Studies &
Opinions
A Review of Dan
Warner’s “Traffic Targeting and Wastage”
Alex Tajirian
May 10, 2006
INTRODUCTION
The model in Warner’s article
is not well structured and lacks predictive power.
Below, I deduce the intended structure of the model,
sequentially point out some of its unclear issues, examine its predictive
power, and briefly outline an alternative prescriptive model to
capture the relationship between type of traffic sought by advertisers
and knowledge about a user’s (or searcher’s) intent.
For easy access, a copy of the article’s summary
diagram is provided in the Appendix.
THE PRESUMED GENERAL MODEL SETTING
The logic behind the model is not explicitly
stated in the article. Thus, one has to deduce it from the quote
below and the article’s general content.
The premise of the article is provided by the following
paragraph: “The enemy to advertisers and traffic affiliates alike
is traffic wastage. Traffic wastage is caused by a misalignment
between the exact intent of the user and the match with an advertiser’s
intent.”
The premise can be translated into modeling
the following hypothesis: If one can
predict the optimal type of traffic to send to advertisers, the
resulting traffic will reduce wastage.
To answer the above question, the article uses
a simple, yet potentially powerful, modeling technique, namely,
the 2x2 matrix of scenarios below.
Matrix
of Scenarios |
X2 Variable |
Low |
I |
II |
High |
III |
IV |
|
|
Low |
High |
|
|
X1 Variable |
|
The low/high states of the X1 and X2 input variables
can be equally representing binary combinations of good/bad, no/yes,
and weak/strong. <<
A SEQUENTIAL LOOK AT THE ARTICLE
The first sentence of paragraph 2 states
that the “key to traffic conversion is the balance of two things; revenue generated for
the traffic provider and sales generated for the advertiser.” First,
“traffic conversion” is not defined. Presumably it refers to conversion
of traffic to revenue. Second, the use of the term “balance” is
misleading. It suggests a tradeoff between the “two things.” However,
the key to traffic conversion - the objective of the traffic affiliate
- should be to maximize the revenue to the traffic provider, and
thus, there is no issue of tradeoff.
The second sentence of paragraph 2 states: “The
Internet traffic available to advertisers is a scarce commodity.”
Not true, unless “Internet traffic” refers to value generating traffic,
as one can have a bot that sends unlimited unwanted traffic to any
website.
The third sentence of paragraph 2 states: “There
are more advertisers wanting traffic than there is traffic available
for them to purchase.” The opposite is true, as each user can be
paid by a large number of companies to visit each of their sites.
Thus, there is more traffic available for purchase than advertisers
wanting traffic.
The fourth sentence of paragraph 2 states that “scarcity
drives a need to maximize the use of all traffic to its fullest
potential.” “Use” for what purpose? Whose “need” is being maximized?
The third paragraph, which is the focus of the article, states: “Traffic wastage
is caused by a misalignment between the exact intent of the user
and the match with an advertiser’s intent.” However, the statement
ignores the fact that wastage can also occur when the
domain name is more suitable for forwarding/leasing than parking,
as indicated by its brand-to-traffic (B/T) ratio. More details on
modeling traffic wastage are discussed below.
After setting up the premise of “traffic wastage,”
the article starts building the model. It casually refers to the
parties involved, but does not spell them out. The parties are intermediary
(such as Google), advertisers, traffic affiliate, and technology.
The model seems not to consider users as one of the relevant parties.
Moreover, it separates the intermediary and the traffic affiliate,
when they can be both viewed as an intermediary, if one poses the
model within the context of multi-platforms. Furthermore, lumping
the traffic affiliate and the owner into one party eliminates interesting
questions for the domain name owner in deciding which monetization
platform to join. In a multi-platform setting, the various “parking”
service providers would be considered representing one “parking”
platform. The model is also not clear on why technology is considered
a separate party!
The seventh paragraph sates: “It should be considered that the traffic affiliate is in control of a scarce
commodity, and has many avenues to sell this commodity at comparable
revenue returns.” First, the traffic
affiliate is leasing, not selling the commodity. Second, how does
the author know that the traffic affiliate “has many avenues to sell this commodity at comparable revenue returns”? So, how
does the traffic affiliate choose among the “comparable revenue
returns”? What precludes the traffic affiliate from having higher
revenue options?
Matching User and Advertiser Intentions
(a) The Model’s Input Variables
In the diagram of “Matching User and Advertiser
Intentions,” X1 refers to “Advertiser’s Specific Intent” and X2
refers to “User’s Specific Intent.” The axes represent the input
variables for the model.
The diagram presented in the article is not well
constructed. It is a two dimensional matrix with four axes represented
by the arrows around the box: “Advertiser’s Specific Intent,” etc.
The four axes would collapse into the standard 2x2 matrix only if
the top and right axes are perfectly negatively correlated with
the axis on the opposite side of the matrix. See the “AN ALTERNATIVE
2x2 MODEL” section below for an alternative selection of input variables.
(b) The Model’s
Output Variables
From the content of the quadrants, one can infer
that the outcome variables of interest are the “conversion rate,”
the “degree of alignment of user’s intent with advertiser’s intent,”
and pricing. However, not every quadrant has a description of all
three outcomes!
Let’s consider Pricing. This is an exogenous variable,
i.e., the traffic affiliate is a price taker (for varying levels
of traffic quality) from the competing ad agencies. Each ad agency
provides pricing that maximizes its profits, not the profits of
the traffic affiliate. Thus, even if the traffic affiliate develops
a proprietary ad agency, the traffic affiliate is still a price
taker with the potential of conflict of interest between the two
sides of the business.
It is not clear why it would be in the best interest
of all traffic affiliates to choose cost-per-acquisition (CPA) over
cost-per-click (CPC) pricing in the respective quadrants of the
diagram. There is no a priori reason to believe that a CPA is better
for the traffic affiliate than, say, a low CPC. It is an empirical
question. Thus, if the traffic affiliate has a choice between CPA
and CPC, the choice should be determined by the option that maximizes
value to the traffic affiliate. Therefore, it should not be considered an outcome and should
be left out of the diagram.
(c) Quadrant Specific Issues
Under the lower quadrant of the X2-axis, User’s
Specific Intent, the lowest possible intent for, say, Cold Calling,
would be “don’t call me.” Thus, the CPA, under this scenario, would
be zero and thus, an uninterestingly predictable outcome.
It is not obvious why under the “Untargeted” quadrant
there should be “little alignment to achieve needs.” If the advertiser
wants to target the general public, presumably because of lack of
specific information, the alignment becomes exact.
Examining the top-left and the bottom-left quadrants
exhibit inconsistencies. How can the X2-axis measure “User’s Specific
Intent,” while the top and bottom quadrants refer to the same degree
of intent, namely, “user’s intentions is likely to be undetermined”
and the bottom refers to “poorly defined user intent” [Emphasis
added]. The axis has to represent dichotomous extreme outcomes.
The top-right quadrant describes an “Exact Alignment”
when “domain (domain phrase) is matched with the same advertiser
bided phrase.” However, the article does not point out that this
is only a necessary condition, but not sufficient, as there is uncertainty
as to, for example, whether the user is looking for information
or a product (or service) related to the keyword.
(d) Modeling Requirements Issues
In the paragraph immediately following the diagram,
the model requires that the “number of advertisers on every phrase
would be deep enough that a truly active and honest auction could
occur for the commodity (traffic), enabling the traffic provider
to be paid the highest value for their goods.” What modeling problem
would a monopoly create when prices reflect the advertiser’s true
willingness to pay? A viable monopoly is bad for the traffic affiliate,
but great for the advertiser and, in the short-term, the traffic
affiliate is a price taker from the ad agency.
The next paragraph describes an imperfect system
where there are “millions of phrases that have no advertisers bidding
for them.” Does that mean that in a perfect world there cannot be
any phrases with no advertisers? What would such an imperfection
cause? In the next sentence “the volume of traffic matched to advertisers
rarely meets the volume of traffic they would like to obtain while
meeting their conversion criteria.” This sounds like excess demand,
so either the advertisers have to pay higher price for traffic or
lower their conversion criteria, i.e., their conversion criteria
may be inefficient.
In the definition of arbitrage, the mis-priced
assets bought and sold simultaneously have to be either identical
but available on different markets or a process of buying and selling
equivalent assets - an asset and a replicating portfolio, for example.
It is meaningless to simultaneously buy and sells the same asset
in the same market, as at a minimum, one would incur the cost of
the bid-ask spread. Thus, arbitrage, as defined in the article,
represents a riskless loss, not a riskless profit. Moreover, it
is not clear what is being arbitraged or what is being bought and
sold. Furthermore, the described process sounds like a pooling concept
rather than an arbitrage activity. Nevertheless, riskless arbitrage
opportunities should disappear once they become public. Thus, one
should now assume that the opportunity has been arbitraged out.
With reference to a traffic pricing system, the
article notes that the “requirement for equity is not to change
the bidding system, but instead to ensure that the advertiser pays
only the fair value for the traffic.” This is a meaningless statement;
as noted earlier, the objective of the traffic affiliate is to maximize
profit. Moreover, the fairness concept is later negated by pointing
out that the “system is flawed because it discourages traffic affiliates
from cleaning up their traffic to a quality grade higher than what
the constant’s own traffic represents.”
PREDICTIVE POWER OF THE MODEL
With low user specific intent, should the traffic
affiliate serve the user’s general interest or high intent ads?
It is not clear what the model predicts. Presumably, it is predicting
low ads, which minimize wastage. But, why not serve high target
ads if that increases the traffic affiliate’s profit? There is no
theoretic or a priori reason to believe that serving low/general
ads is optimal. Thus, the model does not provide a prescription
for the best action.
Now, suppose that the traffic affiliate notices
an increase in conversion rates. Would that be a result of better
matching? Unfortunately, the model cannot answer the question. Thus,
either the modeling technique is not the correct approach to answer
the question or, due to the complexity of the environment, we need
a different model for different sets of questions.
AN ALTERNATIVE 2x2 MODEL
In
the spirit of the article’s modeling framework, an alternative simple
prescriptive model is to assume that there are two types of information
that are available about the intent of the user – low and high –
and two types of advertising scope – general and targeted. A two-by-two
matrix can be constructed by assuming that the objective of the
traffic affiliate is to maximize traffic profit, subject to the
available inventory of domain name types. The X1-axis would represent
“Targeting Quality of Available Domain Names” and the X2-axis would
represent “Knowledge about User’s Intent.”
Under such a model, one would analyze the path
to maximizing revenue under each of the 4 quadrants. Using the model,
one can also investigate the following interesting questions:
-
How does one find the user’s Intent?
Should one use one of the proposed solutions to intent revelation,
albeit in a different context, or does parking require a different
approach?
-
What
can one do to limit the market power of the adverting agents?
CONCLUDING REMARKS
Setting
up the model within a multi-sided market framework with the correct
objective of maximizing profit given the availability of domain
name types suitable for the advertisers’ intent, yields a model
with a broader scope and predictive power.
SOURCE: The article is available online at http://dnjournal.com/articles/series/warner-traffictargeting.htm.
|