Anderl et al. (2016): helping firms reduce complexity in multichannel online data: a new
taxonomy based approach for customer journeys.
This paper proposes a new way for retailers to better understand how channels are being used
along the customer journey. Also, they explain how these funnel give inferences about the
underlying decision making process.
Introduction: More research is needed to understand how retailers can use information about
multichannel behaviour to predict purchase intentions and develop individualized marketing
strategies. This paper mainly focus on these three questions:
1. How can firms use multichannel online customer journey data to improve their
marketing strategy?
2. Which model is most useful to handle the complexity of online multichannel data?
3. Do interrelationships exist across the channels of a customer journey, revealing
progress in the purchase decision process?
In short, the results show that, due to the significant interaction effects along the customer
journey, it’s important for marketeers to not analyse their channels in isolation. This may lead
to erroneous conclusions about the effectiveness of a channel and can lead to suboptimal
decisions.
Theoretical background: the theory of choice set has been empirically validated. Whereas
multistage decisions making in online stores has been subject to extensive research, this paper
finds a new application for the theory of choice set in multichannel online marketing. The
theory of choice set states that consumers are aware of only a limited number of alternative
brands that satisfy their goals (the awareness set). A subset of the awareness set is the
consideration set, these are the brands that are actively considered.
Proposed taxonomy: marketing is traditionally initiated by a firm, but online marketing has
given us the opportunity to have customer-initiated contact. Thus, is a first step, this paper
differentiates between firm-initiated channel (FIC’s) (e.g. display). In customer-initiated
channels (CIC’s), such as SEA, potential customers trigger the communications with an active
search. Thus the origin of the contact is an important thing that should be taken into account.
Previous study show that CIC’s are more elastic. Some also find spill-over effects of CIC’s,
which reduce the costs of visiting through FIC’s.
On a second level, this paper distinguishes between CIC’s according to their brand usage. In
some CIC’s, such as direct type-ins, customers actively use the brand name to initiate contact.
With is referred to as branded CIC. Other CIC’s, like price comparisons, are generic, meaning
they don’t feature a brand name. Overall, Brand-focused keywords appear more effective than
non-brand focused keywords. There is no differentiation withing FIC’s because customers do
not actively trigger these.
Classifying channels along these two dimensions reveals several inferences about the
underlying choice set formation processes. The paper assumes that subsequent clicks in
different channels show progress, especially if a store visitor switches from FIC’s tot CIC-
channels. Related to Ailawadi & Farris who say that different attribution models can lead to
different results. They researched the importance of offline metrics with multi/omni channel.
,FIC to CIC should mean a customer is narrowing down his choice by actively searching for
new information. A switch from generic to branded CIC’s may signal the inclusion of the
advertiser’s brand in the consideration set. Generic search would raise the awareness that the
brand is relevant to the search. The customer can then become more interested in additional
information. Switching from a branded to a generic CIC may indicate that the customer is
looking for other alternatives.
The proposed taxonomy looks as
follows:
The channels were classified
along two dimensions: contact of
origin (firm or customer) and for
customers initiated contact they
differentiated between brand
usage or not.
To test the taxonomy, they test it
against alternative proposed
approaches.
,Method: the research is based on individual-level data from a German online fashion retailer.
The data is collected using cookies to better identify individual users. Cookies do have
limitations however: an inability to track across multiple devices or a bias due to cookie
deletion. Still they remain the best thing to use. For each website visit, data was collected on a
multitude of online marketing channels, sea, organic, display, email, etc. .
Model development: they used a proportional hazard model with time-varying covariates.
Because the time for the observation was fixed, in the majority of the sample, nobody
purchased something, but they might do in the future. They therefore used the time-varying
covariates. Hazard rate modelling, which calculated the time it takes for an outcome to
happens, allows for the inclusion of those possible purchases. The model of the author is
compared with other proposed models such as the AIC-model and the BIC. Log-likelihood is
used to measure compare the models of fit.
Results:
Past purchase has the strongest positive effect on purchase (Pastpurchase b =1.568),
increasing overall purchase probability by 57%. There also is a significant effect with time
(pastpurchase x time b = 0.54). Customers with prior purchases were more likely to buy after
longer journeys. Current clicks in CICs positively predicted purchases. Website visits through
email and retargeting also exerted positive effect, whereas display clicks had a negative
impact on immediate purchase probability.
Interaction effects:
- There is a strong positive interaction effect between past clicks in FIC’s and clicks
in CIC’s. If a potential customer uses a generic CIC after visiting a website trough
FIC, the probability of a purchase on that day increases by 80,7%. Switching from
a FIC to a CIC is a therefore a good measure for progress in the customer journey.
- A switch from CICbranded to CICgeneric instead reduces the probability of a
same-day purchase by 13%. It may indicate that a potential customer is in the
process of identifying a set of alternatives.
- In contrast to Rutz and Bucklin (2011), this paper found a small negative spill-over
effect from CICgeneric to CICbranded.
- Staying in the same group does not do anything
- Clicks in CICgeneric followed by a current click in FIC’s positively predicts
purchase probability. This might related to the retargeting strategy employed by
the retailer.
For every value ( x -1) * 100, where x is the value.
To evaluate the robustness of the results, the authors did a logit transformation, the dependent
variable was whether the journey ended in a conversion. All the independent referred to the
journey level. In doing so, the authors were able to determine the goodness of fit. It was found
that the results of their proportional hazard model had the best model fit compared with other
models.
Below you can find an overview of all the table and an overview of all the relevant results.
The first table shows the potential change in overall purchase probability. So these findings
can be different than what was mentioned above.
,
