 90.5% +11.7% +14.8% 9
Table 8: Logit Regression (Aided Brand Awareness)
Variable
Parameter
Estimate Wald c2 Pr >c2
Intercept -1.2862 413.70 0.0001
Le Mél 1.2145 243.93 0.0001
Alta Vista 4.3612 1045.96 0.0001
Nomade 3.0034 1054.84 0.0001
Voilà 2.5864 655.96 0.0001
Tout en Ville 0.0966 1.34 0.2472
Lycos 2.7124 950.62 0.0001
Cortal -0.3280 13.98 0.0002
Nouba -0.6627 48.19 0.0001
Yahoo! 5.0099 833.75 0.0001
TEST -0.0499 0.90 0.3435
EXP 0.1139 38.32 0.0001
Regression 7243.60 0.0001
40
Table 9: Logit Regression (Aided advertising recall and brand recognition)
Aided Advertising Recall Brand Recognition
Variable
Parameter
Estimate Pr >c2 Parameter
Estimate Pr >c2
Intercept -0.7503 0.0001 0.3436 0.0001
Dummy -0.2293 0.0021 0.3716 0.0047
N 3597 1078
Table 10: Size and Orientation Significance Level
Construct H5a: Size H5b: Orientation
Pr >c2 N Pr >c2 N
Aided Advertising Recall 0.9483 3926 0.0003 918
Brand Recall 0.6397 1018 0.1716 240
Unaided Advertising Recall 0.3056 2360 0.9314 552
Brand Awareness 0.8785 4146 0.1714 962
Overall No Effect Weak Effect
Table 11: Contrast, Execution, and Animation Significance Level
Construct H6a: Contrast H6b: Animation H6c: Message
Pr >c2 N Pr >c2 N Pr >c2 N
Aided Advertising Recall 0.0012 1500 0.0989 900 0.0001 1182
Brand Recall 0.5407 540 0.6084 216 0.0022 418
Unaided Advertising Recall 0.1194 902 0.2726 540 0.9297 716
Brand Awareness 0.2719 1582 0.6447 944 0.6114 1266
Overall Weak Effect No Effect Effect
Table 12: Overall Banner Exposure Effects
Advertising Construct Effect
Exposure 100.0%
Aided Advertising Recall 30.1%
Aided Brand Recognition 18.5%
Unaided Brand Recall 11.4%
Increased Brand Awareness 2.8%
Click-through* 0.6%
* From Nielsen//Netratings 1999
41
Figure 1: Voilà Answer page
42
Figure 2: Voilà Answer zone definition
43
Figure 3: Frequency distribution of the number of banners looked at
0
2
4
6
8
10
12
14
1 2 3 4 5 6 7 8
Number of Banner Seen
Number of Respondents
44
Figure 4: Experts vs. Novices
45
Figure 5: Logit Response Function
Figure 6: Logit Response Function
Post awareness for different exposure level
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.01 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.99
Pre Awareness
Post Awareness
0
1
2
9
Post-Awareness for different pre-awareness
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 2 4 6 8 10 12 14 16 18 20
Exposures
Post Awareness
0.25
0.5
0.75
Pre-Awareness