Internet Meda Plannng: An Optmzaton Model Janghyuk Lee 1 and Laoucne Kerbache 2 HEC School of Management, Pars 1 rue de la Lbératon 78351 Jouy-en-Josas, France 1 Assstant Professor of Marketng emal: lee@hec.fr phone: +33-1 3967 9440 2 Assocate Professor of Logstc and Producton emal: kerbache@hec.fr phone: +33-1 3967 7212 Acknowledgements: We are gratefully ndebted to Vjaya Chebolu for her research assstance and partcularly apprecate fnancal support from "La Fondaton HEC" and data from KoreanClck.
ABSTRACT Of the varous meda vehcles avalable for advertsng, the Internet s the latest and the most rapdly growng, emergng as the deal medum to promote products and servces n the global market. In ths artcle, the authors propose an Internet meda plannng model whose man objectve s to help advertsers determne the return they obtan from spendng on Internet advertsng. Usng avalable data such as Internet page vew and advertsng performance data, the model contrbutes to attempts not only to optmze the Internet advertsng schedule but also to fx the rght prce for Internet advertsements on the bass of the characterstcs of the exposure dstrbuton of stes. The authors test the model wth data provded by KoreanClck, a Korean market research company that specalzes n Internet audence measurement. The optmal duratons for the subject stes provde some useful nsghts. The fndngs contrast wth current Web meda plannng practces, and the authors demonstrate the potental savngs that could be acheved f ther approach were appled. Key words: meda plannng; optmzaton, advertsng repeat exposure, probablty dstrbuton.
1. Introducton For busnesses that sell goods and servces, advertsng often represents the frst means to make the publc aware of them. Among the varous meda vehcles avalable for advertsng, the Internet s the latest and most rapdly growng (Bell and Tang 1998). Already a major communcaton channel n many developed countres, the Internet has attracted the attenton of marketng managers not only because of ts rapd adopton but also because t s an nteractve communcaton medum that provdes a wde varety of sze, locaton, and technology optons (Novak and Hoffman 1997). Especally the rapd evoluton of data transmsson speed on the Internet makes consumers spend more tme on the Internet compared to other tradtonal mass meda. Accordng to a recent report from Jupter Research (2004), the broadband Internet s challengng TV n Europe as 40% of consumers havng broadband access at home sad that they were spendng less tme watchng TV. Therefore, companes ncreasngly have started to rely on Internet advertsng to acqure new customers and mprove ther brand mage. Accordng to an Internet advertsng revenue report from the Interactve Advertsng Bureau (2004), U.S. companes ncreased ther spendng on Internet advertsng from $907 mllon n 1997 to $7,267 mllon n 2003, wth a 21% growth rate between 2002 and 2003. Although ths percentage ncrease ncludes 3% of total meda spendng, the study fnds a hgh concentraton of advertsng spendng on major Web stes. In the fourth quarter of 2003, the top 10 Web stes accounted for 71% of total advertsng spendng, and the top 50 Web stes encompassed 96%. Internet advertsng offers more accurate measurement and more flexble plannng than do tradtonal meda (Drèze and Zufryden 2000). For example, each ste can measure systematcally the sze of ts audence and the frequency of exposure. Ths mproved accuracy enhances the transparency of a return on nvestment (ROI) analyss because the drect mpact of an Internet advertsement on sales can be assessed and even lnked to a commercal Web ste. The Internet also allows for content modfcatons and schedule flexblty. If, for example, an Internet advertsng campagn was unsuccessful n ts early stages,, ts content and schedule could be modfed for the rest of the campagn. For ths overall approach, advertsers need a decson-makng tool such as meda plannng. 1
Meda plannng determnes the subject, tmng, and locaton of advertsements. Decsons regardng the establshment of a meda plan nvolve understandng and then ntegratng marketng objectves, the dynamcs of the market, target audence, and the avalable meda vehcles wth ther assocated costs and characterstcs. Because these data often are ncomplete and uncertan, meda plannng problems tend to be probablstc n nature. However, n contrast to tradtonal meda publcty vehcles, such as magaznes, newspapers, and televson, the Internet provdes much more data that enable advertsers to measure exactly the exposure of consumers to the dsplayed advertsements. In turn, Internet data make t possble to understand the mpact of publcty on the consumer and ncrease ts effectveness. To measure the mpact of advertsng, researchers need nformaton about the number of consumer exposures and consumers attenton levels to a partcular advertsement. The past decade has wtnessed the wdespread adopton of meda models for estmatng the reach and frequency dstrbuton of exposures of tradtonal meda. Thus, several methods, both publc and propretary, are now avalable to meda planners to use to estmate the proporton of the target audence that wll be exposed once, twce, and up to n tmes by a combnaton of n nsertons n m meda vehcles (Lttle and Lodsh 1969, Aaker 1975, Rust 1985). Consderng the nnovaton and technologcal benefts that the Internet has over tradtonal meda, there s an undenable need to adapt prevous models of meda plannng effectvely to the Internet envronment, n whch a vstor can be exposed to an advertsement many tmes durng a fxed perod. In ths artcle, we propose an Internet meda plannng model to deal wth ths ssue, through whch we focus specfcally on Web page vew statstcs and the repeat exposure effect. The man objectve of ths model s to help advertsers assess the ROI of spendng on Internet advertsng through the use of avalable market data such as Internet page vews and advertsng performance data. We organze ths artcle as follows: In secton 2, we present the theoretcal background related to exposure dstrbuton and the repeat exposure effect. Secton 3 s devoted to developng our Internet meda plannng model. In secton 4, we present a real case study and analyze the results by explorng varous scenaros that yeld nterestng manageral nsghts and demonstrate the robustness of our 2
Internet meda plannng model. Fnally, we provde some conclusons and suggestons for further research n secton 5. 2. Theoretcal Background The man concern of advertsers s how many people they can reach and how often. Advertsng agences attempt to acheve optmal plannng about the number of placements and the choce of meda to maxmze the reach/frequency of a campagn wth a gven budget f other thngs (e.g., attractveness, creatvty) are equal. In the followng, we address the major conceptual ssues central to our research: exposure dstrbuton and the repeat exposure effect. 2.1. Exposure Dstrbuton In tradtonal meda such as magaznes and televson, the queston of how to capture the exposure dstrbuton of people across varous meda was explored frst by Metherngham (1964), who used a bnomal dstrbuton to capture the reach and frequency of exposure for a sngle vehcle (.e., magazne) wth a fxed number of nsertons and then developed the beta bnomal dstrbuton (BBD) to ntegrate varyng probabltes of exposure that represent heterogenety across consumers. A more flexble model that treats more than one vehcle smultaneously s based on Drchlet multnomal dstrbuton (DMD) (Leckenby and Ksh 1984, Danaher 1988a). Rust and Leone (1984) extend the DMD wth a hypergeometrc adjustment that accommodates the case of an unequal number of nsertons n all vehcles. Smlarly, Danaher (1988b) develops a log-lnear model to handle three or more vehcles at a tme, whch outperforms BBD and DMD n error measurements. Despte ts enhanced performance, ts computatonal burden remans a problem, whch Danaher (1989) eased by usng an approxmaton of the mnmum deteroraton of performance. However, on the Web, the nature of advertsng exposure changes drastcally because of consumers unlmted access to advertsng, n contrast wth ther passve, lmted exposure n the case of predetermned advertsng schedules such as those used on televson. Therefore, t s no longer vald to apply dstrbutons such as BBD and DMD that fx the total number of nsertons; each nserton on a Web page can generate unlmted exposures. Because Web advertsements are posted 3
on a specfc page for a fxed duraton, the focus of exposure dstrbuton shfts from estmatng the number of exposures accordng to the number of nsertons to determnng the number of exposures an nserton generates durng a fxed perod of tme. The number of exposures to a Web page durng a fxed perod s a stochastc process that follows a Posson dstrbuton for average exposures. Because each consumer may have a dfferent level of average exposure, exposure frequences can be ft by the negatve bnomal dstrbuton (NBD) (Ehrenberg 1959), a mxture of the Posson and gamma dstrbutons. The Posson dstrbuton represents the exposure rate for a fxed duraton, whereas the gamma dstrbuton captures the heterogenety of the exposure rate. Because the number of future purchases a pror depends on whether a customer s actve, Schmttlen, Morrson, and Colombo (1987) refne the NBD by ntroducng the probablty of beng actve as represented by the Pareto dstrbuton that mxes the exponental and gamma dstrbutons. The death rate of the customer s captured by the exponental dstrbuton, whose heterogenety s embedded n the gamma dstrbuton. In ths research, we use the NBD because we assume that all users stay alve for the relatvely short (e.g., four weeks) duraton of an advertsng campagn. Ths assumpton sgnfcantly reduces the computatonal complexty of the meda plannng optmzaton model. 2.2. Repeat Exposure Effect of Advertsng Advertsements can nfluence the consumer's three-stage (generaton, consderaton, selecton) brand choce process. They also can alter the content of brand nformaton on two dmensons accessblty and value stored n memory, ncludng the brand name, the value and valence of brandattrbute belefs, and the valence of brand atttudes (Nedungad, Mtchell, and Berger 1993). By ncreasng the accessblty of product-attrbute belefs and brand atttudes (Berger and Mtchell 1989) and actvatng brand nformaton, advertsng repetton can enhance the performance of advertsements. The two-factor theory proposed by Berlyne (1970) suggests that the mpact of exposure frequency s medated by two factors: habtuaton (learnng) and tedum. Habtuaton can mprove an advertsement s performance, whereas tedum deterorates t. If the tedum factor overwhelms ts 4
counterpart after the number of exposures passes a threshold, repeat exposures may take the form of nverted-u curves, n whch two opposng psychologcal processes operate smultaneously: postve habtuaton and negatve tedum. Smlar explanatons of an nverted-u curve functon for repeated exposure have been proposed for atttudes (Cacoppo and Petty 1979) and learnng (Pechmann and Stewart 1989). Pechmann and Stewart (1989) use the terms wearn and wearout n ther elucdaton of the nverted-u curve response to advertsng and suggest that wearn occurs durng approxmately the frst three exposures, after whch postve thoughts outnumber negatve thoughts. The wearout stage begns wth approxmately the fourth exposure, when message recpents start to become bored and consequently generate negatve repetton-related thoughts, whch undermne the persuasve mpact of the advertsement. Krugman (1972) provdes a dfferent perspectve on the effects of frequency. He proposes a threeht theory, whch posts that an advertsement reaches ts maxmum effectveness at the thrd exposure. The frst exposure elcts a cognate response to the nature of the stmulus. The second exposure s more evaluatve and personal and rases questons about the meanng of the ad. But the thrd exposure represents the true remnder because the vewer has already gone through hs or her cogntve process. Krugman further argues that addtonal exposures smply repeat the thrd-exposure effect wthout ncremental mprovements. Thus, the three-ht theory could be graphcally depcted as an S-shaped or concave response curve wth a plateau after the thrd exposure. Some prevous research also supports the clam that atttudes, purchase ntentons, and postve cogntve responses peak at the thrd exposure n the case of televson advertsng (Cacoppo and Petty 1979, Calder and Sternthal 1980, Belch 1982). For the Internet, Drèze and Hussherr (2003) fnd postve repeat exposure effects for three major measures: aded brand awareness, unaded advertsng recall, and brand awareness. They test the repeat exposure effects for a sample of 807 respondents who were surveyed both before and after ther exposure to Internet advertsng of 10 brands. The number of exposures ranged 0 9 tmes durng 24 hours. They detect a statstcally nsgnfcant forgettng phenomenon. 5
2.3. Prevous Approaches to Meda Plannng In the 1960s, lnear programmng emerged as an approprate modelng and optmzaton tool for allocatng advertsng to varous meda. Mller and Starr (1960) and Day (1962) establshed the crtera to apply lnear programmng prncples for selectng meda through questons about when (tme) and where (space) advertsements should appear accordng to the budgetary constrants. Addtonal constrants that guarantee a mnmum spendng level of a meda class or an ndvdual medum can be taken nto account, as can the mnmum exposure of specfc market segments. These models rely on the ratonale that advertsng creates an advertsng exposure that n turn creates sales; that s, the purchase ntenton of consumers can be elevated by ther enhanced awareness and postve atttude toward the brand. Along these lnes, Lee (1962) developed a lnear programmng model to optmze the number of advertsements of a gven tme length to ensure the requred level of awareness. Starsch (1965) extended the model to select markets n whch the advertsement should appear and ntegrated frequent dspartes between the sales potental, whch s specfc to each market, and the crculaton of canddate meda that serve those markets. Brown and Warshaw (1965) ntroduced the noton of nonlnear response to advertsng, n whch the response prompts dmnshng, S-shaped returns n an exponental form wth a saturaton level (Vdale and Wolfe 1957). Accordng to ths theory, the number of advertsements used per perod can be modeled as a decson varable, fractoned nto regons that have dfferent response levels. To refne and extend prevous models, Lttle and Lodsh (1969) developed MEDIAC, n whch they nclude the market segments, segment-specfc sales potentals, and exposure probabltes of each meda opton. The decson varables are boolean varables that ndcate the nserton of a gven advertsement n a specfc medum at a specfc tme. On the bass of the nserton varable, the authors can asses the total market response as a functon of the level of current and prevous exposures of consumers and ther sales amount, weghted by segment. Subsequently, Aaker (1975) proposed a meda plannng model wth a dfferent approach. Hs ADMOD model focuses not on the aggregate vehcle audence but rather on sample populatons selected from the varous segments and thereby examnes the lkely mpact measured as a change n cognton or purchase ntenton of a partcular 6
nserton on each consumer n the sample. The change n cognton s assessed as a functon of the number of exposures, dependng on the meda schedule, whch then s extrapolated to the real populaton to provde the total expected results (e.g., proft generated by the meda schedule). A bnomal dstrbuton captures the dstrbuton of exposures because the ADMOD model ncludes a lmted number of ad nsertons. Fnally, Rust (1985) suggested a televson plannng model (VIDEAC) that provdes standard data such as program avalablty, cost, and ratng (by segment) drectly to advertsng agences. In VIDEAC, the exposure dstrbuton s estmated by the BBD to capture the heterogenety of the populaton exposure rate. 3. Model Development Compared wth those of prevous meda plannng models, whch were developed manly for televson, rado, and magazne advertsng and have a dscrete format wth a lmted number of nsertons, the decson varables of our model represent the duraton of advertsng (weeks) on selected Web stes. The goal of our model s to assess the repeat exposure effect of advertsng on the Web, where most stes attract vstors repeatedly. For the sake of modelng smplcty, we start by consderng contnuous decson varables and do not lmt the number of nsertons a pror. However, we examne other scenaros subsequently. In our model, the objectve functon drectly measures the number of ndvduals (.e., consumers recallng the ad message) who can be nfluenced by advertsng. It conssts of two parts: the repetton functon and the exposure frequency dstrbuton. In addton, t assesses performance at the level of exposure frequency. For ad message recall, the objectve functon sums the total number of subjects who recall the ad message, obtaned from the probablty of message recall after beng exposed k tmes and the number of vstors exposed k tmes. Thus, the functon maxmzes the number of nfluenced consumers by choosng the duraton of stes accordng to the exposure dsparty n the number of unque vstors and ther repeat exposure dstrbuton across selected stes. The repetton functon can be obtaned from pretest results that calbrate the probablty of ad performance n terms of exposure frequency. In prevous meda plannng models, the repetton functon seems to show dmnshng returns at hgh exposure levels. Lttle and Lodsh (1969) adopt a 7
functon n an exponental form that descrbes the fracton of sales potental realzed as a functon of exposure level, n whch there s a mnmum fracton for no ad exposure and an upper bound of 1 for achevng the full potental of sales. In contrast, ADMOD (Aaker 1975) uses a repetton functon wth lower and upper bounds that are lnked through a power functon of exposure frequency. The VIDEAC model (Rust 1985) adopts a smple form of the square root of the number of exposures for the repetton functon. In our model, we use the repetton functon to represent the probablty of ad performance (message recall rate) and assume t to be a logt functon of the exposure frequency, p( X 1 = k) =. Its lower bound depends on the value of the constant a for the case of no 1+ exp[ ( a + bk)] exposure, and t ncreases monotoncally. The shape of ths repetton functon depends manly on the coeffcent of exposure frequency b. The greater the coeffcent b, the greater s the performance dfference n the low range of exposures because the slope of the repetton functon gets steeper. As the number of exposures tends toward nfnty, the probablty approaches 1. For exposure frequency probablty, we frst descrbe t by a Posson dstrbuton wth the mean exposure rate of λ, f ( X k λ λ e = k λ) =. To estmate the exposure dstrbuton flexbly over a varyng k! tme duraton, the exposure frequency probablty changes to k ( λt ) ( λt) e f ( X ( t) = k λ) =, whch k! ncludes the duraton varable t to represent extended or shrunken duraton. Because t s realstc to ncorporate the heterogenety of the exposure rate λ, whch vares across ndvduals followng a dstrbuton of g(λ), f ( X ( t) = k λ = f ( X ( t) = k λ) g( λ) dλ, whch we ncorporate by usng the gamma 0 dstrbuton. The exposure frequency dstrbuton can be estmated by a mxture dstrbuton of Posson and gamma that leads to a NBD wth two parameters, γ as the shape parameter and α as the scale parameter, n addton to the duraton varable t, so that k t t e e Γ + k t f X t = k = ( λ ) γ γ 1 αλ ( λ ) α λ ( γ ) α ( ( ) ) dλ =. The number of consumers exposed k 0 k! Γ( γ ) Γ( γ ) k! α + t α + t γ k tmes can be obtaned by multplyng the total populaton (M) by the exposure frequency dstrbuton. 8
Thus, our objectve functon to assess the number of ndvduals nfluenced by the advertsng becomes a product of the probablty of ad performance and the number of ndvduals exposed k K tmes, p( X = k) * M * f ( X ( t) = k), for a sngle ste at whch the decson varable s the ad duraton t. k = 1 In case of N dfferent stes, the objectve functon becomes N K = 1 k = 1 p( X = k) * M * f ( X ( t ) = k), and we fnd a set of the duraton for selected stes (t) that maxmzes the objectve functon. However, the repetton functon and frequency exposure dstrbuton may be specfc to each Web ste that has ts own parameters. Along wth the precedng objectve functon, our Internet meda plannng model ncludes the followng set of constrants: The amount of budget allocated to an ad campagn. On the Internet, a frequently appled method to fx ad rates s based on the total frequency of exposures. Among ad practtoners, ths rate s called the cost per thousand exposures (CPM). 1 In our model, ths rate s noted r because t may be specfc to each ste. Therefore, the expresson of the ad campagn cost, obtaned from the cost rate and the total number of exposures of the lsted stes, N r 1000 K = 1 k = 1 M * k * f ( X ( t ) = k) A, should be smaller than the campagn budget (A). The maxmum duraton of selected ste(s). Even for cost-attractve stes, an advertser cannot run an ad campagn for more than a certan perod of tme. Ths duraton constrant often s related to the tmng of the ad campagn. In our model, we easly ncorporate t as t T. Our model can summarzed as follows: The objectve functon, N K p ( X = k) * M * = 1 k = 1 Γ( γ ) k! Γ( γ + k) α α + t γ t α + t k, s subject to the budget constrant, 1 The Interactve Advertsng Bureau (2004) reports that 47% of ad revenues were generated on the bass of CPM or mpressons (ncludng sponsorshp) n 2003 for the U.S. market. 9
N K r Γ( γ + k) α t M * k * k k t t 1000 = Γ( γ ) α + α + = 1 1! γ k A, and the tme duraton constrant, t T, where the decson varable s t = advertsng duraton of ste, and the other parameters are as follows: p (X = k) = ad performance of ste at k exposures, γ = shape parameter of the NBD capturng the exposure frequency of ste, α = scale parameter of the NBD capturng the exposure frequency of ste, r = advertsng fee rate (CPM), and A = total advertsng budget amount. Our s clearly a nonlnear programmng optmzaton model wth contnuous varables and a complex objectve functon. The objectve functon s nonlnear, and therefore, the search for optmal or nondomnated solutons wll be computatonally tme consumng, especally for large problems. Because of the partcular characterstcs of the Internet, advertsers cannot use prevous meda plannng models to choose the optmal combnaton of slots from a predetermned schedule. Our model nstead enables advertsers to plan an ad campagn n a contnuous manner to reduce or ncrease the duraton of advertsng on the lsted stes to maxmze ther objectves, as represented by the objectve functon. Also, our model fully assesses the margnal contrbuton of ad duraton for each ste and captures the ad effect across the range of exposure frequency. Fnally, ths model can be vewed as a platform that may be modfed to ncorporate other key ssues of meda plannng, such as segmentaton and the nteracton effect. To ncorporate segments, we can splt the exposure frequency dstrbuton nto S segments. Each segment has ts own parameters, γ and α, of the NBD dstrbuton, whch enables us to assess the exposure frequency wth dfferent patterns (shape and scale), f γ s k Γ( γ + k) α t s s s ( X ( t ) = k) =. Ths process s exactly the same as that used to Γ( γ ) k! s t α + s t α + s 10
estmate the exposure frequency of stes, except that we splt t by segment and add a weght (w s ) to N represent the sze of the segment, p( X = k) * M * w * f ( X ( t ) = k). S K = 1 s= 1 k = 1 The ncorporaton of the nteracton effect s another feature of our model. For tradtonal meda, the nteracton effect of the ad copy and the vehcle (e.g., magazne, televson or rado program) s of nterest (Ray and Sawyer 1971, Ray and Strong 1971). In our model, we ncorporate ths effect by provdng a specfc repetton functon that depends on the ste or even on the segment, s s p ( X s 1 = k) =. 1+ exp[ ( a + b k)] s s 4. MODEL APPLICATION In the precedng secton, we developed a meda plannng model adapted to handle Internetspecfc characterstcs, such as hgh repeat exposure and decson makng over a contnuous duraton. By usng real Internet page vew statstcs and a repetton functon that shows the ad performance by exposure frequency, we obtan useful nsghts wth regard to enhancng the effcency of Internet advertsng. Our fndngs contrast wth current Internet advertsng practces, presented prevously, whch reflect hghly concentrated ad plannng devoted to a lmted number of popular portals or search engnes that provde a wde reach and hgh repeat exposures. 4.1. Data Descrpton Our Internet data are provded by KoreanClck (www.koreanclck.com), a Korean market research company that specalzes n Internet audence measurement. KoreanClck mantans a panel of Internet users, selected on the bass of stratfed proportons n South Korea, between 10 and 65 years of age. Canddates for the panel are contacted by a random dgt dalng method. After the person agrees to be a panel member, he or she receves authorzaton from KoreanClck by both e-mal and regular mal to regster as a panel member. The panel member s counted as an effectve member f he or she connected to the Internet at least once durng the four precedng weeks. The Internet usage behavor of the panel member s measured by a module, called Track, that captures the use of the actve Internet browsers by the panel member at hs or her home or offce. 11
There are several major performance ndcators of Internet usage, ncludng page vew, vstor, unque vstor, and reach (Novak and Hoffman 1997). A page vew s the act of browsng a specfc Web ste. When a vstor accesses a Web page, a request s sent to the server hostng the page; a page vew occurs when the page s fully loaded. At ths pont, an mpresson takes place because the consumer s exposed to the page contents, ncludng advertsements. The page vew measurement s equvalent to exposure n the case of tradtonal meda such as televson, rado, and magaznes. Used manly to determne advertsng fees, page vew llumnates the volume of browsng on the Web ste. The vstor ndcator reflects a person who recorded at least one page vew of a specfc ste usng the Internet browser, whereas the unque vstor ndcator relates to the net count of vstors when multple vsts by the same person are elmnated. Fnally, reach s the number of unque vstors among the total Internet users durng a fxed perod. The reach explans the capacty of the Web ste n terms of how wdely t covers the total Internet populaton. KoreanClck provdes relable Internet data that mnmze the measurement problems rased by Dréze and Zurfryden (1998). The Track module dentfes the vstor at all ponts of the Internet as long as t s nstalled; however, t can mss some vst data f the panel member accesses the Internet from a publc place, such as schools or Internet cafés. Ths loss of nformaton probably s margnal compared wth the panel member s major Internet usage at ether the workplace or home. Furthermore, Track captures vst data that are cached by the proxy server of the panel member. If Web pages have a frame, Track counts page vews of only the destnaton page and does not double count t as page vew of the framed page. We gathered data for four weeks, March 3 30, 2003. The effectve panel members number 4149 for week 1, 4195 for week 2, 4125 for week 3, and 4148 for week 4. We retan a total of 3492 panel members who were effectve durng the entre four-week perod. Of these effectve panel members, 36% are women, and they average 32 years of age. To measure ther Internet usage behavor, we use page vews of the ndex pages (smlar to the cover page of magazne) of ten selected Web stes n three major categores: communty portal, news, and search engne (see Table 1). <Insert Table 1: Ste Profle around here> 12
We selected these ten stes for ther popularty among all types of Internet users and because they have a relatvely wde reach (f a Web ste has a small reach, ts page vew data become more volatle). All stes experenced smlar gender proportons among ther vstors except news stes, whch receve vsts from more men. Portal stes and search engnes are hghly frequented; for example, the communty portal ste 1 reached more than 80% of the total Internet users durng the week of March 3. To measure the average page vews, we dvded all page vews by the total Internet users, whch represents a measure smlar to the gross ratng ponts frequently used by tradtonal meda. The average page vews ndcate the overall exposure rate of the gven ste to all Internet users. However, the most mportant ndcator s the average page vews per vstor, whch captures the exposure frequency among members who actually vsted the ste. Ths ndcator tends to ncrease when the ste reaches a wder range of the Internet populaton. For example, stes wth very wde reach (e.g., stes 1, 8, and 9) record more than 20 page vews per vstor. In addton to the exposure frequency dstrbuton, we use a repetton functon of Internet advertsng to complete our meda plannng model. Lee and Brley (2004) report on a repetton functon of Internet advertsng that measures the repeat exposure effects of the ad recall rate for a hgh exposure frequency usng 10 onlne ad performance surveys of 10,667 observatons. They fnd a statstcally sgnfcant message recall functon n terms of the exposure frequency and the probablty of ad message recall after k exposures, 1 p ( X = k) =. 1+ exp[ ( 1.141+ 0.187 ln( k + 1))] Ths monotoncally ncreasng functon has a lower bound of 21.41% and a upper bound of 100%. Unlke ADMOD (Aaker 1975), our model does not need an upper bound of less than 100% because the logt form repetton functon provdes a plateau wthn the range of plausble exposures. For example, the message recall rate of ths functon s expected to be 53.77% after 1000 exposures. As we llustrate n Fgure 1, the probablty of message recall ncreases more rapdly n the low exposure frequency area (.e., fewer than 10 exposures) than n the hgh exposure frequency area. <Insert Fgure 1: Repeat Exposure Effect of Message Recall Rate around here> Because the performance of an Internet advertsement does not ncrease lnearly wth the ncrease of the exposure frequency (smlar to other meda), the ncreased page vews per vstor should alert 13
the advertser of ts possble wasteful spendng, accordng to the low advertsng spendng effectveness among consumers who experence hgh repeat exposures. 4.2. Exposure Frequency Dstrbuton We apply the NBD to capture the exposure frequency dstrbuton. As we mentoned n the model development secton, the NBD s a mxture of Posson and gamma dstrbutons. Whereas the Posson dstrbuton estmates the dstrbuton of events (ad exposures) over a fxed duraton wth one parameter λ to represent the mean of events, and thereby captures the dstrbuton of ad exposures n a dscrete manner, the gamma dstrbuton ntroduces the heterogenety of consumers average exposure rate λ. When the Posson s mxed wth the gamma, t becomes an NBD wth two parameters: γ as the shape parameter and α as the scale parameter. The mean exposure rate therefore s computed as γ/α. The NBD provdes two major advantages because of ts flexble nature. The NBD Ft of One-Week Data The frst flexblty of the NBD s ts reasonable ft of Internet page vew data, even though the exposure rate heterogenety s embedded. On the bass of the maxmum lkelhood prncple, we obtan parameter estmates of our ten lsted stes, as we show n Table 2. We use the commercally avalable software MATLAB for the maxmum lkelhood estmaton and compare t to parameter estmates obtaned wth MS Excel Solver. Both software programs provde smlar values for the two parameter estmates, so for our remanng analyss, we use the estmates obtaned from MATLAB. < Insert Table 2: NBD Parameter Estmates (MS Excel Solver and MATLAB) around here> To check the goodness of the NBD ft for our sample of N = 3492, we proceed wth the Kolmogorov-Smrnov (K-S) test (Massey 1951) nstead of the Pearson ch-square test, whch s napproprate for a sample of large observatons because ts value s too senstve to the number of data ponts. The K-S test, a nonparametrc test, compares the goodness of ft of a sample dstrbuton S N (x) wth that of a populaton by measurng the absolute dstance between the two dstrbutons. The maxmum absolute dstance between the sample and the populaton cumulatve dstrbutons s d = maxmum F 0 (x) S N (x). If the dstance s smaller than the crtcal value at a sgnfcant level of α%, 14
the sample provdes an approprate ft to the populaton at that sgnfcance level. The dstance should get smaller as the sze of the sample N ncreases. In our case, we suppose that the exposure dstrbuton of the populaton of Internet users, F 0 (x), follows an NBD. <Insert Table 3: Kolmorov-Smrnov Test around here> The crtcal value to test the goodness of ft s gven by 1.22/ N for α = 10%, 1.36/ N for α = 5%, and 1.63/ N for α = 1%. The crtcal values are 2.06%, 2.30%, and 2.76%, respectvely. As we show n Table 3, the goodness of NBD ft s acceptable at α = 10% except for stes 1 and 9. Usng the conservatve standard of α = 1%, all 10 stes have an approprate ft of the NBD. The K-S dstance tends to correlate wth the varance of both the exposure dstrbuton of stes and the reach. For stes wth a narrow reach, the NBD can mnmze the K-S dstance f t effectvely captures those nonvstors that represent the greatest dstrbuton densty. However, stes wth a wde reach have more dspersed dstrbuton and greater varaton among vstors. For these, the NBD must ft not only nonvstors but also vstors across ther exposure frequences. Ths greater varaton n the exposure frequency dstrbuton s ndcated by the larger dstance of K-S. Extenson of the Duraton After the NBD captures the exposure frequency for a fxed duraton, t can generate the exposure frequency dstrbuton for a flexble duraton wth the same dstrbuton parameters, γ and α. In turn, the duraton can be used as a decson varable n the optmzaton model. To modulate the duraton varable, we must have reasonably stable page vew data or else ncorporate addtonal parameters to correct the exposure dstrbuton. However, n our applcaton, the page vew data are stable because n South Korea, the Internet nfrastructure s hghly advanced and the market s mature. Durng our data collecton perod, 49.4% of the total populaton used the Internet, whch shows that Internet usage had reached a mature level. Therefore, we can apply the NBD across flexble perods. In addton, as we show n Table 4, the NBD weekly parameter estmates are stable for the four-week perod. As a consequence, the average exposure frequency and ts varance are very smlar. <Insert Table 4: NBD Parameter Estmates for Week 1, 2, 3, and 4 around here> 15
As we expected, when we extend the duraton of the exposure frequency dstrbuton from one to four weeks usng the same parameters of the NBD, t generates more errors. The maxmum dstance of K-S ncreases substantally from 1.27% to 7.77%. As n the one-week case, stes wth wder reaches generate more errors. Although we lack a sold explanaton of the estmaton errors and ther drecton (.e., under- or overestmaton), the K-S dstance for the four-week extenson provdes nformaton about the range of errors that may be generated when researchers use one-week parameter estmates across multple weeks (see Table 3). Ad Effcency Curve Because our model s flexble enough to measure ad performance by varyng the campagn duraton, we must check the effcency of advertsng across our lsted stes. The effcency curve of an advertsement can be obtaned as the combnaton of the cost and the effectveness functon of a decson varable. For example, Danaher and Rust (1994) report an effcency curve as a functon of gross ratng ponts that can measure effectveness, as a rato to cost, n terms of reach, effectve reach, ncremental sales, or awareness, dependng on the goals of the campagn. For our purposes, because the advertser pays only for vald exposures on the Internet, we can measure the effcency of an ad by computng the cost of ncreasng the effectveness measure by a unt as a functon of the campagn duraton. We obtan the cost functon by multplyng the ad rate by the number of exposures. The ad rate s gven by r (on a CPM bass), whch means that the advertser pays $r for 1000 exposures at ste. The number of exposures s computed by summng the number of the total populaton exposures accordng to the exposure frequency dstrbuton, K k = 1 M * k * f ( X ( t ) = k ). The cost functon can be wrtten as r 1000 K k = 1 M * k * f ( X ( t ) = k ). The effectveness functon, p( X = k)* M * f ( X ( t) = k), s the sum of K k= 1 the values of the effectveness measure, whch represents message recall and can be computed by multplyng the probablty of recall at k exposures, p(x = k), by the number of consumers exposed k 16
tmes, M*f(X(t) = k), as a functon of campagn duraton t. As a result, the ad effcency curve, 1000 * K k = 1 K r k = 1 p ( X k * = k ) * f ( X ( t ) = f ( X ( t ) = k ) k ), can be structured as a functon of campagn duraton. In Fgure 2, we present a graph that dsplays the effcency curve of the number of consumers who recall the ad message when $1 s spent. < Insert Fgure 2: Advertsng Effcency Curve around here> < Insert Table 5: Advertsng Effcency around here> As the fgure shows, ad effcency deterorates as the duraton of the campagn ncreases, manly due to the lmted margnal ncrease of ad effectveness that cannot compensate for the ad cost ncrease for hgh repeat exposures. Across stes, those wth lower average exposures per vstor tend to enjoy hgher ad effcency. As we show n Table 5, ste 4 outperforms all other stes because t has the fewest average page vews per vstor (5.6), whch mnmzes spendng for consumers who experence hgh repeat exposures. In contrast, ad effcency s low among stes wth hgher average page vews per vstor (e.g., stes 1, 8, and 9), because the advertser must pay to expose the same consumers to the advertsement repeatedly and therefore experences low returns. However, no one ste s systematcally domnated by another ste that performed better n the short term. Because each Web ste has a dfferent type of exposure frequency dstrbuton, each provdes a dfferent effcency curve projecton as a functon of duraton. For example, ste 1 records a better ad effcency than ste 2 for the frst 3.5 weeks, but ste 2 outperforms ste 1 for longer perods because t accommodates more vstors n the effectve range of low exposure frequency. From Fgure 2, advertsers could magne a horzontal lne of so-effcency that compares ad effcency across stes. If an advertser wants to lmt the budget per vstor, t can fx the duraton of lsted stes at that specfc so-effcency lne. An optmum set of ste duratons can be algned along one such soeffcency lne. 4.3. Optmzaton Results The specfc data for our optmzaton model are as follows: The total Internet populaton (M) s estmated as 23,658,097, accordng to KoreanClck. 17
The advertsng cost s $1 per 1000 exposures (CPM, r), close to the market prce. We apply the same ad performance functon of repeat exposure (p(x = k)), as reported by Lee and Brley (2004), to all ten stes. The value of our objectve functon represents the number of panel members who recall the ad message after beng exposed n tmes. Those who recalled the ad message wthout beng exposed (24.21% of the total populaton) are not taken nto account n our objectve functon. Our resultng Internet meda plannng model, whch combnes the exposure frequency dstrbuton of the lsted stes and the message recall rate functon of repeat exposure, s solved usng the specalzed Lngo 8.0 software to obtan the optmal set of lsted ste duratons that maxmzes the number of consumers who recall the ad message. We present the optmzaton results for three budget levels ($300,000, $500,000, and $700,000) n Table 6, along wth the margnal ncrease (dual prces) of message recall vstors. <Insert Table 6: Optmal Ste Duraton around here> As we dscussed prevously, ths nonlnear programmng model has a complex objectve functon that requres a relatvely long computaton tme; for ths model, t took a computer powered by an Intel Celeron 2.2 GHz processor wth 224 MB RAM almost ten mnutes to run t. The optmal duratons of the lsted stes provde some useful nsghts. Frst, all ten stes should focus on maxmzng the number of message recalls, not on an optmum determned by the lmted number of stes wth wde reach. The selecton of stes depends largely on the shape of the ad performance functon. As the margnal return decreases, stes could enter the optmal set f ther duraton s farly short, but unless there s a mnmum duraton requrement, all stes can be used to maxmze the ad performance. Second, the optmal ad duraton s much longer for stes wth low average exposures per vstor (stes 4, 5, 6, and 10) than for those wth hgh averages (stes 1, 8, and 9). Ths result s a logcal consequence of the ad effcency curve presented prevously, n that ad effcency s much greater on stes wth low average exposures per vstor than on ste wth hgh ones. Thrd, the effcency of an ad campagn deterorates substantally, manly due to dmnshng returns, as the ad budget ncreases, as exemplfed by the stuaton n whch the prce of the budget 18
decreases from 17.53 ($300,000) to 8.19 ($700,000). These fndngs ndcate that the advertser can add another 17.53 consumers who recall the ad message by ncreasng the budget by $1 when the ntal budget s $300,000 but can add only 8.19 consumers when the startng budget s $700,000. Ths fndng enables advertsers to compute both ther ROI for the pont at whch ad performance wll begn to deterorate substantally. Our fndngs appear to be n conflct wth the current practce of Internet advertsng. Advertsers n Europe selected 1.9 to 2.4 stes for a campagn for an average duraton of 6 8 weeks. 2 That s, advertsers tend to lmt ther number of stes. Accordng to our results, these advertsers are wastng ther budget substantally, because they have concentrated ther campagn on a small number of stes for a long perod, whch generates too many consumers who are exposed too many tmes. To evaluate ths potental waste, we compare ad performance (number of message recall vstors) for three cases: all stes are programmed (optmal), fve stes are, and only three stes (one from each category) are, as n Table7. To determne the performance dfference accordng to the campagn duraton and n lne wth current ad campagn practce, we dvde the three-ste cases nto two subgroups each: wth a fourweek constrant on the maxmum duraton and wthout. In all cases, we establshed a budget of $500,000 and determned the optmal set of ste duratons to maxmze the number of message recall vstors. <Insert Table 7: Ad Performance Comparson around here> At frst glance, the combnatons of only three stes are largely domnated n performance by the optmal soluton of all ten stes. When all stes are used, the advertser can capture approxmately 17 mllon vstors who recall the ad message. In contrast, for the combnaton of stes 1, 4, and 8, t captures approxmately 10 mllon vstors, a drop of 39.8%, and for the combnaton of stes 3, 5, and 9, t captures only 7.5 mllon vstors, only 57.9% of the optmal case. Our fndngs regardng ad performance therefore demonstrate the terrble amount of waste that takes place n current Internet ad spendng practces that lmt the number of used stes. As the number gets smaller, the ad performance 2 These results are based on Internet ad campagns from 3130 stes n 14 countres n Europe (LemonAd 2002). Its Internet lnk s unfortunately no longer avalable. 19
deterorates because of the greater exposures n a less effcent, hgh repeat exposure zone. The reduced number of message recall consumers represents the magntude of waste n the three-ste cases, and the unattractve dual prce reflects ther neffcency. The optmal soluton wth the ten stes suggests 10.94 as the dual prce of budget spendng; that s, the advertser captures 10.94 vstors who recall the ad message for any extra $1 n ad budget spendng. Each combnaton of the three stes costs the advertser twce as much n ad budget than the optmal case to capture the same number of ad recall vstors. Accordng to the Interactve Advertsng Bureau (2004), a phenomenon of hgh ad spendng concentraton has occurred among Web stes wth wde reach, n whch the top ten Web stes account for more than 70% of the total ad spendng. Ths phenomenon renforces the magntude of potental waste that runs rampant n current Internet ad practces. 5. Manageral Implcatons and Model Lmtatons The man objectve of our research s to provde marketng managers and ad agences wth an optmzaton tool that maxmzes the ROI of ther advertsng budgets by hghlghtng the optmal combnaton of stes and the deal campagn duraton for each ste. The advantages of usng our model for Internet meda plannng, especally the flexblty of exposure dstrbuton, are multfold. Because each ste needs only two parameter estmates for the NBD to generate the exposure frequency of any duraton, the computatonal burden s greatly reduced for a large number of Web stes; t s not necessary to generate the exposure frequency at each step of the optmzaton accordng to the combnaton of chosen stes. In addton, t mnmzes complexty when an advertser wants to conduct segment-level meda plannng for whch t s necessary to obtan addtonal parameter estmates for each segment. The smplcty of the data s another advantage of our model. The exposure frequency data that we use can be obtaned easly from any market research company that keeps a user panel. Varous types of exposure frequency data can be generated from the raw panel page vew data, and the parameters of the NBD can be estmated easly by standard software such as MS Excel Solver or other packages such as MATLAB. In addton, the pretest of ad effectveness becomes more and more 20
affordable on the Internet. Companes such as DynamcLogc and DoubleClck offer ths servce for less than $2,000 per ad. Fnally, our model helps advertsers calculate ther ROI for Internet advertsng by provdng concrete numbers about ad performance and effcency. Our model enables advertsers not only to optmze ther Internet ad schedule but also to fx the rght prce for ther Internet advertsements on the bass of the characterstcs of the exposure dstrbuton of stes. Our fndngs contrast wth the current Web prcng practces, because the ad rate should be based on the average exposures per vstor rather than on ts reach. Despte these major advantages, our model does not nclude some aspects that should be addressed to refne ts performance. Frst, we do not take nto consderaton exposure duplcaton across stes. As a result, our objectve functon may overestmate ad performance. The magntude of ths overestmaton may ncrease when the duplcaton rate of chosen stes ncreases or the plannng unt s restrcted to nteger values. The complex nature of Internet meda plannng, whch requres varyng duraton varables and multple stes, does not allow us to use a smple weght between two stes to reduce the duplcaton, as Headen, Klompmaker, and Teel appled (1976). A possble soluton may be to compute the overlapped exposure dstrbuton among stes. Park and Fader (2004) fnd a substantal mprovement n predctng the ntervst behavor of two-ste cases when they use a Sarmanov famly of multvarate dstrbutons (e.g., exponental tmng process and gamma mxng dstrbuton). Exposure dstrbutons n a canoncal form should be developed to correct the ad performance by assessng the wdth and depth of overlapped exposures across stes. Second, our model does not address the forgettng effect. Whereas MEDIAC (Lttle and Lodsh 1969) ntegrates the forgettng effect as a memory constant by updatng the exposure level at each perod, ADMOD (Aaker 1975) does not ncorporate t drectly. In our model, the forgettng effect may not need to be ncluded due to the relatvely short campagn duratons, for whch the forgettng effect s not statstcally sgnfcant (Drèze and Hussherr 2003). However, more refned research on repeat exposures of an Internet advertsement wth varyng condtons, such as context and tme lap, should be conducted to enhance our model performance. 21
Thrd, our Internet meda plannng optmzaton model has a complex nonlnear objectve functon. If for small problems (as the one solved here), the computatonal tme s not an ssue, the search for optmal or non domnated solutons wll be computatonally tme consumng for large problems. In ths case, we would need to revert to the development of a heurstc approach to solve the problem n reasonable computatonal tme. Ths s part of an on-gong research project. 6. Concluson The results of our Internet meda plannng model provde useful nsghts that can enhance the effcency of Internet advertsng. An advertser must consder as many stes as possble because advertsng on a wde selecton of stes mnmzes wasteful spendng. If a campagn s concentrated on only two or three stes, the advertser must extend the campagn duraton of those chosen stes. Ths extenson substantally penalzes the effcency of the campagn, because t becomes more expensve to get vstors to recall the ad message. If the advertser uses meda plannng tools developed for tradtonal meda, t must carefully choose the proper ndcator to select ts stes. In the case of tradtonal meda, an advertser would prefer stes wth a wde reach and hgh average exposures, as long as the ad rate s the same. But n the case of Internet, the advertser must pay attenton to another ndcator: the average exposures per vstor. Because the prcng practce for an Internet advertsement s based on the number of exposures (page vews), buyers should consder the effcency ssue frst. In turn, because the ad effectveness functon of exposure frequency decreases margnally, the choce of a Web ste wth hgh average exposures per vstor mnmzes the effcency of the campagn. Therefore, the advertser must favor those stes wth low average exposures per vstor, as long as these stes meet the mnmum reach requrements. 22
REFERENCES Aaker, Davd A. 1975. ADMOD: An advertsng decson model. Journal of Marketng Research 12 37-45. Belch, George E. 1982. The effects of televson commercal repetton on cogntve response and message acceptance. Journal of Consumer Research 9 56-65. Bell, Hudson and Nelson K.H. Tang. 1998. The effectveness of commercal Internet Web stes: A user's perspectve. Internet Research: Electronc Networkng Applcatons and Polcy 8(3) 219-28. Beauvllan, Olver. 2004. Evoluton of meda use n Europe, Jupter Research, vson report. Berger, Ida and Andrew A. Mtchell. 1989. The effect of advertsng repetton on atttude accessblty, atttude confdence/certanty and the atttude-behavor relatonshp. Journal of Consumer Research 16 269-79. Berlyne, Danel E. 1970. Novelty, complexty, and hedonc value. Percepton and Psychophyscs 8 279-86. Brown, D.B. and M.R. Warshaw. 1965. Meda selecton by lnear programmng. Journal of Marketng Research 2(1) 83-88. Cacoppo, John T. and Rchard E. Petty. 1979. Effects of message repetton and poston on cogntve response, recall, and persuason. Journal of Personalty and Socal Psychology 37 97-109. Calder, Bobby J. and Bran Sternthal. 1980. Televson commercal wearout: An nformaton processng vew. Journal of Marketng Research 18 173-6. Danaher, Peter J. 1988a. A parameter estmaton for the Drchelet-multnomal dstrbuton usng supplementary beta-bnomal data. Communcatons n Statstcs A17 (6) 777-88. Danaher, Peter J. 1988b. A log-lnear model for predctng magazne audences. Journal of Marketng Research 25 356-62. Danaher, Peter J. 1989. An approxmate loglnear model for predctng magazne audences. Journal of Marketng Research 26(November) 473-9. Danaher, Peter J. and Roland Rust. 1994. Determnng the optmal level of meda spendng. Journal of Advertsng Research 34(1) 28-34. 23
Day, R.L. 1962. On methods: Lnear programmng n meda selecton. Journal of Advertsng Research 2(2) 40-44. Drèze, Xaver and Franços-Xaver Hussherr. 2003. Internet advertsng: Is anybody watchng? Journal of Interactve Marketng 17(4) 8-23. Drèze, Xaver and Fred Zufryden. 1998. Is Internet advertsng ready for prme tme? Journal of Advertsng Research 38 (3) 7-18. Drèze, Xaver and Fred Zufryden. 2000. Internet advertsng: The medum s the dfference. Consumpton, Markets & Culture 4(1) 23-37. Ehrenberg, A.S.C. 1959. The pattern of consumer purchases. Appled Statstcs 8 26-41. Headen, Robert R., Jay E. Klompmaker, and Jesse E. Teel Jr. 1976. TV audence exposure. Journal of Advertsng Research 16 49-52. Interactve Advertsng Bureau. 2004. Internet Advertsng Revenue Report 2003. http://www.ab.net/resources/adrevenue/pdf/iab_pwc_2003.pdf. Krugman, Hubert E. 1972. Why three exposures may be enough. Journal of Advertsng Research 12(6) 11-14. Leckenby, John D. and Shzue Ksh. 1984. The Drchlet multnomal dstrbuton as a magazne exposure model. Journal of Marketng Research 21(1) 100-106. Lee, A.M. 1962. Decson rules for medan schedulng: Statc campagns. Operatonal Research Quarterly 13 229-242. Lee, Janghyuk and Donnel Brley. 2004. Repeat exposure effects of Internet advertsement. HEC Workng Paper. Lttle, John D.C. and Leonard M. Lodsh. 1969. A meda plannng calculus. Operatons Research 17(1) 1-35. Massey, Frank J., Jr. 1951. The Kolmogrov-Smrnov test for goodness of ft. Amercan Statstcal Assocaton Journal 46 68-78. Metherngham, Rchard A. 1964. Measurng the net cumulatve coverage of a prnt campagn. Journal of Advertsng Research 4(4) 23-28. 24
Mller, D.W. and M.K. Starr. 1960. Executve Decsons and Operaton Research. Prentce Hall, Englewood Clffs, NJ. Nedungad P., Andrew A. Mtchell, and Ida Berger. 1993. A framework for understandng the effect of advertsng exposure on choce. A.A. Mtchell, ed. Advertsng Exposure, Memory and Choce. Lawrence Erlbaum Assocates, Hllsdale, NJ. 89-116. Novak, Thomas P. and Donna L. Hoffman. 1997. New metrcs for new meda: Toward the development of Web measurement standards. World Wde Web Journal 2(1) 213-46. Park, Young-Hoon and Peter S. Fader. 2004. Modelng browsng behavor at multple Webstes. Marketng Scence 23(3) 208-303. Pechmann, Cornela and Davd W. Stewart. 1989. Advertsng repetton: A crtcal revew of wearn and wearout. James H. Legh and Claude R. Martn Jr., eds. Current Issues and Research n Advertsng 1988. Unversty of Mchgan, Ann Arbor, MI. 285-89. Ray, Mchael L. and Alan G. Sawyer. 1971. Repetton n meda models: A laboratory technque. Journal of Marketng Research 8 20-9. Ray, Mchael L. and Edward C. Strong. 1971. Frequency effects revsted. Journal of Advertsng Research 11 14-20. Rust, Roland T. 1985. Selectng network televson advertsng schedules. Journal of Busness Research 13 483-94. Rust, Roland T. and Robert P. Leone. 1984. The mxed meda Drchlet-multnomal dstrbuton: A model for evaluatng televson magazne advertsng schedules. Journal of Marketng Research 21 89-99. Schmttlen, Davd C., Donald G. Morrson, and Rchard Colombo. 1987. Countng your customers: Who are they and what wll they do next? Management Scence 33(1) 1-24. Starsch, Stanley F. 1965. Lnear programmng and space-tme consderatons n meda selecton. Journal of Advertsng Research 5(4) 40-46. Vdale, M.L. and H.B. Wolfe. 1957. An operatons research study of sales response to advertsng. Operatons Research 5 370-281. 25
Table 1 Ste Profle Ste Type Reach Average Average Page Male Female Page Vews Vews/Vstor Vstors Vstors 1 Communty portal 81.8% 20.7 25.3 61.5% 38.5% 2 Communty portal 14.4% 2.0 13.7 61.4% 38.6% 3 Communty portal 17.3% 2.1 11.9 62.9% 37.1% 4 Communty portal 11.2% 0.6 5.6 64.2% 35.8% 5 News 9.7% 0.8 8.6 78.2% 21.8% 6 News 8.5% 0.7 8.6 77.8% 22.2% 7 News 5.3% 0.7 13.8 81.9% 18.1% 8 Search engne 49.2% 10.1 20.5 62.5% 37.5% 9 Search engne 60.3% 12.6 20.9 61.8% 38.2% 10 Search engne 30.2% 3.2 10.5 67.1% 32.9% Table 2 NBD Parameter Estmates (MS Excel Solver and MATLAB) Ste 1 Ste 2 Ste 3 Ste 4 Ste 5 Ste 6 Ste 7 Ste 8 Ste 9 Ste 10 γ Excel Solver 0.482 0.040 0.052 0.042 0.031 0.027 0.014 0.165 0.222 0.102 MATLAB 0.486 0.040 0.052 0.042 0.031 0.027 0.014 0.167 0.227 0.104 α Excel Solver 0.023 0.020 0.025 0.067 0.037 0.036 0.019 0.017 0.018 0.033 MATLAB 0.024 0.020 0.026 0.067 0.037 0.036 0.019 0.018 0.020 0.035 Pagevew Mean Excel Solver 20.6 2.0 2.0 0.6 0.8 0.7 0.7 9.9 12.1 3.1 MATLAB 20.3 2.0 2.0 0.6 0.8 0.7 0.7 9.4 11.2 3.0 Pagevew Varance Excel Solver 898.3 99.3 83.1 10.1 23.7 20.7 40.1 603.0 675.4 97.6 MATLAB 870.6 100.0 78.3 10.0 23.8 20.7 40.0 537.1 567.8 89.7 Table 3 Kolmorov-Smrnov Dstance Ste 1 Ste 2 Ste 3 Ste 4 Ste 5 Ste 6 Ste 7 Ste 8 Ste 9 Ste 10 Average 1 week 2.52%*** 1.17% 1.66% 1.46% 0.25% 0.22% 0.34% 1.03% 2.20%** 1.86% 1.27% 4 weeks 11.88%* 3.61%* 7.30%* 9.62%* 4.23%* 5.87%* 3.40%* 11.91%* 10.30%* 9.59%* 7.77% *** Sgnfcant at α = 10%. ** Sgnfcant at α = 5%. * Sgnfcant at α = 1%. 26
Table 4 NBD Parameter Estmates for Weeks 1 4 Ste 1 Ste 2 Ste 3 Ste 4 Ste 5 Ste 6 Ste 7 Ste 8 Ste 9 Ste 10 γ Week1 0.486 0.040 0.052 0.042 0.031 0.027 0.014 0.167 0.227 0.104 Week2 0.498 0.038 0.058 0.046 0.032 0.029 0.015 0.180 0.235 0.110 Week3 0.483 0.044 0.054 0.044 0.032 0.033 0.018 0.185 0.234 0.108 Week4 0.467 0.038 0.053 0.042 0.030 0.030 0.017 0.186 0.220 0.105 α Week1 0.024 0.020 0.026 0.067 0.037 0.036 0.019 0.018 0.020 0.035 Week2 0.025 0.019 0.029 0.075 0.040 0.037 0.020 0.018 0.021 0.037 Week3 0.024 0.022 0.026 0.068 0.037 0.037 0.025 0.018 0.020 0.037 Week4 0.024 0.018 0.028 0.067 0.037 0.040 0.026 0.019 0.020 0.038 Mean Week1 20.3 2.0 2.0 0.6 0.8 0.7 0.7 9.4 11.2 3.0 Week2 19.8 2.1 2.0 0.6 0.8 0.8 0.7 10.0 11.3 3.0 Week3 19.9 2.1 2.1 0.6 0.9 0.9 0.7 10.0 11.5 2.9 Week4 19.2 2.1 1.9 0.6 0.8 0.8 0.6 9.6 11.1 2.8 Varance Week1 870.6 100.0 78.3 10.0 23.8 20.7 40.0 537.1 567.8 89.7 Week2 810.6 111.9 72.6 8.6 20.4 21.5 37.9 565.9 553.8 83.0 Week3 838.0 97.0 82.3 10.0 24.6 25.4 30.0 555.3 578.1 81.6 Week4 804.1 117.8 70.1 10.1 22.2 20.3 26.0 508.4 577.0 75.7 Table 5 Advertsng Effcency Duraton (week) 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Ste 1 61.88 33.85 23.52 18.09 14.73 12.45 10.81 9.58 8.63 7.88 Ste 2 47.91 27.67 20.00 16.01 13.63 12.08 11.00 10.21 9.62 9.15 Ste 3 116.91 70.86 52.10 41.65 34.92 30.19 26.67 23.96 21.80 20.04 Ste 4 218.17 140.61 106.42 86.64 73.56 64.21 57.14 51.60 47.12 43.41 Ste 5 82.45 48.93 35.62 28.32 23.68 20.48 18.15 16.40 15.04 13.96 Ste 6 87.97 52.39 38.20 30.40 25.42 21.97 19.44 17.53 16.04 14.85 Ste 7 75.78 44.97 32.77 26.09 21.88 18.99 16.92 15.37 14.17 13.22 Ste 8 47.23 26.20 18.38 14.29 11.80 10.16 9.02 8.18 7.55 7.06 Ste 9 47.07 25.83 18.01 13.91 11.42 9.77 8.61 7.76 7.13 6.63 Ste 10 82.76 48.03 34.48 27.13 22.47 19.26 16.92 15.14 13.76 12.65 27
Table 6: Optmal Ste Duraton wth Three Budget Amounts Duraton (weeks)/budget Amount ($) 300K 500K 700K Ste 1 0.17 0.26 0.34 Ste 2 0.33 0.59 0.91 Ste 3 0.42 0.73 1.12 Ste 4 1.10 1.95 3.00 Ste 5 0.62 1.11 1.74 Ste 6 0.62 1.12 1.77 Ste 7 0.33 0.60 0.97 Ste 8 0.22 0.36 0.50 Ste 9 0.22 0.36 0.49 Ste 10 0.48 0.83 1.20 # of Message Recall 15050880 17787410 19667450 Dual Prce (Ad Effcency) 17.53 10.94 8.19 Table 7 Ad Performance Comparson All Stes Stes 1,3,5,6,7 Stes 1, 4, 8 Stes 3, 5, 9 No Constrant Max 4 weeks No Constrant Max 4 weeks No Constrant Max 4 weeks Ste 1 0.26 0.50 0.53 0.50 0.53 0.00 0.00 Ste 2 0.59 0.00 0.00 0.00 0.00 0.00 0.00 Ste 3 0.73 2.82 3.27 0.00 0.00 6.41 4.00 Ste 4 1.95 0.00 0.00 7.83 0.00 0.00 0.00 Ste 5 1.11 5.42 4.00 0.00 4.00 15.03 4.00 Ste 6 1.12 5.86 4.00 0.00 0.00 0.00 0.00 Ste 7 0.60 4.31 4.00 0.00 0.00 0.00 0.00 Ste 8 0.36 0.00 0.00 0.86 0.95 0.00 0.00 Ste 9 0.36 0.00 0.00 0.00 0.00 1.21 1.54 Ste 10 0.83 0.00 0.00 0.00 0.00 0.00 0.00 Message Recall Indvduals 17787410 9852930 9846052 10706070 10682060 7483794 7428267 Underperformance -44.6% -44.6% -39.8% -39.9% -57.9% -58.2% Dual Prce (Ad Effcency) 10.94 5.45 5.19 5.53 5.21 4.02 3.29 Underperformance -50.2% -52.6% -49.5% -52.4% -63.3% -69.9% Fgure 1 Repeat Exposure Effect of Message Recall Rate 45% 40% 35% 30% 25% 20% 0 5 10 15 20 25 30 35 40 45 50 Number of Exposures 28
0,5 1,5 2,5 3,5 4,5 5,5 Fgure 2 Advertsng Effcency Curve Message Recall/$ 500 450 400 350 300 250 200 150 Ste1 Ste2 Ste4 Ste6 100 50 0 1 2 3 4 5 6 Duraton (Weeks) 29