Customer loyalty is the central thrust of marketing efforts (Dick and Basu 1994; Evanschitzky et al. 2012), and U.S. firms spend dramatically to build and manage customer loyalty. For example, annual loyalty program outlays have grown 27% since 2010 to exceed $48 billion across 2.7 billion program enrollees in the United States alone, yet less than half of the 22 memberships per household are active (Berry 2013). However, the financial returns of many loyalty-building efforts fail to meet expectations (Henderson et al. 2011; Nunes and Dréze 2006). Even though the concept of “customer loyalty” has been debated for more 60 years (Brown 1952), the mixed returns of loyalty efforts still stem, in part, from divergent theoretical and operational approaches, such as the varied use of attitudinal loyalty without behavioral loyalty or the use of modified word-of-mouth measures as proxies for customer loyalty (Dick and Basu 1994; Keiningham et al. 2007; Oliver 1999; Reinartz and Kumar 2002). To test the consequences of this heterogeneity empirically, we synthesize extant loyalty research to provide parsimonious guidance to both academics and managers seeking to understand the strategy → customer loyalty → performance process.

Although there is no consensus definition of loyalty, extant research generally agrees that it represents a mix of attitudes and behaviors that benefit one firm relative to its competitors (Day 1969; Dick and Basu 1994; Melnyk et al. 2009). Within this conceptual umbrella, researchers often selectively examine loyalty as an attitude, purchase behavior, or multidimensional construct (e.g., attitudes and purchase behaviors, word of mouth). Beyond this broad description there is significant variation in the conceptualization and operationalization of loyalty, which may explain heterogeneity in loyalty-related effects. We map 163 studies published in marketing journals since 1980, at the measurement item level, to capture their heterogeneous research approaches. We then evaluate the degree to which this heterogeneity leads to disparate empirical results, by examining the structural relationships among loyalty, its antecedents, and its outcomes according to meta-analytic data (Zablah et al. 2012). Finally, we examine possible moderators of the effect of loyalty on outcomes with a meta-regression (Rubera and Kirca 2012).

The results of these analyses produce three main contributions. First, we reconcile the differential effects of two theoretical elements of customer loyalty—attitudes and behaviors—according to meta-analytic results. The antecedents differentially build attitudinal and behavioral loyalty, and attitudinal and behavioral loyalty differentially influence managerially relevant outcomes such as word of mouth and performance (i.e., sales, share of wallet, profit performance, and other measurable changes). The common research practice of using single-element measures of loyalty (i.e., only attitude or behavior) thus leads to mixed guidance regarding the effect of loyalty on performance. This concern is especially problematic when we note that most research (65% of studies in our sample) examines loyalty as an end outcome, such that it serves as a potentially misleading proxy for performance.

Second, we examine the moderating role of the measurement composition (e.g., ratio of attitudinal vs. behavioral loyalty items, inclusion of word-of-mouth items) and study-specific characteristics (e.g., temporal orientation) to explain additional variance in loyalty-related effects, using a meta-analysis of the sources of variation identified by our literature review and item-level coding procedure. We find for example that measures composed of combined attitudinal and behavioral items are more effective than attitude-only or behavior-only measures. Varying measurement compositions and study-specific characteristics of loyalty also produces different effects, depending on the type of loyalty and the outcome of interest. These analyses help clarify the heterogeneity in loyalty effects that stem from conceptualizations and the study context.

Third, we provide prescriptive guidance for researchers and managers. In particular, we recommend strategies that build attitudinal and behavioral loyalty, the use of loyalty items obtained from the measures that are most predictive of performance and word of mouth, and consideration of contextual information (e.g., business vs. consumer markets) that can leverage the effect of customer loyalty. We thus attempt to add clarity to the varied conceptual and empirical findings achieved through decades of research by answering three simple but important bquestions: What is customer loyalty? (theoretical), What are researchers doing? (measurement approaches), and What actually matters? (empirical results).

Theoretical domain of customer loyalty

The rich early history of customer loyalty research allowed Jacoby and Chestnut (1978) to cite more than 50 definitions of it. Following more recent, elaborate theoretical expositions (Dick and Basu 1994; Oliver 1999), current theories most often delineate attitudinal loyalty and behavioral loyalty as customer loyalty’s primary elements (Chaudhuri and Holbrook 2001).

Loyalty as favorable attitudes and purchase behaviors

Attitudes are the first element of customer loyalty. People are motivated information processors who use information to form their attitudes (Ahluwalia 2000; Moorman et al. 1993). Attitudinal loyalty then is a “cognition” or “pleasurable fulfillment” that favors a particular entity (Oliver 1999, p. 35; see also Chaudhuri and Holbrook 2001). Strong, loyal attitudes result from systematic evaluations (Petty and Cacioppo 1986) and influence many customer performance–related behaviors (Park et al. 2010; Petty, Haugtvedt, and Smith 1995). Strong positive attitudes induce “defensive processes” in the face of competition that cause customers to resist competitive offers, even when they are objectively better (Ahluwalia 2000, p. 230). Oliver (1999, p. 34) captures this idea when he notes that loyalty persists “despite situational influences and marketing efforts [that have] the potential to cause switching behavior.”

Purchase behaviors, the second element of loyalty, are also central in loyalty research (Ailawadi et al. 2008; De Wulf et al. 2001). Behavioral loyalty entails repeated purchases that stem from a conation or action orientation involving a “readiness to act” to the benefit of a particular entity (Oliver 1999, p. 35; see also Chaudhuri and Holbrook 2001; De Wulf et al. 2003). Various examinations of customer loyalty focus on measuring behaviors, such as repeated purchase behaviors, that have obvious benefits for a firm’s financial performance. This view has given rise to stochastic models of loyalty, including the use of recency, frequency, monetary theory; churn/retention; and purchase sequences. Prior research also demonstrates the benefits of behavioral loyalty. For example, Gupta et al. (2004) find that the impact of a 1% improvement in customer retention is five times greater than the effects of a similar increase in margin. Yet purely behavioral approaches are often agnostic about the psychological processes associated with customer action. They ignore the real possibility that repetitive purchase behavior arises from situational constraints, such as a lack of viable alternatives, or usage situations, such as habit (Henderson et al. 2011). Regardless of their cause, customer behaviors can explain financial outcomes as loyalty-based purchase activities.

Measurement composition and study-specific characteristics in loyalty research

Despite clear delineations between attitudes and purchase behaviors, theories of customer loyalty suggest both are integral (Dick and Basu 1994; Oliver 1999). Furthermore, some researchers characterize loyalty as a general orientation reflected by non-purchase behaviors, such as advocacy (Jones et al. 2008), willingness to pay a premium (Chaudhuri and Holbrook 2001), or continued silence in the hope that things get better (Hirschman 1970). As a result, loyalty measures are often fuzzy; despite its duration, literature on loyalty presents ad hoc measures that are sometimes composed of attitudinal items, sometimes composed of behavioral items, or both—or even both together with items that measure ancillary constructs. For example, many researchers include word of mouth (WOM) items in their operationalization of customer loyalty (Evanschitzky et al. 2012), despite both theoretical (Dick and Basu 1994) and empirical (de Matos and Rossi 2008; Söderlund 2006) arguments for their separation.

In addition, researchers examine customer loyalty in various research settings, creating research-specific measurement characteristics that may exert influences on results. We consider two key characteristics: temporal orientation and target. Temporal orientation refers to whether loyalty is measured as a past account or future predictions of loyal attitudes and behaviors. For example, a customer might be asked to recall previous instances of being loyal (Ailawadi et al. 2008; Davis-Sramek et al. 2009) or else estimate future intentions to be loyal (Johnson et al. 2006; Wagner et al. 2009). The target is the attribution customers make about to whom or what they are loyal. For example, loyalty may be “owned by” or directed toward the selling firm or a salesperson (Palmatier et al. 2007). We consider the effects of both measurement composition and study-specific characteristics in the subsequent sections.

Conceptual model and hypotheses

To understand how loyalty may vary depending on its operationalization, we first consider how linkages in the antecedents → customer loyalty → outcomes framework vary depending on the use of attitudinal or behavioral loyalty. By examining attitudinal and behavioral loyalty as separate mediators, we can isolate relative differences in both their main and interaction effects. Therefore, we evaluate how the theory underlying each element supports distinct predictions. We only include constructs in our conceptual framework if prior studies have demonstrated at least three empirical links between the antecedent and attitudinal and behavioral loyalties, to enable our tests of the differences through meta-analysis. We also include measures that feature both attitudinal and behavioral loyalties to support moderation analyses. We identified four antecedents (commitment, trust, satisfaction, and loyalty incentives) and two outcomes (WOM and performance) that meet these criteria. Table 1 summarizes the constructs included in the model, their definitions, and common aliases. Furthermore, we evaluate other sources of heterogeneity identified in our literature review (measurement composition, study-specific characteristics) and probe for additional sources of heterogeneity by coding study-level factors to test their effects on loyalty-to-outcome linkages. Consistent with our research questions, our hypotheses focus on comparative differences among the links in our model.

Table 1 Review of construct definitions, aliases, and representative papers
Full size table

Loyalty antecedents

Commitment, trust, satisfaction, and loyalty incentives (e.g., reward programs, perks, favorable treatment) have all been positively linked to customer loyalty, but we expect them to have differential effects on attitudinal and behavioral loyalties. Attitudinal loyalty results from positive evaluations of a seller based on previous exchange experience (Brakus et al. 2009; Liu-Thompkins and Tam 2013). Drivers of loyalty that primarily enhance a customer’s evaluation of the exchange should have a stronger effect on attitudinal than on behavioral loyalty. Conversely, behavioral loyalty results from situational triggers and habit (Gustafsson et al. 2005; Johnson et al. 2006), which may not involve a strong attitudinal component. Thus drivers of loyalty that primarily operate as situational triggers in an exchange should have a stronger effect on behavioral than on attitudinal loyalty. Commitment, or the desire to maintain a valued relationship (Moorman et al. 1992), trust, which is confidence in the reliability and integrity of a seller (Morgan and Hunt 1994), and satisfaction, which is the perceived difference between prior expectations and actual performance (Tse and Wilton 1988), all contribute to a customer’s positive experience. Commitment and trust create the sense that customers are in a pleasurable relationship rather than a passing transaction (Palmatier et al. 2006); satisfaction provides a comparative basis (prior expectation versus actual experience) on which to develop attitudes (Geyskens and Steenkamp 2000). Therefore, commitment, trust, and satisfaction should have stronger effects on attitudinal loyalty than on behavioral loyalty. Alternatively, loyalty incentives are additional “extrinsic” enticements meant to encourage repeat patronage (De Wulf et al. 2001), so they might operate as repurchase reminders that reduce effortful purchase considerations and encourage habitual purchasing or as rewards for the positive behavior of repurchasing (Henderson et al. 2011). Thus we expect loyalty incentives to operate primarily through behavioral rather than attitudinal loyalty. Finally, as the theory of planned behavior predicts and prior research demonstrates, we expect attitudinal loyalty to affect behavioral loyalty positively (Ajzen and Fishbein 1980; Chaudhuri and Holbrook 2001).

  1. H1:

    (a) Commitment, (b) trust, and (c) satisfaction have stronger positive effects on attitudinal loyalty than on behavioral loyalty.

  2. H2:

    Loyalty incentives have stronger positive effects on behavioral loyalty than on attitudinal loyalty.

  3. H3:

    Attitudinal loyalty positively affects behavioral loyalty.

Loyalty outcomes

We anticipate that the evaluation- and action-based mechanisms underlying attitudinal and behavioral loyalties differentially influence WOM and performance outcomes. Because attitudinal loyalty is associated with positive evaluations of a seller, and evaluations are abundant and easy to communicate, the effect of attitudinal loyalty should be strongest for WOM. Behavioral loyalty instead might not include a strong, accessible attitudinal component, which provides the basis for WOM (Berger and Schwartz 2011). Therefore, the effect of behavioral loyalty on WOM should be weaker than that of attitudinal loyalty. For performance, we expect an opposite pattern of effects. Attitudinal loyalty tends to be based on conformity (Berger and Heath 2008) and may exist despite situational constraints (e.g., financial, location) that impede actual purchases (i.e., loyalty can be aspirational), so we expect its effect on performance to be weaker. Instead, behavioral loyalty, which is based on a conation or readiness to act and is tied directly to purchase, should have a stronger effect on performance.

  1. H4:

    The effect of attitudinal loyalty on WOM is greater than the effect of behavioral loyalty.

  2. H5:

    The effect of behavioral loyalty on performance is greater than the effect of attitudinal loyalty.

Role of measurement composition and study-based characteristics

Researchers use different compositions of loyalty measures, which should moderate the effect of loyalty on outcomes. Dick and Basu (1994) provide a strong conceptual argument that neither a relatively high attitude nor a behavioral inclination to purchase repeatedly are sufficient to capture customer loyalty fully. Customers with high attitudinal loyalty go to greater lengths to support a seller, and their cognitive biases help them resist competitive persuasion attempts through mechanisms such as avoidance or counterarguments (Ahluwalia 2000; Park et al. 2010). However, these customers also may lack the ability or opportunity to support the seller (e.g., financial constraints). Behavioral loyalty directly increases seller revenues through frequent repurchasing and demonstrates the customer’s ability and opportunity to support the seller. However, these benefits may be short lived if customers lack the motivation to continue their purchase behaviors when their environment changes. Behaviorally loyal customers with low attitudinal loyalty also may be more likely to exploit their relative importance and seek to extract extra concessions from the seller. Thus, customer loyalty should capture both behavioral and attitudinal aspects, to reflect customers’ desire, opportunity, and ability to support the seller financially while avoiding competitors. Because measures that combine attitudes and behaviors capture the variance accounted for by each aspect of loyalty, we expect combined measures of loyalty to exert a stronger effect on performance outcomes than either attitudinal or behavioral measures alone.

Yet broad measures of loyalty that include WOM items might attenuate the predictive effect of loyalty on performance outcomes, because WOM is a socially complex phenomenon that involves self-image concerns, consideration for others’ interests, and serendipitous accessibility (Berger and Schwartz 2011; De Matos and Rossi 2008; Söderlund 2006). Therefore, even though WOM and loyalty are correlated, measures of loyalty that include WOM items may capture unrelated or even countervailing effects (e.g., customers may be very loyal to a condom company but unlikely to recommend it), so they will be less effective at predicting overall performance.

  1. H6:

    Measures of loyalty that include both attitude and behavioral items have a stronger positive effect on WOM than separate measures of attitudinal or behavioral loyalty.

  2. H7:

    Measures of loyalty that include both attitude and behavioral items have a stronger positive effect on performance than separate measures of attitudinal or behavioral loyalty.

  3. H8:

    The positive effect of loyalty on performance appears weaker when loyalty measures include WOM.

We also find systematic variation in the characterization of customer loyalty across extant literature, particularly in terms of the temporal orientation and loyalty target(s). Although research questions or contexts might restrict researchers’ choices of forward-looking versus backward-looking loyalty measures, we expect backward-looking loyalty measures to exhibit a stronger effect on both objective performance and positive WOM. First, backward-looking loyalty assessments tend to be more accurate, because customers avoid the difficulty of imagining obstacles (e.g., price) that might interfere with future purchase behaviors (Ajzen 2002; Zimbardo and Boyd 1999). Second, forward-looking loyalty relies on top-of-mind factors, whereas backward-looking loyalty benefits from subtle psychological mechanisms that offer powerful predictors of future behavior (e.g., cognitive dissonance, habit, sunk costs, switching costs; Henderson et al. 2011). Therefore, both attitudinal and behavioral loyalty should have greater influences on WOM and firm performance when measured as backward-looking rather than forward-looking orientation.

  1. H9:

    Backward-looking loyalty exhibits a stronger effect on (a) objective performance and (b) word of mouth than does forward-looking loyalty.

Although targets of loyalty frequently vary, research often ignores the potential implications of this variance. Loyalty to the firm should be a better predictor of objective performance than loyalty to a salesperson, for several reasons (Palmatier et al. 2007). First, loyalty to a salesperson creates dangers, in that salespeople frequently change positions, so the time when firms can benefit from this loyalty is relatively short; this loyalty also might transfer to a competitor if the salesperson switches firms. Second, salespeople may act opportunistically as agents of the firm and offer their favorite customers unnecessary discounts or perks to gain personal favor (Palmatier et al. 2009). Third, the salespeople to whom the customer is most loyal represent only a limited portion of the firm’s total offering, such that customers usually must deal with multiple salespeople, brands, and locations. This narrow representation restricts the benefits that might accrue if the same amount of loyalty were directed to the firm as a whole. Thus, we expect firm loyalty to have a greater impact on performance than salesperson loyalty does.

However, loyalty to a salesperson may be a better predictor of WOM, because customers can be more confident that others following their recommendations will enjoy a similarly positive experience if they recommend a specific salesperson. Customers view salespeople as more consistent or entatitive targets than firms, which comprise multiple salespeople, brands, and locations. Thus, when assessing an individual as opposed to a firm, “customers are quicker to form judgments, believe the judgments more strongly, and are more likely to act on the beliefs” (McConnell et al. 1997, p. 759). Loyalty to salespeople therefore should have a greater impact on positive WOM than loyalty to a firm, because it is based on a more consistent target.

  1. H10a:

    The positive effect of customer loyalty on objective performance increases when loyalty targets the firm rather than the individual salesperson.

  2. H10b:

    The positive effect of customer loyalty on word of mouth increases when loyalty targets the individual salesperson rather than the firm.

Empirical study

To determine “What is customer loyalty?” we begin by describing our data collection process. Then, to investigate “How is loyalty measured?” and “What actually matters?” we code the composition of the measurement items that we find in individual studies and perform a random effects meta-analysis of reliability adjusted, r-to-Z–transformed correlations between the constructs in our conceptual model. With these meta-analytic data, we perform a structural path analysis (Model 1), followed by a multivariate moderation analysis, or “meta-regression” (Models 2 and 3), to test our hypotheses.

Methodology

Data sample and criteria for inclusion

We used several approaches to identify potential studies for inclusion in our analyses. In the search process, we referred to the EBSCO database and reviewed journals ranked in the highest tier (Polonsky and Whitelaw 2006), namely, Journal of Marketing, Journal of Marketing Research, Marketing Science, Journal of the Academy of Marketing Science, Journal of Consumer Research, and Journal of Retailing, during 1980–2013. For each article of each volume of these journals we evaluated whether the authors measured any construct with “loyalty” in its name (e.g., “consumer loyalty,” “customer loyalty,” attitudinal loyalty,” “behavioral loyalty”), with the exception of employee loyalty or similar unrelated constructs. We then performed a more targeted search of these and related journals, dissertations, and working papers using the Business Source Premier EBSCO, Social Science Research Network (SSRN), ABI/Informs, and PsychINFO global databases. To find studies that investigated issues related to customer loyalty, we used search terms such as “loyalty,” “attitudes,” “repurchase,” and related synonyms (see Table 1) across all scholarly, peer-edited marketing and management journals and dissertations that were electronically available. Finally, we inspected the reference lists of the major narrative and empirical reviews of customer loyalty and related research to identify any potentially missing studies.

The criteria for inclusion required that survey-based studies provide the exact item wording and observational studies provide exact measurement definitions, or else reference to their origins. So that we could examine the empirical implications of heterogeneity across loyalty research, we included a study if it reported a Pearson correlation coefficient or other statistical information (e.g., β, univariate F, t-statistics, χ 2) that we could use to calculate a correlation coefficient according to the formulas provided by Hunter and Schmidt (1990) or Peterson and Brown (2005).

Our sample does not include studies that rely on choice modeling to estimate the extent to which loyalty exists in a particular context. Such studies often estimate the amount of brand/retail patronage loyalty by detecting a brand–household-specific utility (e.g., Guadagni and Little 1983; Horsky et al. 2006). This stream of literature is large and well understood, but the estimates of brand loyalty are rarely tied to other theoretical constructs, which precludes them from entering our analyses.

Item-level coding

We examined the exact item wording of the measures or the citation that provided the source of the measures to categorize the type of loyalty studied by the researcher accurately. We extracted, coded, and categorized each measurement item (i.e., questions responded to by each study’s participants, objective measures), following a predefined set of rules (Kolbe and Burnett 1991). Definitions that reflect the types and elements of customer loyalty outlined in our prior theoretical review are summarized in Table 1. Guided in part by our literature review, we deconstructed each sentence and coded the language of the items in each study for common content (e.g., attitudes, behaviors, WOM, temporal orientation, target). Two coders independently evaluated each article. Fewer than 6% of the effects differed across the double coding procedures, and disagreements were resolved through discussion.

With this procedure, we could examine the bundle of items that each researcher chose to measure loyalty and thereby determine their choice of loyalty conceptualization and the extent to which they maintained content validity. The insights from this content analysis stemmed from our determination of how each sample conceptualized loyalty, using (1) only attitudinal measures, (2) only behavioral measures, (3) both attitudinal and behavioral measures as separate constructs, or (4) attitudinal and behavioral measures in the same scale as a single construct. The loyalty conceptualization thus reflects the researchers’ choice and bundling (e.g., mean average) of items. Table 7 in the Web Appendix provides a summary of the final database of 163 studies and their corresponding coding decisions.

Meta-analysis

To examine the differential effects of attitudinal and behavioral loyalty and to maintain construct validity, we used a subsample that included attitude-only and behavior-only measures (i.e., no mixed or “inclusive” loyalty measures). In addition, we included constructs only if we had at least three effects between each construct and all other constructs in the model in our effort to develop the input correlation matrix (Palmatier et al. 2006). Thus, the structural path analyses used to test H1–H5 were based on 126 studies, 151 separate samples, and 713 effects using attitude-only and behavior-only measures of loyalty.

To this subsample, we applied meta-analytic techniques to generate a correlation matrix of all constructs to use as input for the structural path models (Rubera and Kirca 2012; Zablah et al. 2012). After compiling the data, we adjusted every correlation for reliability (attenuation correction) by dividing it by the product of the square root of the reliabilities of the two constructs (Hunter and Schmidt 2004). We transformed the reliability-corrected correlations into Fisher’s z-coefficients, and then performed a random-effects meta-analysis on the Fisher z-coefficients. Following standard procedures (Shadish and Haddock 2009), we next transformed the z-scores back to r-correlations to obtain the revised, sample-weighted, reliability-adjusted correlation coefficients and 95% confidence intervals with associated t-statistics (Hedges and Olkin 1985; Zablah et al. 2012). The random-effects approach provides more realistic, less inflated estimates of average effect sizes; accounts for variability in true effect sizes across studies; and is generalizable to a population of potential studies (Raudenbush 2009). We addressed the potential problem of selective publication bias in several ways. First, we computed and report the Q statistic (d.f. = n – 1) test of homogeneity (Rosenthal 1979). Second, we tested for publication bias with funneling and trim-and-fill analyses (Homburg et al. 2012), neither of which suggested publication bias was an issue.

Structural path analysis

Meta-analytic correlations between all constructs in the model produced by the analysis and formed into a meta-analytic correlation matrix provided the input for the structural path analyses in Mplus 7.11 (Zablah et al. 2012). Our conceptual model includes commitment, trust, satisfaction, and loyalty incentives as antecedents and WOM and performance as outcomes. Model 1 tests our hypotheses in the presence of paths from all antecedents to both attitudinal and behavioral loyalties and with both loyalties linked to both outcomes, such that attitudinal loyalty is modeled as an antecedent of behavioral loyalty. All the constructs are observed variables, antecedents may covary, and we used the harmonic mean (n = 5671) across all correlations as the sample size (Rubera and Kirca 2012). Our choice to use the harmonic mean provides more conservative testing than the use of the arithmetic mean or median because it give less weight to substantially large cumulative samples sizes typical of meta-analyses, and as such is a common and strongly recommended decision for meta-analytic structural equation models (Pick and Eisend 2014). We examined within-model hypothesized differences in path coefficients by setting the corresponding paths to be equal and testing for significance using Wald Chi-square tests.

Multivariate moderation analysis

To examine the moderating effects of sources of heterogeneity, we expanded the sample that we used for the structural path analysis to include “inclusive” measures of loyalty. With this approach, we can compare the effects of attitudinal-only, behavioral-only, and various inclusive forms of customer loyalty elements with a hierarchical linear model (HLM), in which the effects are nested within studies. We performed our analysis by regressing the moderator variables on correlations (meta-regression) to account for within-study error correlation between effect sizes (Homburg et al. 2012; Rubera and Kirca 2012), coding for the specific qualities of each construct we examine. To evaluate how the effect of loyalty on WOM and performance varies across (1) loyalty aspects (H4–H7), (2) other conceptual features (H8–H10; WOM inclusive, temporal orientation, and target), and (3) study-level factors that served to test robustness (e.g., business vs. consumer markets, common method susceptibility), we included studies with any type of loyalty that reported an effect on either WOM or performance. By including all forms of loyalty in this analysis, we established a sample that is sufficiently large to support tests of all moderation effects simultaneously while also accounting for loyalty aspects. Thus, the multivariate HLM moderation analyses were based on 32 (30) studies, 41 (32) separate samples, and 68 (57) effects for WOM (performance), using all measures of loyalty.

As a replication and robustness test, we also evaluated the differences between loyalty operationalizations by dummy coding an effect as equal to 1 if the loyalty measure was a mixture of attitudes and behaviors in a single scale, consistent with 35% of our sample, and 0 if it included only attitudes or behaviors. In addition, as an alternative test of H4 and H5, we captured the moderating effect of the mixture of attitudinal and behavioral items by coding the ratio of attitudinal to total items used in the measure of loyalty (1 = all attitudinal, 0.5 = half attitudinal and half behavioral, 0 = all behavioral). These two moderators reflect extant research, where (0, 0) indicates behavioral loyalty, (0, 1) is attitudinal loyalty, and (1, ratio) serves as a loyalty proxy with varying mixtures of attitudinal and behavioral items, as is common.

We also evaluated other conceptual features by dummy coding whether a loyalty effect was WOM inclusive to test H8 (1 = measure included at least one WOM item, 0 = measure did not include any WOM items). To evaluate the moderating effect of temporal orientation, we coded the ratio of forward-looking to total loyalty items for each study effect to test H6 (1 = all forward-looking, 0.5 = half forward- and half backward-looking, 0 = all backward-looking). For the moderating effect of the target, we dummy coded the measure for each study effect for the loyalty target to test H10 (1 = individual salesperson, 0 = ambiguous, −1 = the firm).

Finally, to understand and control for potential influences of study-based features on our results, we coded and tested the moderating effects of six study characteristics (Albers et al. 2010; Homburg et al. 2012; Sethuraman et al. 2011). Specifically, we used a dummy to code whether a loyalty effect was from business or consumer markets (1 = business, 0 = ambiguous, −1 = consumer), pertained to brand loyalty (1 = “brand” was included in the construct name, context, or any items, 0 = not brand-related in any way), appeared in an unpublished journal article or dissertation (1 = unpublished or dissertation, 0 = published), and was susceptible to common method bias (1 = used a method or sample susceptible to common method bias, 0 = not susceptible to common method bias). To account for the last two study features, we used mean-centered continuous coding schemes. We evaluated any longitudinal effects according to the four-digit year of the study’s publication date; for journal quality, we coded the Journal Eigenfactor Scores for the source of each loyalty effect, such that unpublished journal and dissertation effects were assigned the sample mean value (West and Bergstrom 2013).

Results

Item-level content analysis

Table 2 shows the breakdown of the loyalty constructs in our sample, coded at the item level, which helps address our second research question. Overall, we find that though many researchers maintain attitudinal loyalty and behavioral loyalty as separate constructs (65% of studies), a substantial set combine attitudinal and behavioral items together (35% of studies). Furthermore, many researchers examine only behavioral (43% total) or attitudinal (7%) loyalty, but only 15% examine both attitudinal and behavioral loyalties in the same study, as separate constructs. Studies with behavior-only or attitude-only measures of loyalty, typically labeled “loyalty” without qualifiers, potentially misrepresent the effects of the antecedents on loyalty and loyalty’s impact on outcomes, because attitudes and behaviors have divergent effects (Baker et al. 2002; Brexendorf et al. 2010; Burton et al. 1998; Maxham and Netemeyer 2002).

Table 2 Item-level content analysis
Full size table

Among the 35% of researchers examining customer loyalty using both attitudinal and behavioral measures in the same construct, more than half (18% of total sample) operationalize loyalty with WOM as an indicator of loyalty (Table 2, Panel A). Although WOM and loyalty are related, the prevalence of WOM as a measure of customer loyalty conflicts with both theoretical (Dick and Basu 1994) and empirical (De Matos and Rossi 2008; Söderlund 2006) arguments for their separation. In addition, the use of forward- versus backward-looking loyalty measures is unevenly divided across studies, at 59% and 41%, respectively (Table 2, Panel B). Prior research shows that the abstract cognitive processes of reconstructing the past and constructing the future influence decision making (Zimbardo and Boyd 1999), so the selection of a particular temporal measurement may create variance in empirical results. The selling firm is the primary target of customer loyalty (72%), with the remainder of the research attention divided approximately equally between loyalty to a brand (15%) and to an individual salesperson (13%). However, researchers typically fail to qualify or acknowledge the loyalty target (Table 2, Panel C), even though firm- and salesperson-based loyalty reflect different decision processes that differentially affect performance (Palmatier et al. 2007).

Meta-analysis and structural path analysis (Model 1)

Table 3 contains the sample-size weighted mean meta-analytic correlations, total Ns, number of effects, coefficient t-values, 95% confidence intervals, and corresponding Q statistics among the antecedents, attitudinal and behavioral loyalty, and outcomes in our model. We used this information to calculate our structural path analysis in Mplus 7.11. The modification indices suggested modifications to the initial hypothesized model to improve fit, by allowing the conceptually related constructs of commitment, trust, and satisfaction to correlate with outcomes. Importantly, prior research supports the inclusion of these links considering the well documented influence of all three constructs on performance and word of mouth (Morgan and Hunt 1994; Palmatier et al. 2006). These changes resulted in acceptable fit statistics for Model 1: chi-square (χ 2 (2)) = 174.75, comparative fit index (CFI) = 0.99, and standardized root mean residual (SRMR) = 0.04. Figure 1 shows the results of our final structural path analysis using meta-analytic data (Model 1), with standardized beta coefficients and construct R-square values.

Table 3 Results: meta-analysis of direct effects
Full size table
Fig. 1
figure 1

Results: structural path model analysis (Model 1)

Full size image

Examining our model in relation to our third research question, we find that trust (βAtt = 0.27 vs. βBeh = 0.22, χ 2(1) = 16.41, p < .01) and satisfaction (βAtt = 0.25 vs. βBeh = 0.04, χ 2(1) = 115.06, p < .01) have stronger positive effects on attitudinal than on behavioral loyalty, in support of H1b and H1c, respectively. However, commitment is equally powerful for building both attitudinal and behavioral loyalty (βAtt = 0.34 vs. βBeh = 0.35, χ 2(1) = 2.80, p > .05), so we cannot confirm H1a. Furthermore, loyalty incentives (βAtt = −0.08 vs. βBeh = 0.01 (n.s.), χ 2(1) = 31.16, p < .01) do not have a significantly stronger positive effect on behavioral loyalty than on attitudinal loyalty, so we reject H2. Consistent with H3, attitudinal loyalty has positive, significant impact on behavioral loyalty (βAtt = 0.10, χ 2(1) = 174.75). We also find support for H4, in that the effect of attitudinal loyalty on WOM is stronger than the effect of behavioral loyalty on WOM (βAtt = 0.41 vs. βBeh = 0.19, χ 2(1) = 94.06). On the flipside, the effect of behavioral loyalty on performance is greater than that of attitudinal loyalty (βAtt = 0.02 (n.s.) vs. βBeh = 0.27, χ 2(1) = 133.98), in support of H5.

Multivariate moderation analysis (Models 2 and 3)

Table 4 shows the results of our two moderation analyses (WOM in Model 2; performance in Model 3). By using an expanded sample (all measures of loyalty linked to outcomes rather than just attitude-only and behavior-only measures) for these two paths and simultaneously controlling for other potential moderators, we retested some hypotheses to increase confidence in our findings related to our third research question, “What matters?” The positive effects of loyalty on WOM (β = 0.33, p < .05) and performance (β = −0.34, p < .05) are significantly moderated in opposite directions by the ratio of attitudinal items in the loyalty measure, corroborating our support for H4 and H5. A higher percentage of attitudinal items enhances the effect of loyalty on WOM while simultaneously suppressing its effect on performance. Similarly, the positive effects of loyalty on both WOM (β = 0.31, p < .05) and performance (β = 0.11, p < .05) are significantly moderated by measures that use a mix of attitudinal and behavioral items in the same construct rather than attitude- and behavior-only loyalty measures, in further support of H7 and H8. A non-significant intercept of 0.35 further reflects that the effect of loyalty on WOM is dependent upon its conceptualization.

Table 4 Results: multivariate meta-regression of moderation effects
Full size table

The positive effect of loyalty on performance decreases (β = −0.44, p < .05) in loyalty measures that contain at least one WOM item, which implies that WOM-inclusive loyalty measures are less predictive of performance, in support of H6. We also find support for H9a and H9b, focused on the temporal orientation of loyalty, because the effects of loyalty on WOM (β = 0.13, p < .05) and performance (β = −0.20, p < .05) are significantly moderated in opposite directions by the ratio of forward-looking items in the loyalty measure. A higher percentage of forward-looking, relative to backward-looking, loyalty items enhances the effect of loyalty on WOM while simultaneously suppressing its effect on performance. However, we did not find support for H10a and H10b, which proposed that the effect of loyalty on outcomes would be moderated by the target (individual versus firm) (WOM: β = 0.10, p > .10; performance: β = 0.03, p > .10).

With respect to other (non-hypothesized) study features, we find that the effect of loyalty on WOM is slightly stronger in business (vs. consumer) markets (β = 0.10, p < .05), and the effect of loyalty on WOM has grown stronger over time (β = 0.03, p < .05). However, the effect of loyalty on outcomes did not vary significantly for brand loyalty, unpublished sources, methods susceptible to common method bias, or journal quality. These findings increase our confidence that our sample is representative and unlikely to suffer from potential inclusion or method biases that emerge when aggregating a sample of studies for meta-analytic research (Sethuraman et al. 2011).

Post hoc analysis (Model 4)

Researchers often conceptualize customer loyalty as a “favorable correspondence between attitudes and behaviors,” stemming from an underlying motivation to maintain a relationship with a particular entity (Dick and Basu 1994, p.102; see also Brady et al. 2012; Sirdeshmukh et al. 2002). Our moderation analysis (Model 2 and 3) shows that the positive effects of loyalty on both WOM and performance are significantly and positively moderated by measures that use a mix of attitudinal and behavioral items in the same construct (WOM β = 0.31, p < .05; performance β = 0.11, p < .05); we investigated this finding with a post hoc structural path analysis that parallels Model 1. Specifically, we modeled attitudinal and behavioral loyalties as reflective indicators of a latent construct (i.e., loyalty), with all antecedents and outcomes linked to it according to the data we used in our first structural path analysis. As with Model 1, we modeled all constructs as observed variables (except for loyalty as a latent construct), allowed the antecedents to covary, and used the harmonic mean (n = 5671) across all correlations as the model’s sample size (Rubera and Kirca 2012).

Figure 2 contains the results of our post hoc structural Model 4, which provides slightly improved model fit statistics: χ 2(4) = 130.38, p < .05, CFI = 0.99, and SRMR = 0.01. To gain insight into the nature of customer loyalty, we compared the 95% confidence intervals of the standardized beta coefficients from Model 1 against our post hoc model with loyalty as a latent construct. Mirroring the results from our moderation analyses, modeling customer loyalty as a single latent construct results in stronger standardized coefficients for performance (βLoyalty = 0.51, [0.46, 0.57]) compared with either attitudinal (βAtt = 0.02, [−0.01, 0.06]) or behavioral (βBeh = 0.27, [0.24, 0.30]) loyalty alone. We find a similar pattern of results for the effect of loyalty on WOM (βLoyalty = 0.54, [0.47, 0.62]; βAtt = 0.41, [0.39, 0.43]; βBeh = 0.19, [0.17, 0.21]); the model that includes both attitudinal and behavioral elements to capture customer loyalty results in better fit and stronger effects than models that maintain either element separately. These findings support the concept that firms benefit most from “true customer loyalty,” involving a positive cognitive state (attitudinal loyalty) manifested as positive behavioral actions (behavioral loyalty) (Dick and Basu 1994; Oliver 1999; Sirdeshmukh et al. 2002).Footnote 1

Fig. 2
figure 2

Post Hoc: structural path model analysis (Model 4)

Full size image

Strengthening the loyalty framework for researchers and practitioners

From this precise inventory and examination of the primary conceptualizations of customer loyalty, we derive theoretical and practical insights to guide marketing research and practice. By synthesizing decades of research, we (1) provide a single conceptual definition of loyalty, (2) describe how researchers should approach empirical definitions to clarify crucial aspects of their treatment of loyalty and reduce measurement heterogeneity, and (3) provide exemplary loyalty measures, which we summarize in Table 5. We also provide a database of recent loyalty-related studies in marketing (Web Appendix, Table 7) and a summary of best practices in terms of study design and methodological choices (Table 6).

Table 5 Summary of key findings and implications for researchers and practitioners
Full size table
Table 6 Benchmark of methods in meta-analyses published in marketing journals since 2012a
Full size table

From a conceptual standpoint, customer loyalty is a collection of attitudes aligned with a series of purchase behaviors that systematically favor one entity over competing entities. However, empirical definitions should append a temporal aspect (backward-looking vs. forward-looking), because of its influence on how loyalty gets processed psychologically and its ultimate impact on performance outcomes. To address various combinations of empirical definitions, we offer 10 exemplary items, extracted from various studies that broadly capture elements (attitudes and behaviors) and study-specific characteristics (temporal aspect, target) that influence loyalty. We thus offer specific advice to researchers and practitioners to help them capture the customer loyalty construct more accurately and reap its benefits.

Guidance for researchers

Every loyalty study should consider four primary conceptual guidelines. First, if researchers seek to understand how antecedents create loyalty, loyalty must be measured and reported as an attitude or behavior separately, because the antecedents differentially build each element. Satisfaction (Model 1) has little effect on behavioral loyalty (β = 0.04, p < .05) but a strong effect on attitudinal loyalty (β = 0.24, p < .05), for example. Other antecedents that we did not consider in the current study plausibly should exert similarly distinct effects, and ignoring such differences could produce misleading results that depend more on the loyalty element measured than on the actual efficacy of the loyalty-building strategy. Second, if researchers seek to understand the effect of loyalty on objective performance outcomes (e.g., revenue, profit), they must measure loyalty as both an attitude and a behavior, because this composition offers the strongest effect on objective performance (β = 0.11, p < .05; Table 1, Model 3). Third, if research aims to investigate WOM outcomes, attitudinal loyalty may be the best predictor, because behaviors reflect potential constraints (e.g., size of wallet, store location) that are less important for WOM outcomes, whereas technology and other social shifts appear to enhance the effect of loyalty on WOM over time (β = 0.03, p < .05; Table 4, Model 2). Although researchers are cognizant of fusing WOM items into their measures of loyalty, they should work to avoid doing so, according to our empirical evidence this common practice undermines the linkage between loyalty and performance, as well as the theoretical reasons for the separation (de Matos and Rossi 2008; Dick and Basu 1994; Söderlund 2006). Fourth, researchers studying loyalty will be better served by employing backward-looking measures, which also help guard against potential inflation of the link between loyalty and WOM (β = 0.13, p < .05; Table 4, Model 2). In Table 6 we summarize best practices derived from recent meta-analyses published in Journal of Marketing, Journal of Marketing Research, Journal of Consumer Research, and Journal of the Academy of Marketing Science for key research decisions (sample, measurement, heterogeneity, and analyses), which we used to guide our methodological approach.

Guidance for practitioners

We offer three primary practical recommendations. First, even if assessed with just two questions (e.g., “What is your attitude about X relative to its competitors?” and “How often do you purchase from X instead of its competitors?”), customer loyalty measures need to reflect both attitudes and behaviors, because both aspects of loyalty together have a stronger effect on objective performance than either alone. The loyalties expressed by a firm’s loyal customers often differ in composition: some customers are only attitudinally loyal or only behaviorally loyal, and others are both simultaneously, and each group exerts significantly different effects on outcomes (Dick and Basu 1994). Thus, the value of loyalty for a seller depends on not only the level of loyalty but also its composition in the customer portfolio. A customer with very high attitudinal loyalty could report a high net promoter score (NPS), which is a popular metric among Fortune 500 companies for measuring loyalty (Reichheld 2003), but a lack of corresponding behavioral loyalty may reduce the effect of this NPS on objective performance outcomes (Keiningham et al. 2007). From this perspective, the ultimate effect on a specific outcome depends largely on which loyalty (attitudinal or behavioral) the firm considers, such that marketing investments may be misallocated simply because of the type of loyalty used to evaluate the investment. Similarly, assessments of customers’ future value may be biased by the loyalty metric used as an intermediate indicator of future performance.

Second, our findings suggest that customer loyalty cannot be bought using incentive strategies but can be built with relational strategies (commitment, trust, and satisfaction). The $48 billion spent on U.S. loyalty programs likely is not building “true” loyalty (Berry 2013). For an average customer, adding another loyalty card to the dozens he or she already owns may be less effective for building attitudinal and behavioral loyalties (β = −0.08 and 0.01) than building relationships through commitment (β = 0.34 to 0.35) and trust (β = 0.27 to 0.22) or improving transaction performance though satisfaction (β = 0.25 and 0.04) (Model 1).

Third, strategies for capitalizing on WOM should be separate from strategies aimed at increasing customer loyalty. Including WOM in loyalty measures detracts from the construct’s accuracy for predicting performance (β = −0.44, p < .05, Model 3). This point is especially important in light of our finding that customer loyalty has a stronger effect on WOM, but not performance, in business markets than in consumer markets, which likely reflects the greater interrelatedness in business relationships. Customers with high attitudinal loyalty likely spread WOM (β = 0.41, p < .05, Model 1) but might not contribute much to a seller’s bottom line (β = 0.02, p > .05) though their behaviors. Therefore, managers who take a portfolio approach to marketing investments—such that they recognize customer referral value as separate from customer lifetime value (Petersen et al. 2009)—should invest in attitudinal loyalty only insofar as it maximizes their overall customer portfolio lifetime value. This implication may be especially relevant for service settings and business-to-business markets, in which the effect of loyalty on WOM is much stronger.

Limitations and directions for research

Typical of meta-analyses, this study has several limitations. First, we attempted to include many loyalty constructs and samples across publication outlets, but we may have overlooked some. Second, the constructs we include and our results are limited to variables for which there exist enough data for analysis. Our framework is a summary of important loyalty-related constructs, not an exhaustive list. Our primary objective was to ascertain the implications of heterogeneity in extant loyalty conceptualizations and measurements, which required a thorough examination of loyalty’s relationship to a few key constructs rather than all constructs. Third, the heterogeneity in effect sizes that was not accounted for by our moderation analysis suggests that including other, unmeasured moderating factors might influence the reported effect sizes.

Further research thus might expand the constructs included in our customer loyalty framework to examine how they differentially affect attitudinal loyalty, behavioral loyalty, and customer loyalty. Dependence, cooperation, communication, conflict, and unfairness all might exert distinct influences on types of loyalty. Additionally, new research may also consider whether and how various types of financial outcomes (e.g., Tobin’s Q vs. ROI) are differentially impacted by attitudinal, behavioral, or combined loyalty. Clarifying these effects would provide a richer set of options for managers to tailor their marketing actions to enhance WOM and performance, given their unique circumstances.

Conclusion

Practitioners recognize the importance of repeat patronage, but “few say they have cracked the code on building long-term loyalty” (Weissenberg 2013). Despite elegant conceptualizations (Dick and Basu 1994; Oliver 1999), academics have failed to demonstrate consistently how loyalty builds and when it is most effective. It is therefore no surprise that many of the promises associated with building customer loyalty remain unrealized. We find evidence in support of the premise that this failure stems, in part, from a systematic divergence between the conceptualizations (What is customer loyalty?) and measurement (How is it measured?) of loyalty. We aggregate more than three decades of loyalty research to address this divergence and explicate when differences between theory and practice influence the strategy → loyalty → performance process (What actually matters?).

Our results offer clear evidence in support of construct divergence. Although loyalty is primarily conceptualized as the alignment of attitudes and behaviors, items used to measure loyalty often include extraneous constructs (Table 2). In addition, study-specific characteristics get incorporated into conceptualizations and/or operationalizations of loyalty, often with little or no discussion of their potential effects. We have assessed the moderating effect of several aspects of loyalty across 163 studies published in marketing journals since 1980 that measure loyalty as an attitude, a behavior, or both to determine when loyalty is most effective for predicting performance outcomes.