MIT LIBRARIES
DUPL
3 9080 02618 0114
"'
11
m
Digitized by the Internet Archive
in
2011 with funding from
Boston Library Consortium
Member
Libraries
http://www.archive.org/details/searchobfuscatioOOelli
HB31
*?"
"DBAfEY
2?
Massachusetts Institute of Technology
Department of Economics
Working Paper Series
SEARCH, OBFUSCATION,
ON
AND
PRICE ELASTICITIES
THE INTERNET
Glenn Ellison
Sara Ellison
Working Paper 04-27
June 2004
Room
E52-251
50 Memorial Drive
Cambridge,
MA 021 42
This paper can be downloaded without c harge from the
Social Science
Research Network Paper Collection
http://ssrn.com/abstract=564742
at
MASSACHUSETTS INST
OF TECHNOLOGY
DEC
7
2004
.LIBRARIES
Search, Obfuscation, and Price Elasticities on the Internet 1
Glenn Ellison
and NBER
MIT
and
Sara Fisher Ellison
MIT
June 2004
1
We
thank Nathan Barczi, Nada Mora, Silke Januszewski, Caroline Smith and
We also thank Patrick Goebel for a valuable
tip on Internet data collection, Steve Ellison for sharing substantial industry expertise, and Drew
Fudenberg for his comments. This work was supported by NSF grants SBR-9818524 and SES0219205. The first author's work was supported by fellowships from the Center for Advanced Study
in the Behavioral Sciences and the Institute for Advanced Study. The second author's work was
supported by fellowships from the Hoover Institute and the Institute for Advanced Study.
would
like to
Andrew Sweeting
for
outstanding research assistance.
Massachusetts Institute of Technology
Department of Economics
Working Paper Series
SEARCH, OBFUSCATION,
ON
AND
PRICE ELASTICITIES
THE INTERNET
Glenn Ellison
Sara Ellison
Working Paper 04-27
June 2004
RoomE52-251
50 Memorial Drive
Cambridge,
MA 021 42
This paper can be downloaded without c harge from the
Social Science Research Network Paper Collection at
http://ssrn.com/abstract=564742
Search, Obfuscation, and Price Elasticities on the Internet 1
Glenn Ellison
and NBER
MIT
and
Sara Fisher Ellison
MIT
June 2004
1
We
thank Nathan Barczi, Nada Mora, Silke Januszewski, Caroline Smith and
We also thank Patrick Goebel for a valuable
tip on Internet data collection, Steve Ellison for sharing substantial industry expertise, and Drew
Pudenberg for his comments. This work was supported by NSF grants SBR-9818524 and SES0219205. The first author's work was supported by fellowships from the Center for Advanced Study
in the Behavioral Sciences and the Institute for Advanced Study. The second author's work was
supported by fellowships from the Hoover Institute and the Institute for Advanced Study.
would
like to
Andrew Sweeting
for
outstanding research assistance.
Abstract
We
examine the competition between a group of Internet
ronment where a price search engine plays a dominant
in this
tailers,
it
less
We
environment, the easy price search makes
though, engage in obfuscation
damaging
to firms
—resulting
in
much
website design,
demand
retail,
elasticities, frictionless
its effects
price-sensitive.
Re-
consumer search or make
less price sensitivity
Observed markups are adequate to allow
Keywords: search, obfuscation, Internet,
frustrate
and examine
an envi-
We show that for some products
demand tremendously
—practices that
discuss several models of obfuscation
empirically.
role.
retailers that operate in
on some other products.
on demand and markups
efficient online retailers to survive.
search engines, loss leaders, add-on pricing,
commerce
Massachusetts Institute of Technology
Department of Economics
Working Paper Series
SEARCH, OBFUSCATION,
ON
AND
PRICE ELASTICITIES
THE INTERNET
Glenn Ellison
Sara Ellison
Working Paper 04-27
June 2004
Room
E52-251
50 Memorial Drive
Cambridge,
MA 021 42
This paper can be downloaded without c harge from the
Social Science Research Network Paper Collection at
http://ssrn.com/abstract=564742
Search, Obfuscation,
and Price
Elasticities
on the Internet 1
Glenn Ellison
MIT and NBER
and
Sara Fisher Ellison
MIT
June 2004
1
We
would
like to
Andrew Sweeting
for
thank Nathan Barczi, Nada Mora,
outstanding research assistance.
We
Silke Januszewski, Caroline
also
thank Patrick Goebel
for
Smith and
a valuable
on Internet data collection, Steve Ellison for sharing substantial industry expertise, and Drew
Pudenberg for his comments. This work was supported by NSF grants SBR-9818524 and SES0219205. The first author's work was supported by fellowships from the Center for Advanced Study
in the Behavioral Sciences and the Institute for Advanced Study. The second author's work was
supported by fellowships from the Hoover Institute and the Institute for Advanced Study.
tip
Abstract
We
examine the competition between a group of Internet
ronment where a price search engine plays a dominant
in this
tailers,
it
less
We
environment, the easy price search makes
though, engage in obfuscation
damaging
to firms
—resulting
role.
retailers that operate in
an envi-
We show that for some products
demand tremendously
price-sensitive.
Re-
—practices that frustrate consumer search or make
in
much
discuss several models of obfuscation
less price sensitivity
and examine
its effects
on some other products.
on demand and markups
empirically Observed markups are adequate to allow efficient online retailers to survive.
Keywords: search, obfuscation, Internet,
website design,
demand
retail,
elasticities, frictionless
search engines, loss leaders, add-on pricing,
commerce
Introduction
1
When
Internet
commerce."
commerce
emerged one heard a
first
Internet technologies would
make
it
about the promise of
lot
easy for consumers to find the exact
product they wanted and to compare prices from competing
demonstration of this potential so far
is
retailers.
A
price search engines.
exact product he or she wants need only type the
"frictionless
name
1
The most
intriguing
consumer who knows the
into Dealtime,
mySimon, Shopper.
Pricewatch, Pricescan or Kelkoo to receive instantaneously and without charge a nicely
sorted
list
by dozens of
of the prices charged
e-retailers.
have tens of millions of users. 2 Internet search
industries. For example, a recent J.D.
new
is
Once a
curiosity, these sites
many
also affecting
"brick
3
The
diffusion of Internet search technologies has
so that their potential impact
The
may be
goal of this paper
is
is
much
is
been
greater than the impact
to provide
some
insights into
affected as price search technologies improve
basic messages
how
make/model
sufficiently slow,
we have
all
seen so
however,
far.
online and traditonal retail
and become more popular. Our most
that economists should think about "obfuscation" as well as search: as
search technologies improve firms will have an incentive to put
market by creating an environment
of a threat to profitability.
We
in
which price search
we develop some formal models; and we
environment
in
is
some
more
"friction"
back
in
the
difficult or at least less
explore the interaction of search, price search engines,
obfuscation with a variety of methodologies:
retail
and mortar"
Power survey indicated that about one quarter of
car buyers reported that Internet research had a "big impact" on their
choice.
now
we present some informal
and
descriptive evidence;
carry out several econometric analyses of an online
which a search engine plays a central
role.
1
For example, in 1999, The Wall Street Journal quoted an economist with Lehman Brothers who estimated that e-commerce would cut inflation by half a percentage point by 2002 because "the Internet
eliminates local monopoly power." See "New E-conomy: If E-commerce Helps Kill Inflation, Why Did
Prices Just Spike?" The Wall Street Journal, October 18, 1999, Al.
2
ComScore Media Metrix reported that Dealtime was visited by 13.8 million distinct U.S. users in August
of 2003, which is about 9% of all U.S. Internet users. CNET's 2002 annual report indicates that its shopping
sites, which include mySimon and Shopper as well as sites offering product reviews, received referral fees
from more than 100 million clickthroughs. Dealtime's visitor total is up about 300% from 2000. See White
(2000) for more on the impact of price search engines circa 2000.
The
Power survey result was announced in an October 1, 2003 press release. Another example
demand, which is surely affected by the popularity of sites like Expedia, Travelocity, and Orbitz.
ComScore Media Metrix reports that each had over 10 million unique visitors per month as of October 2003.
is
airline
J.D.
We
We
begin with a brief theoretical discussion of price search engines.
model
potential paradoxes that any
search engine could not
marginal
cost.
We
of price search
make any money
argue that
is is
if it
must
mention some
For example, a price
resolve.
down
induced retailers to compete prices
to
not hard to get around these problems and develop a
theory of price search engines as long as one allows the search engines to have sufficient
pricing instruments.
We
to create a perfectly frictionless environment.
who
why
use a simple reduced-form model to illustrate
search engines will often want
This conflicts with the wishes of
We
earn zero profits in frictionless environments.
argue that
it is
think of search frictions as being determined in a a balance-of-power
engines' efforts to reduce frictions are counteracted
by
retailers,
therefore useful to
game
in
which search
retailers' efforts to obfuscate.
technological improvements affect the equilibrium level of search frictions in such a
is
How
model
clearly ambiguous.
The theory
section concludes with a sketch of two specific models illustrating
which one can formalize the idea of obfuscating actions raising equilibrium
of these uses standard search theory ideas.
It is
is
profits.
in
One
a theory in which obfuscation makes
search more difficult and thereby reduces consumer learning.
(2003),
ways
a competitive price discrimination model.
It is
The
other, detailed in Ellison
a theory in which obfuscation
doesn't actually reduce consumer learning in equilibrium, but instead reduces the degree
to which
consumer learning hurts firm
Firms have, of course, always
profitability.
tried to
may, however, make obfuscation more important
tools
make
firms worse off
it
number
for
two reasons.
that the Internet
feel
First,
improved search
they do not engage in obfuscation. Second, the Internet
make obfuscation cheaper and
hire a substantial
We
thwart price search.
easier in a
number
of ways.
of articulate salespeople
if
One
is
traditional retailers
they want to
offer
nonstandard
may
must
sales
contracts, to use legal "bait-and-switch" techniques, or to offer personalized prices, whereas
Internet retailers can easily create automated sales pitches which cheaply implement
these strategies. 4 Another
4
The high
is
all
of
that e-retail purchases are naturally bundled with shipping,
costs of hiring articulate salespeople probably accounts for
restricted to retailers selling fairly expensive products like cars, appliances,
why such
practices are usually
and mattresses.
which allows products to be offered
for e-retailers to gather
are
to get insights into
many
different prices.
Another
is
that
it
is
easier
information about consumers and personalize prices. Each of these
consumer price search.
practices can frustrate
Most people
at
still
buying most items as they did before the Internet existed. To try
what may happen
in the future if price search engines
popular we have chosen to examine an environent that
the Pricewatch universe. Pricewatch
relatively savvy people shopping for
of small, low-overhead retailers.
is
atypical.
an Internet price search engine that
computer
They
now very
is
do become very
parts. Its universe
is
is
We
call
it
popular with
inhabited by hundreds
attract consumers largely just by keeping Pricewatch
informed about their (low) prices.
Our informal evidence
Some
obfuscation.
section describes various practices that
of these are as simple as
we think
of as forms of
making product descriptions complicated and
creating multiple versions of products so that consumers have to examine the attributes
and prices of a large number
of products to
become ubiquitous
attention to a practice that has
an
(inefficiently)
who
visit
the
site into
For example, the retailer we study
We
will refer to this practice as
from the
lists
Pricewatch universe: firms offer
CPU
two respects:
of Ellison (2003)
of this paper
is
it
them
may be
we mentioned
may
to
means
why
price series
We
loss-leader
and
the adoption of such
of obfuscation.
devoted to formal empirical analyses of the Pricewatch
demand and
within eight specific product categories: four categories of computer
CPUs.
sometimes
involve getting consumers
buy both the
discusses
address several questions via an analysis of
four categories of
it
sold for a slight profit rather than
practices (referred to there as "add-on pricing") can be a
The majority
with an attached cooling fan.
being a "loss-leader strategy" even though
classic loss-leader strategy in
The model
We
call
very low prices on Pricewatch for "bare" CPUs, and
a second physical good; and the "loss-leader"
universe.
We particularly
paying extra to get the product they really want.
to upgrade to a superior product rather than getting
at a loss.
being offered.
in the
convince consumers to instead purchase a
tries to
differs
is
low quality product at a very low price to attract customers and then try
to talk consumers
then
know what
gather data from two sources.
by using GolZilla to repeatedly conduct
We
substitution patterns
memory modules and
obtained year-long hourly
price searches
on Pricewatch
in each
of those eight product categories.
We
matched
private firm that operates several computer parts websites
from Pricewatch
Our
first
the Internet
We
of its sales
a striking confirmation of the hypothesis that price search on
is
dramatically reduce search frictions and lead to extremely elastic demand.
estimate that the firm faces a
quality
and derives most
referrals.
empirical result
may
data obtained from a single
this to sales
memory modules! These
demand
are
some
elasticity of
between -25 and -40
of the largest
demand
elasticities
for its lowest
we have seen
empirically estimated. For single product retailers they would lead to a "Bertrand paradox"
where the equilibrium price would be so low
as to prevent retailers
from covering
their fixed
costs.
Our second main
loss leaders.
We
retailer's sales of
empirical result can be regarded as a contribution to the empirics of
show that charging a low
price for a low quality product increases our
medium- and high-quality products. The reason why
this
happens
is
that
one cannot ask a search engine to find "decent- quality memory modules sold with reasonable
shipping, return, warranty and other terms."
do what
-
it
their preferences.
use Pricewatch to
memory module
of these websites to find other products that better
fit
This result (as well as our findings that website design appears to be a
determinant of the sucess of an
which has discussed
literature,
many consumers
can do - find the websites that offer the lowest prices for any
and then search within a few
critical
Hence,
e-retailer)
may
also
be of interest to the marketing
but does not contain clear empirical evidence of
loss leaders
their effectiveness.
Our
third
itability.
main
set of empirical results
Specifically,
we explore predictions
discussed in our theory section.
by making consumers
examine how
less
of the
two
it is
that obfuscation affects prof-
specific obfuscation
In the search theoretic model, obfuscation raises profits
informed.
In the competitive price-discrimination model, ob-
fuscation raises profits by creating an adverse-selection effect
attracting consumers via price cuts.
we
and we
find that there
consider
find that
We
find evidence of the
many consumers appear
nisms:
is
mechanisms
to have
makes firms
less interested
importance of both mecha-
ended up with limited information;
an important adverse selection problem that firms would want to
when contemplating
a price cut.
we examine an
Finally,
additional data source, cost data, for direct evidence that re-
tailers obfuscation strategies
have been successful
in raising
Given the extreme price
would otherwise be sustainable.
low-quality products, a naive application of single-good
is
We
rules
We
retailers.
sufficient to enable active
The most
We
find that the average
conclude that the
competition between
They use a
who conducted
of thousands of people
several discrete choice models of
and
identify price
buy from branded
retailers
sales at
elasticities is
12% markup, but
elasticities
Two
on Pricewatch
is
and demand that we know of
is
efficient retailers.
price searches for books
on Dealtime to estimate
retailers
appear to be
premia that consumers are on average willing to pay to
&
Noble, and Borders) and from retailers
The one study
of e-retail
we know
of that reports
Chevalier and Goolsbee (2003), which estimates elasticities for book
and
by inferring
after
sales
Noble.
They
collected data
on books'
sales ranks at the
two
an Amazon pricing experiment, and then estimated demand
from the data on
sales ranks.
other studies have provided some evidence on the link between Internet search and
price levels.
Brown and Goolsbee note that
appeared
1996 and term
in
falling for those
life
(which does not
Internet sites comparing
insurance premia
fell
12%
in 1996
life
and
insurance premia
8%
in 1997,
mostly
with higher predicted Internet usage. Scott Morton, Zettelmeyer and Silva-
Risso (2001, 2003) report that people
A
this
dataset containing the click sequences of tens
(Amazon, Barnes
Amazon and Barnes &
sites before, during,
for
markup on memory modules
demand. 5 They note that book
they have patronized in the past. 6
demand
demand
would suggest that
level of obfuscation
closely related study of price search engines
Brynjolfsson and Smith (2001).
differentiated
that
explain, using actual data in
actually about 12%. Pets.com could never have survived with a
should be fine for our
level
not what one should have expected in light of our findings about the
adverse selection effect in demand.
is
sensitivity of the
markup
equilibrium price-cost margins might be just 2.5% to 4%.
an example, why this
markups beyond the
itself
who
received dealer referrals from Autobytel.com
provide any price comparison capabilities) paid
disadvantage of their dataset
is
1%
or 1.5% less for
that they do not observe purchases and must use last clickthroughs
as ai proxy.
They do not report
and reported separately
price elastiticies, but they could presumably
presumably easily
for
branded and generic online bookstores.
be derived from their estimates
their
new
cars than did other customers
who bought comparable
examined price dispersion
of studies have
at e-retailers.
Finally, a
cars.
number
Krishnan
See, for example, Clay,
and Wolff (2002) and Baye, Morgan and Scholten (2003).
Theory of price search engines; search and obfuscation
2
We
why
begin this section with a high-level model that brings out a basic intuition for
search engines
may want
discuss a couple
Any model
ways
of price search engines
A
must avoid two possible contradictions. The
price search engine that caused
The second
little use.
We
then
which one could model search and obfuscation more concretely.
in
the Bertrand paradox.
would be of
to reduce search frictions and retailers to increase them.
what we
is
will call
all retailers
first is
to go out of business
the search engine revenue paradox.
If
a price search engine creates Bertrand-like competition, then retailers cannot pay the search
engine because they are making no profits, and consumers will not pay the search engine
because
if
there
is
no price dispersion they can just go directly to any
easy to avoid these paradoxes
it is
Consider a
result of its searches
and charge
The
parameters: the wholesale price
retailers
are
we
call
at
Q*(w,s) and aggregate
and that
Yahoo! Shopping,
for
game between the
can reach consumers
listed retailers
retailers acquire the
Assume that aggregate
retailer profits are
for small s
retailers
is
made
via a
as a
a referral fee of r for each sale made. 7 Suppose
which
the level of "search frictions."
differentiable
'
w
that
of firms selling a single undiffer-
search engine can monitor sales that are
that the outcome of the price competition
that
number
a large
Suppose that the only way that
price search engine.
We note
the search engine has adequate pricing instruments.
retail sector consisting of
entiated product.
monopoly
if
retailer.
r
TT
(w,s).
we have %=£ <
example, charges retailers
2%
0,
good and a parameter
0,
s
sales in this equilibrium
Assume that
^<
depends on two
these functions are
§^ <
and
of their gross revenues on any sales
3£ >
0.
8
made during
the course of a browser session in which the consumer was referred to the website by Yahoo! Shopping.
8
The one
condition where one would expect "for small s" to bind most quickly
the elasticities
we
we
feel
is
the last one, but given
comfortable assuming that the firms in our data would prefer somewhat
less
A main
observation of Diamond's (1971) search model is that profits can be discontinuous
to the monopoly price with a positive search cost. A large subsequent literature has explored
efficient search.
and jump
report,
variants of the model in which intermediate search costs lead to intermediate outcomes. See Varian (1980),
Burdett and Judd (1983) and Stahl (1989). Bakos (1997) and Janssen and Moraga (2000) examine the
effects of search costs on prices and welfare. Bakos examines a model with horizontally differentiated firms
where consumers must pay a search cost to learn a firm's location as well as its price and notes that lower
Assume that
r
7r
=
(u;,0)
s.
now
for
= D(w) and
Write
that the price search engine can costlessly choose any level of search
c for
is
=
The
0).
rQ*(c
search frictions
(i.e.,
+ r, s).
decreasing in s the optimal choice
is
The
now
Maxns
Given that Q*
sell.
the cost at which retailers acquire the good that they
problem facing the search engine
set set s
Q*(w,0)
0.
Suppose
frictions
prices converge to marginal cost as s goes to zero so
search engine's problem
is
to eliminate
is
all
then simply the standard monopoly pricing
problem,
Maxp (p -c)D(p),
=
+
The search engine
— pm —
monopoly
profit.
Retailers earn zero profits. This illustrates our first observation, that search engines
would
where p
like to
c
reduce
Our
r.
sets r
and gets the
resolution to the search engine revenue paradox
fees, firms
paradox problem
if
retailers
have fixed
natural: the search engine could
if
that
frictions that
is
make
is
that as long as search engines
pass the referral fees on to consumers and search
engines can collect their revenues from retailers. 9
costs, or,
full
frictions.
can charge per sale referral
Now
c
costs.
fixed
The model does have
the Bertrand-
In this case a couple of solutions would be
payments
to the retailers to cover their fixed
not feasible, the search engine could choose the minimal level of search
would
let
the retailers recover their fixed costs.
consider a model in which search engines and retailers must
ments to increase or decrease search
frictions.
make
costly invest-
Search frictions will then typically not be
eliminated, and the effect of technological progress on the level of search frictions
determinate. This
is
is
in-
our second observation. For example, the simplest balance-of-power
game would have the search engine choose an investment
level
x se
at cost g(x se 9) while
;
Moraga examine a Stahl-style
model with undifferentiated products and two types of consumers and emphasize that in some cases prices
may be higher with lower search costs.
As in the literature on vertical restraints, there may be many other contracts that could be used to
extract the monopoly profits. For example, the search engine could refuse to post any price below the
monopoly price and charge each retailer a fixed fee equal to its expected market share times the monopoly
search costs lead to lower prices and improved match quality. Janssen and
profit.
x r at cost h{x r ;8), resulting in the
retailers simultaneously choose
being so
— x se + x r The parameter
.
8 indexes the state of technology.
6 increase or decrease equilibrium search frictions in a
minate:
it
level of search frictions
model
Whether increases
like this is
in
obviously indeter-
depends on whether the technology aids search-improving or search-obfuscating
more.
A
couple variants of the model are worth mentioning.
engine could also charge fixed
fees.
to extract the retailers' profits.
stage prior to
when the
would be unchanged
i's
It
First,
would then charge a
As long
as the fixed
and
suppose the the search
fixed fee of
referral fees
r
7r
(s*)
—
h(x*\8)
were chosen in a
were chosen, the determination of equilibrium search frictions
—the only difference would be that the
would just help them achieve a zero
profit.
engines competing to attract consumers.
retailers' efforts at
obfuscation
Second, suppose that there were multiple search
In a
model with
differentiation
between search
engines a la Hotelling, the degree of differentiation determines the utility consumers receive.
The
In
search engines will want to provide this utility level in the most efficient
many models
them
also lead
The only
Morgan
this
would give search engines an incentive to reduce search
to charge lower referral fees than in the
full
They do not
model that one might think are
if
and
monopoly model.
listing fees
The key
trivial
—the
insight
is
fees.
They
that differences from the Bertrand
presence of an outside option for retailers
—make the standard argument that the Bertrand game has no
equilibria inapplicable. It turns out that the
model has a symmetric mixed
strategy equilibrium in which firms randomize both over whether to
to choose
frictions
consider the possibility of charging referral
nonetheless avoid the revenue paradox.
mixed strategy
possible.
models of price search engines we are aware of are those of Baye and
(2001, 2003).
and/or positive
way
list
and over the
prices
they do. Both retailers and consumers are willing to pay positive fixed fees to
the search engine.
2.1
Incomplete consumer search
The model above
treats "search frictions" very abstractly.
How
might one construct more
concrete micro models in which search frictions are determined by competing investments
by search engines and
retailers?
s
One way would be
to build
on a standard search model.
A
considers a model with two types of consumers.
and always learn the prices offered by
A
fraction 1
—
\i
fraction
fi
Stahl (1989), for example,
enjoy searching for low prices
which there are a
all retailers (of
incur a cost of c for every price quote they obtain.
number).
Consumers are oth-
downward sloping demands. Stahl notes that
erwise identical and have
finite
this
model has a
unique symmetric equilibrium: retailers randomize over prices in some interval; informed
consumers purchase from the firm offering the lowest
chase from the
always
above
is
if
first
store they visit provided that
in equilibrium)
we
stochastic
define s
=
1
The
.
— jx:
solution
is
its
price;
price
continuous in
retailer profits are lower
\i
and the other consumers purbelow some threshold (which
is
and varies
and demand
in the
manner assumed
higher (in a
is
dominance sense) when the fraction of informed consumers
is
it
higher,
first
order
and price
converges to cost in the limit as ^ goes to one. 10
Building on this model, one could provide an explicit model of search frictions by assuming that consumers are heterogeneous in their ability to digest the information presented on
a webpage.
The
se
search engine's investment x
mation and presenting
in a
it
x r would be investments
in
manner that
is
would be an investment
in collecting infor-
easier to understand. Retailers' investments
making pricing more complicated and
difficult to describe.
To
the model above exactly, one could assume that the distribution of consumer abilities
fit
such that after these investments are made, a fraction
to understand the
webpage
fully
and thereby learn
all
liq
analysis of the search
random
x T of consumers are able
firms' prices,
consumers cannot understand the price information at
listed websites sequentially (in a
+ x se —
all
is
whereas the
and must resort to
rest of the
visiting the
order) to understand each retailer's price. 11
and obfuscation game would then follow exactly
The
as described above.
In any interior equilibrium, the model will exhibit positive retail markups, price dispersion
across retailers, and limited knowledge of prices by
1
The
some consumers.
solution has similar comparative statics in the c parameter, so one could alternately take s
the measure of search frictions. There are a
number
of search
models
in
=
c as
the literature, but Stahl's model
stands out for having intuitively appealing comparative statics.
n Stahrs
(1996) paper extends the model to allow for heterogeneity in search costs
among
many
with positive search costs, so that some of the partially informed consumers will learn
others learn just one or a few.
the consumers
prices whereas
Add-ons and adverse selection
2.2
Ellison (2003) describes
retailer profits
price.
which obfuscation raises
in
even though consumers either observe or correctly anticipate every
The model
tion) involves
an alternate approach to obfuscation
two
(as
one would reinterpret
which are
retailers,
it
to
fit
retailer's
the search and obfuscation applica-
Consumers are
slightly differentiated a la Hotelling.
heterogeneous both in their preferences over the two firms and in their marginal
of income.
The
firms are
assumed to
sell
transportation costs) to each consumer.
provides only v
— w
utils.
12
a standard good that provides v utils (minus
They
are also able to sell a
damaged good that
consumers can distinguish the damaged good from
Critically,
the original, but the search engine cannot.
For example,
signed to collect and display shipping details the
standard good accompanied by
utilities
if
the search engine
damaged good could be an
fine print stating that the
is
not de-
offer to sell
the
shipping time will be one month.
Because the search engine cannot distinguish between the damaged and the standard good,
be
in equilibrium all listings will
for the
contain a separate higher price for the
damaged good. Each
retailers' website will also
undamaged goods, and consumers
will
need to incur
the cost of visiting each retailer's website to learn what these "add-on" prices are.
As
Diamond
high:
(1971), the structure of the search costs
retailers charge the
monopoly
makes the price of the add-on very
price for the incremental value of the standard
damaged good. Consumers with
rational expectations anticipate this
are fully informed: they directly observe each firm's price for the
rectly infer the prices that are not posted. Profits earned
partially or completely
The
competed away
interesting case of the
in the
undamaged good while
the firms' profits are higher
would be
if
the
damaged good did not
exist or
if
listing
See Deneckere and McAfee (1996)
for
the het-
to upgrade to
Notably,
prices than they
the search engine were able to distinguish
more on damaged goods.
10
w and
to save money.
damaged good
damaged and undamaged goods. Moreover, the equilibrium
2
cor-
for the listed goods.
some consumers
damaged good
when they compete by
in equilibrium
damaged good and
model occurs when the amount of damage
others settle for the
good over the
on the add-on may therefore be
form of lower prices
erogeneity in consumers' marginal utilities of income lead
the
and
in
profits are increasing in the
amount of damage w. An
intuition for the basic result
is
that
when the
firms
sell
damaged
goods, they create an adverse selection problem that makes firms hesitant to undercut each
other: a price cut will disproportionately attract cheapskates
at a
low price (which
One
is
sometimes below
who buy
the
damaged good
cost).
could provide a second explicit model of "search frictions" by defining s
and considering a game
in
= w
which the search engines invest in developing software that can
recognize and inform consumers about different forms of damage, thereby reducing w. while
retailers invest in inventing
w.
new and
This model would not exactly
reasons.
First,
it
creative ways to
damage products, thereby
framework described above
in the general
fit
increasing
needs some differentiation so prices will not drop
all
the
way
for
two
to cost as
search frictions vanish. Second, the higher prices that search frictions bring do not actually
reduce the quantity sold because consumers are assumed to have unit demands up to a
choke price that
is
not binding in equilibrium. The
and obfuscation game. The second
however, both
make
more
it
is
realistic
important.
and make
first
should not matter for the search
A
slight modification of the
it
fit
model would,
within the general framework: one
would just need to make the consumers' demands downward sloping. 13
3
We
The Pricewatch
study a segment of
The Pricewatch
retailers selling
universe;
e-retail largely
universe
memory
is
memory modules and CPUs
mediated by a price search engine called Pricewatch.
characterized by a large
number
of small, undifferentiated e-
upgrades, CPUs, and other computer parts.
The customers
include
hobbyists building computers, IT support people, gamers, and other relatively savvy users
upgrading old computers, and proprietors of small computer
do
little
stores.
The
retailers
tend to
or no adversiting, to have rudimentary websites, to receive no venture capital, and
We
to run efficient, profit-maximizing operations.
assume that they receive a large fraction
of their customers through Pricewatch.
One can
The
unit
use Pricewatch to locate a product in one of two ways.
demands
in Ellison's
model are primarily there
for tractability
One can
either type a
and to provide a neat contrast
by making add-on pricing completely irrelevant when there is less consumer heterogeneity. The best way to
construct a tractable model without unit demands might be to build on the model around logit demands as
in Verboven (1999) rather than around the Hotelling model.
11
technical product description, such as "Kingston
can run through a multilayered
egories, e.g., clicking
We
menu
PC2100 512MB,"
number
to select one of a
into a search box, or
of predefined product cat-
SDRAM DIMM."
on "System Memory" and then on "PC133 128MB
believe that the latter
is
much more common. Pricewatch
returns a
list
sorted from cheapest to most expensive in a twelve-listings-per-page format.
running through the
menu
that
some
The
categories inescapably
made by
product
many
as 350 listings
encompass products of varying
Note that
and
from 100
different websites.
quality, often including
products
higher and lower quality manufacturers and always including offers in which the
is
bundled with
different levels of service. Typically, the first several
would contain only low quality
buried deep within the
list,
of products
leads to predefined, not user-defined, product categories,
of these categories contain as
one
that for
list
offerings, the higher quality
of 350 products.
Figure
1
on Pricewatch are very low. For example,
in the period
we
list
products at reasonable prices
contains the
PC100 128MB memory modules from October
pages of a
first
12, 2000.
page of a typical
The low
study, generic
prices listed
memory modules
are typically offered for about one-half of the price that Dell charged for a Dell-branded
product.
People
is
who have not
seen previous studies about price dispersion might think that
remarkable that there are so
many
different prices listed for nearly identical items
Pricewatch. In comparison with previous studies, however,
prices are so close together. In the early part of our data
module
sold for about $100, the average difference
12 i/l lowest listed price
was about
$8.
we think
it is
it
on
remarkable that
when a 128MB PC100 memory
between the lowest
listed price
and the
Figure 2 illustrates the range over the course of the
year.
Prices for most computer components have declined sharply. Retail prices for
memory
modules, for example, dropped from about $100 per megabyte at the start of 1990 to about
10 cents per megabyte in September of 2003.
for
Memoiy
prices are also volatile. In Figure
example, we see that although prices declined by about
70%
over the course of the year,
there are two periods of sharp price increases. Prices rose by about
and early July 2000 and by about 25%
for
CPUs
are less volatile
and decline
in less
in a
than two weeks
more orderly
12
2,
fashion.
in
50% between
late
November 2000.
May
Prices
A fascinating aspect of the Pricewatch lists
is
the substantial turnover from day to day and even from hour to hour.
on the
of the twenty-four retailers
down one
place on the
first
list.
Some
Some
retailer
From time
list.
week or more, but there
various websites at the time. Pricewatch
is
to time one
is
no
in
may
change
up
retailers
Our impression
is
observe a firm
rigid hierarchy.
by Pricewatch and the
a database-based system which
in its database.
were setting prices
retailers
will
list
moves several other
typically
of this turnover can be attributed to the technology used
updating their own prices
average, three
websites are clearly big players that regularly occupy
a position near the top of the Pricewatch
sitting in the first position for a
On
two pages of the above-mentioned
Each price change
their prices in a given hour.
or
(which greatly aids our econometric analysis)
that
all (or
relies
on retailers'
almost
of the
all)
Pricewatch manually in the time period we study.
A
typical
has dozens or hundreds of products listed in the Pricewatch database, making
it
impractical to constantly monitor one's place in each predefined product category and the
current wholesale prices for each product. Instead, a retailer might manually examine
its
position on the most important Pricewatch categories a few times a day and might look at
current wholesale prices once or twice a day.
Our
sales
components
for their
and cost data come from a firm that operates several websites seUing computer
in the
PC133 memory modules
in
PC100 memory modules
both 128MB and 256MB; and
MHz, 700 MHz, 750 MHz, and 800 MHz.
The
in
both 128MB and 256MB:
AMD
Athlon processors
in
650
In each of these categories, our firm offers a
of products with varying quality levels, typically three.
first
part of each
speed with which the
costs
examine the pricing of and demand
will
products within eight product categories (each of which corresponds to a popular
predefined search on Pricewatch):
number
We
Pricewatch environment.
because the speed of a
often needed to
Despite
this, sales of
the
PC 100
we
PC100
study, people
vs.
PC133,
CPU. PC133 memory
the two speeds of
memory module must match
In the period
buy
description,
memory communicates with
about the same.
motherboard.
memory module
memory
refers to the
is
better and
are comparable,
the speed of a computer's
who were upgrading
CPU
old computers
and
more
memory. Hobbyists building new computers usually wanted to
13
build systems around
AMD
of the product description
can substitute two
CPUs
Athlon
that required
the capacity of the
is
128MB modules
for
one
list
memory
256MB
and some consumers would be
memory. The second part
Many consumers
megabytes.
in
modules, although some with older
256MB modules found
motherboards could not use the "high density"
Pricewatch
PC 133
at the top of the
The 256MB modules
slot constrained.
are
about twice as expensive. Typically, one spends about $100 on memory when building an
inexpensive computer. In the
of the
demand
at Pricewatch
summer
was
128MB
for
when 256MB modules
of 2000
cost about $250
most
when 256MB
modules. In the spring of 2001
modules cost about $60, the two were about equally popular.
The
retailer that
provided us with sales data
Pricewatch memory category on each of
PC100 128MB memory modules,
well as
by contract terms. Figure 3
is
illustrates
a
sites often
is
a similar quality choice
AMD
presented to
is
websites:
many memory modules
Athlon processors. Since an
are
AMD Athlon
from about $120
for a
500
not yet widely available).
MHz
A
Athlon purchased from any other website. The
CPUs were hobbyists
new computers. CPUs come
in
Athlon to over $1000
typical
way
to
computer
stores)
In early June 2000 prices ranged
for a
make the speed
1000
vs.
MHz
Athlon (which was
price tradeoff
would be to
computer on the CPU. Pricewatch's predefined
for a
categories recognize that one should
compare
will
(or proprietors of small
many clock speeds.
spend one-fourth of one's total budget
We
as
similar to shopping for mattresses.
identical to a bare
primary purchasers of Athlon
speed.
and the board
a branded product, a "bare" Athlon with a given clock speed purchased from
any website should be
building
how
DRAM
three different
A's design. Making quality comparisons within a
other set of products we study are
processor
by quality of the
sells
contain minimal technical specifications. In this regard shopping
memory module can be
The
site
each website
e.g.,
much easier than making comparisons across
unbranded and
for
websites,
differentiated
consumers on a website that copied
website
its
three different products within each
sells
prices for processors with the
examine prices and demand within four
of the categories that
same
clock
were most
popular in the summer of 2000: the 650 Mhz, 700 MHz, 750 MHz, and 800 MHz. Typical
prices for these
CPUs
ranged from $140
for the
Athlon.
1
I
650
MHz
Athlon to $280
for
the 800
MHz
Despite the fact that the 650
MHz
650
category
to be
fail
different levels of service.
MHz
Athlon
homogeneous
Second,
CPUs
CPU.
the
Retailers
may
First, different retailers offer
are often purchased together with a retailer-
$5 and an experienced installer can attach
attachment can be tricky
two reasons.
for
One can buy a decent
installed fan/heatsink.
a branded product, offerings within the
is
quality fan/heatsink combination for about
it
to the
CPU
in less
The
than a minute.
a novice, though, and entails a nontrivial chance of ruining
for
offer to
attach any of several fans, and each CPU-fan combination
should be viewed as a distinct product within the same Pricewatch category. For example,
one of the websites from which we have data typically
a bare
category:
CPU,
a
CPU
offers three
products within each
attached to an AMD-approved fan, and a
CPU
together
with a premium-quality fan (and also bundled with better service terms and an extended
warranty).
To
reiterate, Pricewatch's categories did not distinguish
and without fans
(or
same Pricewatch
the
with better or worse fans) so
,
Over the
One
its
lists
not
last
would appear
in
few years Pricewatch has made a number of enhancements to combat obfus-
of the
most
remain commonplace.
visible search-and-obfuscation battles
early days Pricewatch did not collect information
was fought over shipping
uncommon
for firms to list
with a two-pronged approach:
it
fee
when they went
mandated that
all
no greater than a Pricewatch-set amount ($11
column to
its listing
charges. 14
Many
retailers
adopted an $11 shipping
uncertainty about the cost of
Our empirical work
is
wary of
UPS
was
to check out. Pricewatch fought this
firms offer
for
UPS
ground shipping
memory modules); and
pages that either displayed the shipping charge
or a warning that customers should be
it
its
memory module and inform consumers
a price of $1 for a
and handling"
costs.
on shipping costs and sorted
purely on the basis of the item price. Shipping charges grew to the point that
of a $40 "shipping
it)
with
list.
cation. Practices that frustrate search nonetheless
fee
of these products
CPUs
Observations of obfuscation
4
In
all
between
stores that
fee for
(if
it
for
a
added a
the retailer reported
do not report their shipping
memory modules
in response, but
ground shipping was not completely eliminated because a
based on data from the period when these policies were
15
in effect.
number
of retailers left the
charges.
The meaning
company
explicity stated
column blank
"UPS ground
of
on
or (more
commonly) reported a range
shipping" was also subject to manipulation: one
website that items ordered with the standard
its
of shipping
UPS ground
shipping were given lower priority for packing than other orders and might take two weeks to
arrive.
More
and switched
recently,
Pricewatch mandated that retailers provide
to sorting low-price lists based
on the maximum shipping charge
desired, although
it
shipping: customers
if
with shipping charges
it
on shipping-inclusive prices (which are based
a range
is
This appears to have worked as
given).
only provides the desired information for customers
who wish
to upgrade to 3rd-, 2nd-, or next-day air
who
prefer
ground
must look through
the retailers' websites manually.
One
of the potential
models of obfuscation we discussed involved
firms' trying to increase
customers' inspection costs and/or to reduce the fraction of customers
the firm on the top of the search engine's
list
The most
bundling low-quality goods with unattractive contractual terms,
it
fee
on
all
print.
Another
is
making advertised
listed
available
listed prices.
like
effective
We
seems to be
providing no warranty
on Pricewatch without reading the
prices difficult to find. In 2001
time to find prices listed on Pricewatch on several
found the
buy from
Given the variety of terms we observed,
returns.
would seem unwise to purchase a product
will
without doing any firm-by-firm searching.
observed several practices that might serve this purpose.
and charging a 20% restocking
who
it
retailers' sites. In
fine
took us quite a bit of
a few cases, we never
Several other firms were explicit that Pricewatch prices were only
on telephone orders. Given that phone
calls are
more
costly for the retailers,
we
can only assume that firms either wanted people to waste time on hold or intended to make
them
sit
example,
through sales pitches. Pricewatch has fought these practices in several ways. For
it
added a "buy now" button, which
(at least in theory) takes
customers directly
to the advertised product.
The second obfuscation mechanism we
discussed
is
the adoption of a "loss-leader" or
"add-on" pricing scheme: damaged goods are listed on the search engine at low prices and
websites are designed to convince customers attracted by the low prices to upgrade to a
higher quality product. Such practices are
example. Customers
who
now ubiquitous on
try to order a generic
16
Pricwatch. Figure 3
memory module from Buyaib.com
is
one
at the
price advertised on Pricewatch.com are directed to this page.
which the low priced product
in
markups). Figure 4
from Tufshop.com
is
is
is
inferior to other
A
another example.
It illustrates
products the company
consumer who
several
sells (at
tries to order a generic
ways
higher
module
taken to this page, from which a number of complementary products,
The
upgrades, and services.
figure presents the
webpage
as
it
To avoid
initially appears.
purchasing the various add-ons the consumer must read through the various options and
unclick several boxes.
After completing this page, a Tufshop.com customer
another on which he or she must choose from a long
UPS
paying $15.91 extra to upgrade from
UPS
to
2-day,
Our impression
is
categories
all
on the
of the listings
much more
first
likely to
is
memory
lists
In Pricewatch's
We
are told that
CPU
most of
are very low quality products
have problems than are other modules that can sometimes be
We know
that the wholesale price
occasionally so small as to induce the retailer from which
"medium
to ship
low quality.
few pages were bare CPUs.
purchased wholesale for just one or two dollars more.
difference
upgrade
3-day, $30.96 extra to
next day 15
inefficiently
the "generic modules" at the top of Pricewatch's
that are
UPS
that the practices are also consistent with the add-on pricing model
terms of the low-priced goods being of
in
UPS
and $45.96 extra to upgrade to
taken to
of shipping options. These include
list
ground to
is
quality" generic modules to customers
(without telling the customers) because
it
felt
who
we
got our data
ordered low quality modules
the time cost and hassle of dealing with
returns was not worth the cost savings.
Obfuscation could presumably take
our theory section.
Presumably,
than
it is
profits.
fusing.
15
16
this
One
many forms
would involve
either tricking
worth to them or altering their
is
that
consumers into paying more
utility functions in
many Pricewatch
For example, whereas several
The incremental
we
outlined in
that firms could try to confuse boundedly-rational consumers.
is
Our impression
in addition to those
sites will
a
way
for
a product
that raises equilibrium
e-retailers sites' are intentionally con-
provide consumers with product comparison
costs to Tufshop of the upgraded delivery
methods are about
$4, $6,
and
$20.
16
Following our work on obfuscation, Gabaix and Laibson (2004) have developed a model of consumer
confusion. They assume confused consumers behave exactly as if they had stronger idiosyncratic preferences
for
one firm relative to the other
pricing
and incentives to invest
in differentiation.
in
a standard horizontal differentiation model.
in confusing
consumers
is
identical to
Hence, an analysis of
an analysis of firms that can invest
Welfare consequences are different, however, because the idiosyncratic factors
A over product B are not preferences that contribute to utility.
consumers prefer product
17
that,
make
lists like
that in Figure
description of
3,
we did not
see
any that augmented such a comparison with a
what "CAS latency" means
should or shouldn't care about
to help consumers think about whether they
it.
As mentioned above, Pricewatch
enter their prices into a database.
is
a database-based
facility that requires
that retailers
This scheme for running a price search website seems
to be gaining relative to the alternate technology of using shopbots to maintain a price
database or to search for prices in realtime whenever a consumer requests them.
One
explanation for this must be that the shopbot approach can result in substantial delays
while the site searches.
Another may be that the shopbot approach may be even more
prone to obfuscation. In 2001,
had a much
for
easier time organizing products
searched sites hosted by Yahoo.
through Yahoo! Shopping
at a
Yahoo
than a general search engine because
collected a royalty on
all sales
,
box and sorted on
how
price, the five lowest listed prices
clicks over to the retailer
above the Pricewatch
results
when we typed "128MB PC100
price).
price or items
for
listing
and ordering
into the search
were from merchants who had figured
is
zero even though a
can easily see the price (and see that
The next hundred
were also either products
or so cheapest items
it is
50-100%
on Yahoo's search
which Yahoo's search engine had misinterpreted the
which were inexpensive because they were not the desired product. 1
5
Data
Our
price data were
only
made by merchants
SDRAM DIMM"
to get Yahoo! Shopping's search engine to think the price
human who
it
a monthly fee based on the number of items offered
(as well as
Yahoo! Store) so there must have been some standardization of
mechanics. Nonetheless,
out
example, Yahoo! Shopping search engine should have
'
downloaded from Pricewatch.com. They contain information on the
twelve (sometimes twenty-four) lowest price offerings within each of the eight predefined
11
it was possible to use Yahoo!
Shopping's search engine to get a well-sorted list and find low
one was experienced with it. In 2003 Yahoo! Shopping's search engine conducts a broad search
not restricted to Yahoo! Shops and affiliates. The list returned in response to a corresponding search no
longer begins with a large number of sites that Yahoo thinks have zero prices, but it does begin with a large
number of sites for which it has incorrectly parsed the HTML and inferred the price to be less than it truly
is.
Most of these mistakes appear to be accidental rather than due to a plan by the retailers to fool the
Yahoo! Shopping search engine.
In 2001
prices
if
18
categories mentioned above.
2000 to
The
May
18
These data were collected
at hourly frequency
2001. 19 Recall that Pricewatch's predefined categories are
we downloaded
vast majority of the products for which
prices are
from
somewhat
May
coarse.
what we could
call
"low-quality" products.
In addition to the price data for these "low-quality" products,
quantity data from one Internet retailer.
for all
products that
products
in
fit
The data contain the
The
each category.
will
simply
The raw data
call
May
of 2000.
and
and quantities sold
firm operates several different (but similar) websites,
20
We
which
use data from three websites,
A, B, and C.
are at the level of the individual order and indicate from which website the
customer made the purchase.
in
prices
price
within the eight Pricewatch categories, often three different quality
typically have different prices for the products studied.
which we
we obtained
We
We currently have
approximately a year of sales data starting
aggregate the individual orders to produce daily sales totals for each
product at each website. 21 Our primary price variables are the average transaction prices
for sales of a given
product on a given day. 22
each website on Pricewatch's price ranked
The same
for
We
also record the daily average position of
list.
Internet retailer also provided us with data on wholesale acquisition costs
each product.
The
firm maintains very low inventories.
Some products
are purchased
from wholesalers every few days or once a week. Others are purchased almost every day.
The
firm often has a "negative" inventory of the most popular
them each afternoon
to
fill
memory modules, buying
the backlog of retail orders taken that morning and the previous
18
We collected the twenty four lowest prices for the 128MB PC100 and 128MB PC133 memory modules
and the twelve lowest prices in each of the other six categories.
1
We
used GolZilla to carry out the downloads.
Among
the motivations for having several websites are that different websites
looks and consumers
may have
heterogeneous reactions, that
it
may be
allows the websites to be
given different
more
specialized
(which seems to be attractive to some consumers), that it facilitates experimentation, that it may help
promote private-label branded products, and that it lets the firm occupy multiple places on the Pricewatch
screen.
"1
Here, a "product" includes also the quality level, e.g., a high-quality 128MB PC100 module.
"Transaction prices are unavailable for products which have zero sales on a given day. If low-quality
products ever have zero sales, then we fill the prices in using the listed prices from Pricewatch, since these
prices are quite volatile. Prices for the other products change less often and can be filled in fairly completely
by just assuming that prices are constant in the gaps between times when we observe identical transaction
price observations. We also fill in a few other values by looking at when low-quality prices and prices at
other websites operated by the firm changed.
19
evening.
The data provided
particular product.
Our
analysis of
to us contain wholesale prices
23
memory module demand
uses data from websites
websites have identical product lineups: they
ule category,
Summary
which we
of the
is
number
presumably
A
and B. These two
memory mod-
three products within each
medium-, and the high-quality module. Our
contains between 575 and 683 observations in each category. 24
each of the four categories are presented in Table
statistics for
data are at the
sell
refer to as the low-, the
memory modules
dataset on
whenever the firm purchased a
level of the website-day, so the
LowestPrice
of observations.
for a low-quality
number
is
of days covered
is
Note that the
1.
approximately half
the lowest price listed on Pricewatch (which
memory module.) 25 Rangel-12
the difference between
is
the twelfth lowest listed price and the lowest listed price. Note that the price distribution
PLow, PMid and PHi
fairly tight.
ules at the
sorted
list
two websites.
are the prices for the three qualities of
PLowRank
is
rank of the website's
first
is
memory mod-
entry in Pricewatch's
of prices within the category in question. 26 This variable turns out to allow us to
predict sales
much
better than
we can with simple functions
of the cardinal price variables.
Note that the websites we study are consistently near the top of the Pricewatch
all
four categories.
quantities of each quality of
module
sold
the
10%
of the average lowest
QLow, QMid and QHi
are the average daily
average prices for the low-quality products they
price on Pricewatch in
list:
sell
are within
by each website. The majority
of the sales are the
low-quality modules.
We
have not broken the summary
usually lower than website B's, but there
23
The
24
Data
We
which we were missing data. The
use mid-March rather than
The Pricewatch data
hours using weights that
26
We
no
CPU
only
know
a
May
256MB
as the
are hourly.
reflect the
sites'
gaps
left
website.
rigid relationship.
CPU
fans.
For example, in the
128MB
This does not cause any substantial incom-
we used
to collect data and missing data
Figure 2 to indicate the frequency with
most of the last six weeks, so we chose to
in the price curves in
prices are missing for
end of the
256MB
samples.
Daily variables are constructed by taking a weighted average across
average hourly sales volumes of the websites we study.
if it is among the 24 lowest priced websites for
Pricewatch rank
or the 12 lowest priced websites for
Website A's prices are
fans rarely change.
are occasionally missing due to failures of the program
in the files the firm provided.
25
is
only products purchased infrequently are
pleteness of the data because prices of
down by
statistics
256MB modules
128MB modules
When a site's price
we set PLowRank equal
or Athlon processors.
is
too high
to one one
and it does not appear on the lists we downloaded from Pricewatch
more than the number of firms appearing on the lists we collected, i.e. 25 for 128MB modules and 13 for the
other products. In the 128MB category this happens for fewer than 1% of the observations. In the 256MB
category this happens for 3% of the site A observations and 14% of the site B observations.
20
PC 100 memory
same on 27
module category, website A's low-quality price
and higher on 60 days. In
days,
quite small, e.g.
they are two dollars or
would expect,
70%
Site
A
Our
analysis of
and C. Athlon
demand
sales
more than 70%
less
of the time.
the median
larger:
also has higher sales volumes. For example,
128MB PC100 memory
of the sales of low-quality
drop
patterns for
off
lower on 251 days, the
this category the price differences are usually
somewhat
categories the price differences are
is
256MB
In the
As one
dollars.
is five
responsible for about
it is
modules.
AMD Athlon processors uses data from websites B
sharply in the
of 2000 following
fall
AMD's
introduction of the
improved Athlon Thunderbird. Accordingly, we analyze data from a shorter period,
November
of 2000.
An
additional factor complicating the analysis
not have identical product lineups. Site
category: the "low quality" product
briefly
with selling a
CPU
is
C
that sites
B
CPU
fan. Site
and
B
offers
with a dual fan. The fact that
forces us to analyze site B's sales
of this
a
CPU
and
site
two products
in each category: the bare
CPU
sites
B and C
have separate
C's sales in separate regressions.
be identical across clock speeds except
speed dummies. Table 2 contains summary
as
it is
used in the regressions. The
the second to
site
C's sales.
The
statistics for the
are averages across clock speeds.
650MHz Athlon
first
to
The combination
PHi
in the website
B
is
the clock speed-day Hence, each
pay about $39 more to get a
unit sales of
The minimum
end of the sample. The
reflect that
will also
who buy upgraded
PLow
prices reported
on
site
B
is
B would
prices for
a
little
is
PLow
on average need
CPUs
are lower than
reduce the precision of our estimates.
is
what
price of $302.45
The mean
attached. Unit sales of
products
21
mean
maximum
customers of website
CPU with a fan
memory modules. This
fraction of consumers
price of $105.25 for
cost near the start of the sample.
data
for additive clock-
panel refers to the data on site B's sales and
unit of observation
cost near the
what a bare 800MHz Athlon
and
sets of offerings
Athlon processor data broken
panel summarizes a dataset with four observations per day, and the
a bare
is
and the shorter time period led us to pool data from the four clock-speed categories
restrict the coefficients to
down
do
with a nonapproved fan instead); the "medium quality" product
with a premium quality dual
CPU
C
(although the firm experimented
has an AMD-approved fan/heatsink attached; and the "high quality" product
and the
and
to
products within each processor-speed
offers three
usually a bare
is
May
higher than in the
The
memory
data:
39%
medium-
of the customers of website
of the customers of website
C buy
a
or high-quality product.
Demand
6
B and 56%
In this section
patterns
we estimate the demand
how consumers
elastiticies
Pricewatch e-retailers face and examine
substitute between low-, medium-,
and high-quality products.
We
do this
both to provide new descriptive evidence on the functioning of online markets and to provide
empirical evidence on the theories of obfuscation discussed above.
Methodology
6.1
Our primary
depend on
interest
its
is
demand estimation
for
in
how website w's sales
prices for the three products
suppose that within each product category
from website
w
on day
of low-,
and on
c,
its
medium-, and high-quality products
competitors' prices. Specifically,
we
the quantity of quality q products purchased
t is
tywcqt
—
c
i
with
=
XwctPcq
Pcqo
+ /3cq ilog(l + P Low Rankwct + /3cq2 log(PMidwct + /3cq3 log(PHi wct
)
)
)
12
+/3cq 4log(LowestPrice ct )
In this equation
watch's
list
capture
how
Pricewatch.
offerings.
27
PLowRankwc
t
is
+ pcq5 Weekend + Pcq eSiteB w +
t
^
f3 cqe+s
TimeTrendst
the rank of website w's low-quality offering on Price-
of the lowest prices in category
c.
It is
the primary variable that
we use
to
website w's low-quality price differs from those of the other firms listed on
PMid wct
and
PHi wct
The TimeTrend
that changes every 30 days.
are the prices of website w's
medium- and high-quality
variables allow for a piecewise linear time trend with a slope
We
as an additional control variable.
think of the fourth price variable, LowestPrice c t, mainly
28
2
Medium- and high-quality cannot be defined consistently across retailers. Accordingly, we have no data
on competitors' prices for comparable products to include in the regressions.
We do not treat the variable as providing an estimate of the elasticity of the aggregate demand at
all Pricewatch retailers as a function of their price because we lack data on prices at traditional retailers.
Interpretation would be even more difficult in the medium- and high-quality demand regressions because
the coefficient on LowestPrice will also reflect how decisions depend on medium- and high-quality prices
'
relative to low-quality prices.
22
.
We
estimate the
demand equations
assume that the error term u wcqt
via
satisfies
GMM.
Specifically, for
E{e Uwc i \Xwct )
t
=
most of our estimates we
so that
1
we can estimate the
models using the moment condition
E(Q wcqt e- x ^ t0«> - l\Xwct ) =
These estimates are done separately
for
each product category and each quality
level.
Stan-
dard errors were computed with a Newey-West style approach.
The
on the logs of the prices of the medium- and high-quality modules can
coefficients
be directly interpreted as
elasticities.
a website's low-quality price,
we
compute estimated
when
of
PLowRank
elasticities
To construct
treat log(l
all
+ PLowRank)
variables are at their
with respect to a change in
between the twelve lowest
elasticities of
PLow
demand with
respect to
as a continuous variable
means by
and
setting the derivative
equal to the inverse of the average distance
prices.
Although we
This approach presumes that the price variables are not endogenous.
present estimates in a later section that do not assume exogeneity of prices, these non-
instrumented estimates are our preferred ones for two reasons. First and most importantly,
we think that
little
of the variation in our e-retailers' prices
about demand. The firm appears to have
or times
little
is
due to
its
information about whether particular days
might be good ones to be higher or lower on Pricewatch's
changes in our
e-retailer's ranks, in fact, are
using information
Most
lists.
of the
not due to any decisions: the e-retailer has
limited attention to devote to each of the dozens of Pricewatch
lists
on which
its
products
appear, and most rank changes are due to other firms changing their prices.
The person who
sets prices noted in discussions
his low-quality prices for
the Pricewatch
lists
he typically leaves
and may
it.
if
Less attention
First,
raise or lower a price
if
a few times a day he checks some of
a rank has drifted too
in response to a wholesale price change.
paid to medium- and high-quality prices. These are typically
several weeks at a time.
23
lists
Third, he occasionally
order-processing or shipping employees have failed to show
is
from where
-
far
Second, once a day or so he receives updated wholesale price
and may change a price
raises prices
one of three reasons.
with us that he would typically change
up
for
work.
left fixed for
Second, given our assumed functional form for demand, economic theory implies that
prices should not be endogenous. In our
multiplicatively. In such
marginal costs
maximize
is
profits
before setting
demand model, shocks
increase or decrease
demand
an environment, the profit-maximizing price of a firm with constant
independent of the realization of the demand shock. Hence, prices set to
would not be endogenous even
if
the firm observed the
demand shock
its price.
Despite these arguments, we do also estimate our model using two distinct sets of
instruments for
PLowRank, PMid, and PHi. These
estimation of the
moment
estimates are obtained from a
condition
E(Q wcqt e- x^«> - l\Wwct =
)
where
W
estimation in section
modules. The
the
demand
first
for
128MB PC100
column of the table contains estimates of the demand equation
128MB PC100
for low-quality
discuss this
demand
Table 3 presents estimates of demand from one of our product categories:
memory
We
6.5.
Basic results on
6.2
0,
a vector containing the instruments instead of the prices.
is
wct
GMM
modules. The second and third columns contain estimates of
medium- and high-quality 128MB PC100 modules,
respectively. Coefficient
estimates are presented with f-statistics in parentheses below them.
Our
faces
is
first
main empirical
result
is
extremely price sensitive. This
that the
is
demand
for low-quality
due to a combination of two
modules a website
factors. First, a price
change of a few dollars can move a website up or down several places on the Pricewatch
list.
Second, most people buying low-quality
first
few firms on the Pricewatch
variable in the
a website's
sales
first
list.
The
memory modules buy them from one
coefficient estimate
column implies that moving from from
sales of low-quality
first
on the log(l + PLowRank)
to second on the
modules by 45% and dropping from
by 87%. 29 This feature of demand
is
of the
first
list
to seventh reduces
—we get a
remarkably clear in the data
reduces
i-statistic
of 15.1 in a regression with only 683 observations.
A move
from
first
to second increases log(l
thereby reduces log-quantity by 1.48 * 0.41
~
+ PLowRank)
0.6.
24
from from log(2)
w
0.69 to log(3)
~
1.1,
Table 4 presents demand
The upper
elasticities.
left
number
indicates that the 1.48 coefficient estimate in the low-quality
in the
left
panel
128MB PC100 demand
equa-
tion corresponds to an own-price elasticity of -25.0. In each of the other
categories the corresponding elasticity
128MB PC133
the
even
is
category to -41.4 in the
larger.
The
256MB PC133
upper
memory module
estimates range from -33.0 in
category. These elasticities are a
powerful illustration of the potential that price search has to create a real Bertrand paradox.
The
familiar Lerner index formula for equilibrium
MC)/P =
[P —
retailers,
—1/t], implies that
markups over marginal
their
retailers
A
would
find
it difficult
cost
if
quality
memory
is
is
available,
sumers are doing
list
non-advertising
that low quality
memory
is
means that
is
memory and
for the lowest price at
more medium-quality memory
it sells
memory
if its
are subsititutes. Apparently, however,
the more such consumers
for (low-quality)
attracts.
it
price for low-
what many con-
literature
common
in
first
to seventh
first coefficient in
sales
memory
very large.
on the Pricewatch
as well.
list for
The
-0.41 coefficient
low-quality
memory
by 43%.
marketing to talk about
on the effectiveness of
is
on the
the third column indicates that
to seventh on the Pricewatch
first
would reduce high-quality memory
it is
loss leader effect
is
fur-
reduces a website's sales of medium-quality
the loss-leader benefits include sales of high-quality
moving from
The
moving from
128MB PC100 memory
128MB PC100 memory by 65%. The
Although
+
which
through to those websites to examine their offerings
clicking
-0.76 coefficient estimate indicates that
indicates that
effec-
using Pricewatch to find a set of websites offering the lowest prices
memory,
list,
an
controlling for the prices
and then sometimes purchasing a higher quality product. The higher a firm
Pricewatch
The
efficient,
Ordinarily one would expect the opposite relationship because
lower.
low- and medium-quality
ther,
is
negative and highly significant. This
memory
for low-quality
Even
to survive with such markups.
a website charges for medium- and high-quality
low-quality
only low-quality products,
In the second column of the table, note that the coefficient on log(l
tive loss leader.
is
retailers sold
would be 2.5% to 4%.
second striking empirical result in Table 3
P Low Rank)
markups with competing single-product
loss leaders
loss leaders, the empirical
marketing
has produced mixed results (Walters, 1988;
Walters and McKenzie, 1988; Chevalier, Kashyap and Rossi, 2000).
25
We
are not aware of
any evidence nearly as clear our
A
ative
third noteworthy result
and
is
results.
the fact that the coefficients on the site
They
significant in all three regressions.
B
the same prices, site
cessful at selling high-quality
of the importance
and
owned by the same
memory modules. We
They were designed by the same
are neg-
the two sites charge
It is
especially less suc-
be a striking demonstration
find this to
Website
A
people.
They share the same
difficulty of effective website design.
firm.
when
memory modules.
substantially fewer
sells
indicate that
B dummy
and website B are
telephone operators and packing employees and hence provide exactly the same level of service quality.
it
had
A
few attributes should make website
B more
attractive to
slightly lower shipping charges for part of the sample;
additional products that
may be
it
offers a
prefer to
The
memory. One hypothesis the firm suggested
buy memory from a website
(like site
The
significant.
The other
The one way
log(l
in
+ PLowRank),
which the results
We
with our estimation strategy
128MB PC100
for the
is
is
sold toward the
result,
varies
This problem
full
tables of
is
memory
variables are usually highly
category are unusual
is
that
are precisely estimated.
high-quality
A
memory
particularly severe in the
is
reduced by the fact that most of the
and
It is larger in
some
memory
demand
categories in Table 4, but to save
estimates.
The
elasticity tables reveal that
large loss-leader benefits consistently across cate-
SiteB
finding of a negative coefficient on
somewhat.
memory.
they are highly collinear with the flexible time
report elasticity matrices for the other
The
some customers may
end of the data period.
find large own-price elasticities
gories.
had a
not as effective
find in the other
medium- and
that prices for
categories because the effective sample size
space we have not included
we
what we
Weekend, and SiteB
trends that we have included in the model.
memory
that
specializes in
medium- and high-quality memory
modules change infrequently. As a
256MB
is
it is
it
variables are usually insignificant.
the own-price elasticities of
difficulty
to us
A) that
significance levels in this table are similar to
categories.
broader variety of
purchased at the same time; and at the time
higher customer feedback rating at ResellerRatings.com. Nonetheless,
at selling
some customers:
of the
256MB
Table 8 presents estimates of the demand
for
26
is
consistent, although the
magnitude
regressions.
AMD
Athlon
CPUs
sold through website
C.
The
first
column
gives
The second column
CPU
with an
parameter estimates of the demand for low quality (bare) CPUs.
reports on the
AMD- approved
medium-, and high-quality
CPU
from
in
third
column reports on demand
is
quite similar to the
a number of ways.
demand
The demand
to seventh
is
predicted to lower
when the
many more
sells
of the medium-quality
CPU
puts
it
It
abandoned
LowHasFan
(Moving
to an
a significant loss-leader
CPU/fan combinations. For
this
list it
part of our sample the
experiment after just a few weeks because too
variable in the third
column indicates that there
that the alternate product lineup led to
As
list.
near the top of the Pricewatch
were buying the low-quality product instead of upgrading. The
fans.
for the low-quality
CPU with a non- AMD-approved fan rather than a bare CPU as its lowest-quality
firm sold a
product.
firm's price for a bare
is
for
medium-,
demand by about 80%. This corresponds
own-price elasticity of -27.7 for the low-quality product.) There
benefit:
for the
demand
for low-,
highly dependent on the website's position on the Pricewatch
is
first
The
the website's medium-quality offering: a
with a premium quality dual fan. The
CPUs
and high-quality memory modules
product
for
fan attached.
website's high-quality offering: a
low-,
demand
in
many
many
fewer sales of
many customers
coefficient estimate
is
on the
clear statistical evidence
CPUs
with premium-quality
of the product categories the own-price elasticities of
medium- and high-
quality products cannot be precisely estimated.
To save space we do not report estimates
estimates are qualitatively similar.
(a bare
6.3
The
CPU)
is
The own
of the
demand
for
CPUs
through
site B.
The
price elasticity of site B's low-quality offering
estimated to be -28.3.
The mechanics
of obfuscation: incomplete consumer search
model of obfuscation we discussed
Stahl's
model of
search with heterogeneous consumers. In Stahl's model, consumers do not learn
all prices.
We
first
explicit
in section 2
was
noted that any action that raised the cost of learning about each firm's offerings or
that forced more consumers to conduct firm-by-firm searches could be regarded as a form
of obfuscation.
In this section
we
exploit a feature of our dataset to provide additional
evidence that consumers are not fully informed about the prices charged by the Pricewatch
retailers.
27
Our data do not provide any easy way
low-quality offerings
highly on the Pricewatch
offers
The market share advantage
incomplete.
is
list
does, however, provides a nice opportunity to
higher quality products and prices
quality
is
but
list,
and idiosyncratic consumer
identical products. If all
of the
ory choose between site
dummy
data using a
two
sites
We
incomplete.
would buy
for
A
and
analysis of
site
B we
utility Ui wcq t over
Uiwcqt
A
or site
the website
w
with
=
Pi log(l
site's price for
parsimony).
it
from the cheaper
Differences between
more medium-quality memory. 30
how consumers medium- and
high-quality
effect
mem-
estimate simple logit models on the consumer-level
B
at time
t
and
Ei
makes
w a standard
on the Pricewatch
product categories and different quality
If
sells
site.
i
site
A
(versus site B) as
choosing to buy a category
this choice to
maximimize
list
logit error.
The
ri cq
first
for the category.
+
his
e iw ,
right-hand
The second
is
the product that being purchased. Note that the estimated coefficients
consumers are
fully
levels are constrained to
be equal
(for
informed, one would expect that 0\ would be zero and
would be negative.
first
column of Table 7 reports estimates from a regression run on a dataset contain-
ing observations from
One
offer
medium- and high-
)
a quality-category fixed
for different
30
site's prices for
+ PLowRankwcq t) + 02 \og(Pwcqt + foTimeTrendt +
side variable reflects a site's position
The
have data on two websites that
we assume each customer
site
product from
c quality q
/?2
structure of our dataset
whether each consumer chose to buy from
the dependent variable. Formally,
that
The
examine whether consumer learning about
consumers learned about each
To provide a straightforward
the
could also be attributable purely
preferences.
low-quality prices should not help predict which site
rj cq
it
memory, then we would expect that most of consumers buying medium-quality
memory from one
with
more
of firms ranked
could reflect that some consumers have not examined the
by firms farther from the top of the
to price differences
consumer learning about the
to see whether
all
four
memory
categories
and on both medium- and high-quality
+ PLowRank)
was significant in the medium-quality
than fully informed about prices, but only
because a site's rank on the low-quality product is correlated with its rank on the medium-quality product
(which is not well-defined and, therefore, not included in the regression). This alternate interpretation is no
longer possible when we study the relative shares of the two websites with identical products sold at prices
could imagine that the coefficient on log{\
regressions
we ran
earlier,
not because consumers were
we know.
28
less
The
modules.
coefficients
on log(l
+ PLowRank) and
standard deviation of the difference in log(l
The
log(P) are both significant.
+ PLowRank)
across websites
is
more than
ten times larger than the standard deviation of the difference in log(P), so the coefficient
estimates can be thought of as indicating the a firm's position on the low-quality Price-
watch
list
has a slightly stronger effect than does
its
price for the item being purchased.
This suggests that there are a substantial number of consumers
both
and a substantial number who are
sites' prices
The second and
who
are informed
about
not.
third columns report estimates from the
same
regressions run separately
on medium-quality and high-quality memory modules. In both cases the results indicate
clearly that the low-quality module's position
minant of
sales of
medium- and
on the Pricewatch
an important deter-
list is
we
high-quality modules. In the medium-quality case,
are
unable to find significant evidence that the relative prices of medium-quality modules on
the two sites affects where consumers
for the high-quality
products appear to be playing a more important
The mechanics
6.4
The second
explicit
model
for
is
model
damaged goods and
retailer's prices for
role.
in Section 2
was a variant on the
of Ellison (2003). In this
model the practice of
requiring consumers to visit a site to learn
higher quality goods raises prices even
what each
we discussed
of obfuscation
case, the sites' prices
and adverse selection
of obfuscation: add-ons
rational expectations add-on pricing
posting prices for
buy them. In the high-quality
if
consumers
(in
its
prices
equilibrium) correctly anticipate
the higher quality goods will be.
The main reason
for this
that adopting add-on pricing creates an adverse selection problem that discourages price
cutting:
likely
a firm that cuts prices will disproportionately attract cheapskates
who
are less
than other customers to buy higher-quality products.
In the section on basic results,
we
verified
quality products were attracting customers
in other words, that the low-quality
in negative coefficients
In this section,
on log(l
one implication of
who upgraded
we examine the evidence on the
in the
model, that the low-
to the higher quality products, or,
products were effective
+ PLowRank)
this
loss leaders.
This was reflected
medium- and high-quality
theory's
more important
regressions.
prediction, that the
use of the add-on pricing creates an adverse selection problem (which would be expected to
29
lead to higher prices). In our
on log(l
reflected in the coefficient
than
in the
demand
system, such an adverse selection problem would be
+ P Low Rank)
medium- and high-quality
being larger in the low-quality regression
The
regressions.
larger coefficient indicates that
lowering the price of the low-quality product would lead to a larger proportional increase
in low-quality sales
than
estimates on the log(l
in
medium-
or high-quality sales.
+ PLowRank)
variable in the seventeen
three quality levels in each of the four
CPUs
the regression examining
CPUs
examining
In
all cases,
memory module
sold through site C;
effect.
In every category
medium- and high-quality
offer
would
regressions.
result in
128MB PC100
(i.e.
is
in every
Most
An
alternative
way
driven by functional form assumptions)
sample means. For example, when
256MB memory, 63%
in tenth place, only
we
We
catgories.
are identical.
levels in the regression
of
its
35%
site
A
consistent with the presence of an
column of the
table) the coefficient
of the differences are statistically significant.
A move
column.
of this adverse selection effect (and a
Finally,
ran:
larger than the coefficient estimates from the
64%, 41%, and 25% decreases
quality modules, respectively.
is
we
categories; three quality levels in
is
a sense of the magnitudes suggested by these
estimates from the
list
regressions
sold through site B.
estimate from the low-quality regression
To
demand
and two quality
the pattern of the coefficient estimates
adverse selection
Table 6 reports the coefficient
way
coefficient estimates, consider the
from
first
to third on the Pricewatch
medium-, and high-
in unit sales of low-,
for us to offer intuition for the
magnitude
to do so avoiding any worries about results being
it
to look at the firm's quality
or site
B
is first
unit sales are low-quality
mix using simple
on one of the Pricewatch
memory.
of the unit sales are low-quality
On
days when one of them
memory 31
.
also find Table 6 striking for the consistency of the estimates across
cannot reject the hypothesis that the six coefficient estimates in the
Nor can we
lists for
reject the equality of the five estimates in the
product
first
row
second row, nor of
the five estimates in the third row. Given that the products in the different categories differ
in price, that
31
Medium
40%. Total
moving up or down one rank involves changing the item
quality sales increase from
sales, of course, are
much
16%
lower
of total unit sales to
when
the firm
is
21% and
price by a different
high-quality sales from
21%
to
in the tenth position so these "increases" are
decreases in unit sales of the higher quality products being offset by even larger decreases of the low-quality
product
30
number
of dollars in the different categories,
quality levels differ between
and that the attributes that distinguish the
memory modules and CPUs, we
find this regularity intriguing.
Instrumental variables estimates
6.5
We
have assumed so far that variation in a website's rank on the Pricewatch
medium- and high-quality
prices
exogenous.
is
We
have argued that this
list
and
in its
a reasonable
is
assumption, but in this section will investigate the robustness of our conclusions to two
instrumental variables strategies.
We feel that
there are two primary sources of exogenous variation in our websites' prices
(relative to the prices of
competing websites).
and changing one's prices involves a time
drift as
cost, the websites'
its
prices in response to cost shocks.
unexpected employee absences that make
Another
order flow.
more common, but
is
common
to
all
less clearly suitable for
variable that
and
firms;
may
(especially in the
it
more
One
difficult to
(2) prices of
source of such shocks
process a normal day's
latter are
use as instruments because (1) they will not
if
wholesale prices for low-quality modules are
other firms' higher quality products are an omitted
medium- and high-quality demand
of instruments
PC100 memory modules, which we
call "other
we
log(PMid), and log(PHi)
for
memory. These variables may
will
use to estimate
demand
for
128MB
speed" instruments and "cost" instruments.
+ PLowRank).
the same website but for a different product,
reflect
memory
regressions).
"other speed" instruments are the contemporaneous values of log(l
prices:
The
be correlated with wholesale prices of medium- and high-quality
We have constructed two sets
lists
inattentive. Second,
is
variation in the firm's wholesale acquisition costs.
lead to variation in a firm's Pricewatch rank
The
ranks on the Pricewatch
other sites change prices during periods in which our firm
the firm will raise or lower
is
because monitoring other firms' prices
First,
both sources of exogenous variation
in
128MB PC133
PC100 memory
the website's ranks in the two categories could drift in unison during periods of
inattention;
and
if
employees do not show up for work, the firm
product categories. 32
Memory
If
the
prices have a strong
periods of inattention.
demands
for
may
PC100 and PCT33 memory
downward trend
raise its prices in all
are independent (recall
so a firm's rank in each category will drift
In practice, the correlation
is
strong but far from perfect.
category, for example, the R-squared of a regression of log(l
31
+ PLowRankioo)
In the
on log(l
+
up during long
128MB PC100
PLowRank\33),
PC133 ranks would be
that the products are not substitutes for most consumers), then the
exogenous.
The
"cost" instruments for log(l
+ PLowRank),
log(PMid), and log(PHi) are
derived from our data on the firm's acquisition costs for low-, medium-, and high-quality
modules. 33
128MB PC100 memory
If
the wholesale prices for low-quality
memory
our firm
faces are not perfectly correlated with the prices other firms face, then the low-quality
wholesale price should be correlated with the firm's Pricewatch ranks. 34
128MB PC100 memory modules
Table 8 reports estimates of the demand equations for
(comparable to those
Table
in
low quality regression (in the
log(l
+ PLowRank)
Again we
variable
from their values
obtained using the two sets of instruments. In the
column of the
first
is
demand
find that the
coefficient estimates
3)
table), the coefficient estimate
unchanged and the variable remains highly
for low-quality
products
on the log{l+P LowRank) variable
in the
is
on the
significant.
extremely price-sensitive.
in the next
first
The
two columns are reduced
uninstrumented variables by about one-half. The standard errors
are substantially increased, failing to provide significant evidence either that loss leader
effects exist or that
our previous estimates of the loss leader benefit were too large.
can, however, reject the hypothesis that the coefficient estimates in the
We
medium- and high-
quality regression are as large as the coefficient estimate from the low-quality regression.
Hence, the results confirm that the
The standard
errors
loss leader pricing creates
an adverse selection problem.
on the \og{PMid) and log(PHi) variables are
also increased.
Again, with the "cost" instruments, the log(l + PLowRank) variable remains significant
about extreme
in the low-quality regression confirming our finding
the other rank instruments, and the exogenous variables in the
log(l
+ PLowRankizz)
has a t-statistic of
13.
There
is
much
they can be predicted almost perfectly in such regressions,
PC100
demand equation
less variation in
e.g.
price-sensitivity.
we
is
0.58.
The
Again,
coefficient
on
log(PMid) and log(PHi) and
get an R-squared of 0.995 in the
128MB
data.
33
These data are only available on days on which the firm made wholesale purchases. Gaps of one or a
We fill in the missing data by interpolating linearly. The instruments for \og(PMid)
and \og(PHi) are simply the log of the cost of the cost of the product in question. Our instrument for
P Low Rank is the log of a predicted rank obtained by adding a constant to what the firm's rank would be
few days are common.
if it
34
priced at cost.
In practice, the cost-based instrument for
based instrument.
PLowRank) on
t-statistic
In the
128MB PC100
PLowRank
has predictive power, but
less
than the speed-
category, for example, the R-squared of a regression of log(l
+
the cost-based instruments and exogenous variables has an R-squared of 0.48 with the
on the instrument being about
very strong predictive power,
e.g.
in
the
five.
The instruments
128MB PC100
first-stage regressions.
32
for
log(PMid) and \og(PHi) again have
category we again get R-squareds above 0.99 in
the coefficients on this variable
become
insignificant in the other
two columns leaving us
unable to say that the IV regressions provide additional support for our loss leader results.
The standard
errors
on the log(PMid) and log(PHi) variables become much
We
most of the true exogenous variation
believe that this reflects that the instruments eliminate
in
larger.
these variables (which comes from these prices gradually becoming higher relative to the
low-quality price in between adjustments and then being
much
smaller immediately after
adjustments).
Markups
7
This section examines price-cost margins.
retail
It
has two motivations. First, the future of
depends on the balance between search and obfuscation. At one extreme,
e-
retail firms
would struggle to survive. At the other, search engines wouldn't be worth using. Markups
provide descriptive information about
universe.
how
the battle
is
playing out in the Pricewatch
Second, analyzing markups can provide a more complete understanding of the
mechanisms by which obfuscation
affects prices.
Table 9 presents our estimates of the actual markups.
The
table presents revenue-
weighted average percentage markups for each of the four categories of
and
for
CPUs
sold through each of the
two
sites.
35
The
dollar
memory modules
markups were obtained
by adding the standard shipping and handling charge to the advertised item
then subtracting the wholesale acquisition
cost,
an estimate of marginal labor costs
cost, credit
card
fees,
for order processing, packing,
estimation, omitting observations from
December 2000
and
an approximate shipping
and returns, and an
allowance for losses due to fraud. 36 In each category the sample period
demand
price,
(for
is
that used in the
which we do not have
cost data).
In the two 128
MB memory categories, the markups for low-quality products are slightly
negative. Prices have not, however, been pushed far below cost
by the hope of attracting
customers who can be talked into upgrading. Markups are about 16%
modules, and about
35
27%
The percentage markup
The labor and shipping
some error.
3
for high-quality
is
modules. Averaging across
the percentage of the sale price,
i.e.
100(p
all
for
medium-quality
three quality levels,
— mc)/p.
costs were chosen after discussions with the firm, but are obviously subject to
33
markups are about 8% and 12%
for a
PC100 module and
The
16%
of
firm's average
in the
two categories. This corresponds to about
ten dollars for a
markups
in the
PC133 module.
256MB memory
the two categories. Part of the difference
in
five dollars
due to the
is
13% and
categories were higher:
fact that
a higher fraction
consumers buy premium quality products, but the largest part comes from the markups
on low-quality memory being substantially higher. 37
Finally, the average percentage
B and 6%
at website C.
in dollar terms.
The
markups on Athlon CPUs
difference in
markups
The average markups on CPUs
dollars, respectively.
The
are
much lower: 3%
not as dramatic
is
at the
two
sites are
if
at website
one looks
about
six
largest part of the difference in average percentage
at
them
and twelve
markups
is
due to markups on the upgraded CPU-fan combinations being lower than markups on high-
memory
quality
low-quality
An
modules. Markups on low quality
CPUs
are also lower than
markups on
memory modules.
obvious question to ask
is
whether the markups reported above are what one would
expect under static Nash equilibrium pricing given the strength of the adverse selection
effect
we
identified in the previous section.
(There
are, of course,
many
reasons for prices
to be higher or lower, such as repeated-game collusion or desire to build short-run market
share.) This, however,
reason
is
is
something that one can not do with our demand estimates. One
that we have not been able to obtain precise estimates of own-price elasticities for
medium- and high-quality products.
are using are not well-behaved in
A
some
second
due to functional form choices we made
37
that the functional forms for
of the counterfactuals that
doing a Nash equilibrium analysis. This
interpret.
is
is
partially
demand we
must be considered when
due to data limitations and partially
in order to
make the demand
results easier to
38
also due to the sample period being different. The revenue weighted estimates put a great deal
on the latter part of the data. The 256MB data end 6 weeks before the 128MB data. Markups
on 128MB modules were relatively low in the last six weeks.
38
Examples of unavoidable data limitations are that it is hard to separately identify the effect of the
firm's low quality price and its Pricewatch rank on demand since they are highly correlated and that there
is little variation in medium- or high- quality prices that is not collinear with the time trends. Examples
of choices we made include our use of a firm's Pricewatch rank as the primary explanatory variable, which
is not appropriate if one wants to predict what would happen if the last firm on the list raised its price
toward infinity, and the specification of \og(QLow) and log(QMid) as linear functions of \og(PHi), which
implies that a firm could also earn unbounded profits by setting PLow and PMid above cost and sending
Part
is
of weight
34
Although we cannot provide a structural analysis of markups, we think that we can provide evidence that the observed markups are roughly consistent with firms' approximately
maximizing
We
static profits.
think that our
demand system
at the prices one typically observes in the data.
We
demand
accurately reflects
therefore feel comfortable
computing
projected profits at prices charged in the data and can examine whether there appear to
be strong incentives to deviate from the actual prices to other prices
Consider, for example, the prices from October 12, 2000 in Figure
on
day are
this
Our
fairly typical.
retailer
same range.
the
in
The markups
1.
bought low-quality modules wholesale on the
previous day for $72. 39 Shipping and handling charges are about $1 below the
sum of actual
shipping costs and other marginal costs, so the lowest-priced firm, listing at $68,
is
low-quahty modules at about $5 below cost and the sixth firm,
about $1
above
cost.
average
Given that the low-priced firms make
markup
for low-quality
modules
the sample mean. Site B's markups on
is
illustrate the effect
on a firm's
and $94, and
profits of
from 54.5 units
if site
B
in
1.
occupies the
position. Sales of high-quality
first
Figure
is
just a little
below
it
charged $103.49 and $148.50.
moving positions on the Pricewatch
predicted quantities and predicted profits for site
list
the quantity-weighted
probably about -2%, which
Table 10 contains results from a series of scenarios involving
the twelve places on the
sales,
medium- and high-quality memory on that day were
also typical: the wholesale prices were $77
To
many more
listing at $74, is at
selling
B assuming
site B. In particular, it
site
B had
Predicted sales of low-quality
first
position to 3.4 units
if it
The
sixth
has
occupied each of
memory drop
sharply
occupies the twelfth
modules decline much more gradually, from 2.8 units
position to 1.3 units in the twelfth.
list,
column reports estimates of
in the
site
B's
expected gross profit from being in each position under the assumption that
site
B would
have used the wholesale prices from the previous day to calculate
Site
B would
most have
liked to
be
that profits are fairly
in the fifth position,
flat
where
is
the
would have expected to earn $154. Note
over a wide range of markups. For example, the firm's profits in
the third and tenth positions are nearly equal.
the others
it
its costs.
The one
position that
is
clearly worse than
first.
PHi
toward infinity.
Given that the prices are from 9:01am Pacific time, the $72 figure
most firms had on costs before they set the observed prices.
35
is
likely to
be the
latest information
What would an
more than
e-Nash equilibrium look
by changing
e
are too high this will
they are too low
intermediate
level,
and moving
to another location
is
that no firm can gain
on the
If
list.
because everyone will want to move to the top of the
fail
will fail
it
price
its
The requirement
like?
because everyone
will
markups
list,
and
if
want to move to the bottom. At some
the losses that the top few firms incur selling low-quality modules below
cost could roughly offset the profits they derive from their greater sales of premium-quality
modules and thus we could have an e-Nash equilibrium with price dispersion.
Although
find the
first
site
B
not indifferent between the various places on the
is
position quite unattractive),
suggesting that play
First, profits
is
close to
can claim
this spot only
Site
B
in equilibrium.
cannot switch
its
(and would
of the profit estimates in Table 10 as
an £-Nash equilibrium with a reasonable
need not be exactly equal
40
to prevent deviations.
we think
list
for a
They need only be
price to $74
by charging $73, which would
e
and take the
few reasons.
close
enough
fifth spot.
It
$144 rather than
result in a profit of
$154.
Second, heterogeneity in website design will lead to firms' having different preferences
over positions on the Pricewatch
at the
same
prices,
and
high-quality memory.
profits.
The
fifth
is
more
list.
Recall that site
is
also the
slots.
This
is
A
demand
most attractive
consumers
for firm .4,
In particular,
call
B
it
but
its
preferences over
does relatively better in
A
tends to set
on the phone and talk to salespeople,
functions. It does not
such heterogeneity might be able to account for
the top
site
and B have similar designs. Other websites that employ
different tactics, e.g., requiring that
face quite different
than
buy medium- and
consistent with our earlier observation that site
lower prices than site B. Sites
may
sales
of Table 10 reports estimates of site A's gross
the other positions are somewhat different. 41
the top few
makes higher
successful in inducing customers to
The seventh column
position
A
seem unreasonable
why Computer
Craft
is
to suppose that
willing to
occupy
slot.
Third, cost heterogeneity will also contribute to firms having different preferences over
40
Ellison
and Fudcnberg (2003) provide several examples of how such "market impact
effects"
can create
a multiplicity of asymmetric pure strategy equilibria.
41
The attractiveness of the fifth slot is due to an unusual feature of the prices in Figure 1. The fifth-ranked
markup is a full four dollars above the markup of the third-ranked firm. This may be a situation
where our rank-based demand system is overestimating profits.
firm's
36
positions on the Pricewatch
watch
retailers
may
In addition to the standard cost heterogeneities, Price-
list.
differ significantly in their
expectations about wholesale prices.
example, the most recent time prior to October 11th that our firm
chase of low-quality
listed firms
memory was on Monday October
9th
when
it
made a
For
wholesale pur-
Some
paid $78.
of the
might not have checked wholesale prices since then, or might have made the
mistake of using acquisition costs rather than replacement costs when pricing inventory
The
acquired three days earlier.
would have expected
attractive.
true that
would wait
until the 13th or
lists
the prices that site
used the three-day-old costs. The ninth position
if it
It is also
eighth column of Table 10
many
now
is
B
the most
would not shiporders received on the 12th
retailers
modules
until the 13th to replace shipped
in inventory. It turns
out that before the end of the day on the 13th, low-quality modules could be purchased
wholesale for $68.
The
final
column
had anticipated these lower
if it
of the table lists the profits site
costs.
The
first
position
is
now
other potential explanation for the top-ranked firm's behavior
had
different information
Fourth, the
e
in
about
B would
have expected
An-
the most attractive.
is
thus that
it
might have
costs.
e-Nash equilibrium would include monitoring and decision costs and
should not be thought of as extremely small.
manager. Monitoring dozens of Pricewatch
Our
for
retailer,
example, has just one
a small part of his duties and he
lists is just
does not have nearly enough time to make minor adjustments to every price every hour.
Hence, we should not expect to see
times.
present in the data.
On
1 left their prices fixed
list
came
spot.
some kind
Decision costs of
all
maximizing the
firms
we
calculated at
are necessary to account for the price inertia that
October 12th,
for
example, the top six firms on the
throughout the afternoon. The
later in the day:
profits
Augustus Technology cut
its
first
list in
is
Figure
change to the top half of the
price to $73
and took over the
fifth
42
Although we think that the observed markups are roughly consistent with a
model of competition, two features of the data stand out as very hard
a model.
First, the
There were
and
all
left
five
the top 12.
positions.
Two
markups
other changes
Two
for low-quality
among
256MB memory
firms lower
on the
list.
The
37
first
Nash
to reconcile to such
are higher than one would
eighth-place firm raised
its
price
page in the ninth and tenth
page and one to stay on it.
other firms cut their prices to $76 to enter the
other firms cut their prices to $77, one to enter the
static
first
—a
predict even
if
markups over
markups
The
one ignored the
time.
The two
cost in
Second, there
features are not completely unrelated.
memory
for low-quality
number
firm sold a large
loss leader benefits.
are positive
of low-quality
is
We
256MB modules
price far below the price offered
We
fact that
at
two subperiods.
about ten dollars above
five dollars
by the standard wholesale
six or fewer retailers
competing
above cost in
A
period markups were relatively high in
all
four
in
them 256MB modules
distributors.
43
As a
at a
result, there
two months rather than dozens. 44
in these
do not know why higher markups prevailed
sell
in
average
think we understand what happened in the former period.
small number of retailers found an obscure supplier willing to
were effectively
The
entirely attributable to
September-October 2000 and a large number at about
February-March 2001.
some variation
is
February and March of 2001. In this
memory
categories.
Conclusion
8
In this paper
we have noted that the extent
search costs
not clear. While the Internet clearly facilitates search,
is
to
which the Internet
adopt a number of strategies that make search more
we
see that
demand
is
sometimes remarkably
difficult.
elastic,
will
reduce consumer
also allows firms to
it
In the Pricewatch universe,
but that
this
not always what
is
happens.
The most popular obfuscation
strategy for the products
we study
to intentionally
is
create an inferior quality good that can be offered at a very low price.
Retailers could,
by refusing
to let the search
of course, avoid the negative impact of search engines simply
engines have access to their prices. This easy solution, however, has a free rider problem
if
other firms are listing,
43
I
will suffer
from not being
listed.
What may
help
make the
on July 10. On that day, when the
Cost cuts its retail price to $218
full $51 below the next lowest price. The firm remained all alone with a price at least $37 less than that of
any other firm for three weeks. We infer from this that its acquisition costs must have been very low.
44
We cannot provide any estimates about how the markups of the low-cost retailers varied in August as
the number of low-cost retailers increased from one to (we think) six, because the retailer from which we
The
first retailer
to have found the supplier appears to have found
firm that supplied us with data bought modules wholesale for $270,
it
PC
—
got the data appears to be the last of the six to have found the low-priced supplier. At that point, in early
September, low-quality memory modules were selling for about $10 above cost. Markups remained steady
and no other retailers appear to have found the supplier for the next two months. At the end of October
wholesale prices fell dramatically, apparently eliminating the advantage. At this point several retailers that
had been absent for months jumped to the top of the Pricewatch list and margins on low-quality memory
quickly dropped to approximately zero.
:;s
obfuscation strategy we observe popular
is
that
hard not to copy
it is
it
—
if
a retailer tries
to advertise a decent quality product with reasonable contractual terms at a fair price
be buried behind dozens of lower price
makes
quality aspect of the practice
and
We
and
loss-leader models,
would
also
it
it
offers
on the search engine's
somewhat
seems that
it
different
it
will
The endogenous-
list.
from previous bait-and-switch
would be a worthwhile topic
for research.
45
be interested to see more work integrating search engines into models with
search frictions, exploring other obfuscation techniques (such as individualized prices), and
trying to understand
why adoption
of price search engines has been slow.
Large online retailers have very high expenses. Although they can avoid the expense of
setting
It is
up hundreds of retail
much cheaper
number
off
to ship a truckload of merchandise to a store than to ship a similar
of individual packages with
them
the shelf and bring
shelves
locations, they have to handle every item they ship individually.
and put them
types: they have
UPS.
cheaper to have a customer take packages
It is
to a cashier than
in boxes.
Large online
it is
to hire
been very intensive users of managerial
Given these expenses, large
e-retailers
Many
to pick packages off
have very high expenses of other
retailers also
talent; they
and they have enormous website development
advertising;
someone
costs.
have spent a
lot
on
46
cannot be competitive with Walmart's prices.
may be
willing to
pay a premium to avoid
going to the mall and to take advantage of superior product variety.
A more serious problem
This
is
is
not necessarily worrisome.
people
that managerial, advertising, and website development costs seem
than marginal
costs.
price search engines
If
more
like fixed costs
become more popular, and better sources
information develop on the reliability of retailers, then vigorous price competition
of
may push
online prices below the level necessary to cover these fixed costs.
What
happen
is
will
that
happen
to e-retail
demand can become
break even. While this
is
Simester's (1995)
makes the low
price search flourishes?
One
thing
we have seen can
incredibly elastic, so elastic that retailers could never
a frightening possibility to retailers,
to us. In seems likely that retailers
4
if
would make
model seems the most
it
does not seem very likely
significant efforts to ensure that price search
similar to the practice.
prices on Pricewatch have advertising value
is
We
would imagine, however, that what
that the offerings are sufficiently attractive so
as to force a retailer to set low prices for its other offerings to avoid having everyone
product.
46
Amazon,
for
example, reported more than $250 million
39
in
R&D
costs in 2000.
buy the advertised
engines never work too well, and markups will end up at least as high as they are in the
(unbranded) Pricewatch universe. Similar markups would force large retailers to cut back
substantially on
management, advertising, and website
that they could do
so.
Another possibility
market structure than traditional
retail.
is
costs,
but
it is
that e-retail could end
certainly plausible
up with a very
different
Small e-retailers without venture capital realized
long ago that the Internet allows one to operate a retail store with virtually no advertising
or website development costs. At present, small e-retailers are fax
more famous competitors. Whether they
will
remain more
more
efficient
efficient
and come
to
than their
dominate
price-search mediated markets or whether they will be pushed aside as larger firms copy
their
model
is
another interesting question.
40
References
Yannis (1997): "Reducing Buyer Search Costs: Implications
ketplaces," Management Science 43, 1976 - 1993.
Bakos,
for Electronic
J.
Mar-
Morgan (2001): "Information Gatekeepers on the Internet and
Homogeneous Product Markets," American Economic Review 91,
Baye, Michael R., and John
the Competitiveness of
454-474.
Baye, Michael R., and John
ination on the Internet,"
Morgan
Economics
(2003): "Information Gatekeepers
Baye, Michael R., John Morgan, and Patrick Scholten (2003):
"Price Dispersion in the
Small and the Large: Evidence from an Internet Price Comparison
trial
and Price Discrim-
Letters 76, 47-51.
Site,"
Journal of Indus-
Economics, forthcoming.
Brown, Jeffrey R., and Austan Goolsbee (2002): "Does the Internet Make Markets More
Competitive? Evidence from the Life Insurance Industry," Journal of Political Economy
110, 481-507.
Brynjolfsson, Erik, and Michael Smith (2001):
Shopbot: Brand
Still
"Consumer Decision-making at an Internet
XLLX, 541-558.
Matters," Journal of Industrial Economics
Burdett, Kenneth, and Kenneth L. Judd (1983):
"Equilibrium Price Dispersion," Econo-
metrica 51, 955-969.
Chevalier, Judith,
Online:
and Austan Goolsbee (2003): "Measuring Prices and Price Competition
Noble," Quantitative Marketing and Economics 1,203-222.
Amazon vs. Barnes and
Chevalier, Judith, Anil Kashyap, and Peter Rossi (2000):
Periods of Peak
Demand? Evidence from Scanner Data,"
"Why
Don't Prices Rise During
University of Chicago Graduate
School of Business, mimeo.
Clay, Karen,
Ramayya Krishnan, and Eric Wolff (2001): "Prices and Price Dispersion on
Book Industry," Journal of Industrial Economics 49,
the Web: Evidence from the Online
521-540.
Deneckere, Raymond, and R. Preston McAfee (1996):
nomics and Management Strategy
Diamond, Peter
(1971):
5,
"A Model
"Damaged Goods," Journal
of Eco-
149-174.
of Price Adjustment," Journal of
Economic Theory
3,
156-168.
Ellison,
Glenn (2003): "A Model of Add-on Pricing," mimeo, MIT.
Gabaix, Xavier and David Laibson (2004): "Competition and Consumer Confusion," mimeo,
Harvard University and MIT.
Janssen, Maarten, and Jose Luis
Moraga
(2000):
"Pricing,
of Internet Markets," Tinbergen Institute Discussion
41
Consumer Search and the
Paper 2000-0042.
Size
Scott Morton, Fiona, Florian Zettelmeyer, and Jorge Silva-Risso (2001):
Retailing," Journal of Industrial
Economics
Scott Morton, Fiona, Florian Zettelmeyer, and Jorge Silva-Risso (2003):
Cowards:
Why
Simester,
Duncan
are Internet
(1995):
"Internet
Car
49, 501-519.
"Cowboys
or
Car Prices Lower?," mimeo.
"Signalling Price
Image Using Advertised
Prices,"
Marketing
Search,"
American
Science 14, 166-188.
Stahl, Dale O. (1989): "Oligopolistic Pricing with Sequential
Economic Review
Consumer
79, 700-712.
Stahl, Dale O. (1996):
"Oligopolistic Pricing with Sequential
Consumer Search and Het-
erogeneous Search Costs," International Journal of Industrial Organization
Varian, Hal (1980):
"A Model
Walters, Rochney (1988):
of Sales,"
American Economic Review
"Retail Promotions
Some Key Hypotheses," Journal
14,
243268.
70, 651-659.
and Retail Store Performance:
A
Test of
of Retailing 64, 153-180.
McKenzie (1988): "A Structural Equation Analysis of the
Impact of Price Promotions on Store Performance," Journal of Marketing Research 25,
Walters, Rochney, and Scott
51-63.
White, E. (2000): "No Comparison. Shopping 'Bots' Were Supposed to Unleash Brutal
Price Wars. Why Haven't They?" The Wall Street Journal October 23, 2000, R18.
12
PRICE WATCH® Query
System Memory PC100 128MB
Back To
-
New Components Home
Back To - "No t Exactly New" Home
PRODUCT
BK VV»
PRICE FOR ONLINE
Generic
DESCRIPTION
ORDERS ONLY
128MB PC100 SDRAM DIMM
-
-
8ns
Gold leads
.- * LIMIT ONE Easy installation
-
snip
Mill K
in
$ 68
9.69
DEALER/PHONE
Computer Craft
DATE/HK
INSURED
10/12/00
12:40:05
Inc.
AM 800-487-4910
FL
727-327-7559
CST
stock
PART*
MEM-128100PCT
Online Ordering
Connect
Generic
ONLINE ORDERS ONLY - 128MB
SDRAM PC100 16x64 168pin
-
* LIMIT
ONE
INSUREDS'). 95
$ 69
10/11/00
10:59:56 PM
CST
Computers
888-277-6287
949-367-0703
CA
Online Ordering
1st Choice
Generic
ORDER
PC100 SDRAM DIMM
PRICE FOR ONLINE
-
128MB
-
* LIMIT
ONE
-
-
InStock, 16x64-Gold
$ 70
10.75
Leads
10/11/00
2:11:00 PM
Memory
CST
—
949-888-3810
CA
-
C-
-
CA
-
P.O.'s accepted
Online Ordering
Generic
PRICE FOR ONLINE ORDER - 128mb - *
LIMIT ONE - in
True PC100 SDRAM EEPROM
stock - with Lifetime
DIMM16x64 168pin 6ns/7ns/8ns
Warranty
Gold Leads
IN
Generic
STOCK, 128MB PC100
Pin,
Generic
7/8ns
- * LIMIT ONE Not
compatible with E
3.3volt
SDRAM
unbuffered
-
Gold Lead 168
with Lifetime warranty
$72
CST
3.3volt SDRAM 168
with UFIT1ME
-
*
LIMIT
ONE
$ 74
10.25
Pin,
7/8ns
-
800-382-6678
P.O.'s accepted
Online Ordering
Memplus.com
CST
877-918-6767
626-918-6767
10/9/00
6:53:25 PM
Portatech
800-487-1327
880060
CA
CST
128MB PC100
House
Brand
pcboost.com
AM —
10/11/00
12:44:00 PM
10.95- UPS
INSURED
$ 74
Machine
PRICE FOR ONLINE ORDERS ONLY 128MB True PC100 SDRAM DIMM 8ns Gold - warranty
10/10/00
11:30:39
9.85
-
*
LIMIT
ONE
10/11/00
10:20:23
10.50 FedEx
$ 74
WARRANTY
CST
1st
Compu
Choice
AM 800-345-8880
OH
-
MC
-
800-345-8880
Sunset
Generic
DIMM16x64 168pin
128MB 168Pin TRUE PC100 SDRAM
-
6ns/7ns/8ns Gold
Leads
OEM 16X64
$ 75
10/11/00
2:37:00 PM
$10
CST
Marketing, Inc.
800-397-5050
410-626-0211 P.O.'s
Generic
128MB 16x64 PC100 8ns SDRAM.
-
*
LIMIT
$77
ONE
10/12/00
9:37:45 AM
10.90
CST
accepted
PC COST
800-877-9442
847-690-0103
a
Online Ordering
Augustus
STOCK, PC100, 128MB, 168pins
DIMM NonECC, - with Lifetime
IN
Generic
-
*
LIMIT
$10.95
$77
5
& UP
UPS Ground
warranty
For
10/9/00
5:11:10 PM
CST
Technoloqv, Inc
877-468-5181
909-468-1883
CA
-
IL
-
Online Ordering
Computer Super
Generic
128MB PC100 8NS 16x64 SDRAM
-
one year warranty
*
LIMIT
ONE
$ 78
Ups Ground
$10.62
10/11/00
5:16:36 PM
CST
Sale
800-305-4930
847-640-9710
Online Ordering
PRICE FOR ONLINE
Generic
ORDERS ONLY
-
PC100 128MB NonBuffered, NonECC
16x64 SDRAM DIMM 3.3V 8ns mem
module
-
* LIMIT
lifetime
ONE
-
with
warranty
$ 78
CST
Page
1
Figure
ET
1:
A
sample Pricewatch search
on October
list:
10/5/00
6:29:59 PM
10.95
1
888-485-8872
909-869-8859
Next
128MB PC100 memory modules
43
USA, LLC
of 30
5 10 15 20 25 30
12, 2000.
Jazz Technology
at
12:01pm
CA
ME-
GBP100128
Prices for
128MB PC100 Memory Modules
140
<v
o
May-00
Jul-00
Sep-00
Nov-00
Jan-01
Mar-01
May-01
Date
Figure
2:
Prices for
128MB PC100 memory
modules: the lowest and 12th lowest prices on
Pricewatch and website B's price
11
Memory
Spec. Chart
Samsung/Micron or Major
512MB PC3200 [ADD
-
PC3200
DDR 512MB (Select Your Memory Module)
Industry Standard
PC3200 |ADD
S25]
OEM 512MB
512MB
•
•
CAS
2.5 Latency
•
CAS
•
Hand Picked 5ns
•
Hand Picked 5ns
•
6 Layer
•
6 Layer
PCB
•
•
Low
Noise Shielded
Board
32x8 DRAM Type
Samsung/Micron or Major
Return Shipping Paid
•
No
•
Satisfaction
&
Industry Standard
•
5
Days
Memory
Full
Refund
Tested Before Ship
Out
• Copper Heat Sink
the
Figure
3:
A
Latency
•
OEM DRAM
Downgrade
Chips
DRAM Type
DRAM
32x8
•
3
•
20%
Restocking Fee
According
•
Memory
Days No Restocking Fee
to the
Market Value
Verify Compatibility with
Configurator
•
7
•
Return Shipping not Paid
•
Return Shipping not Paid
•
Improved Compatibility
•
9 Months Warranty
•
Lifetime Warranty
•
Aluminum Heat Sink
Compatibility
Lifetime Warranty
1
Shielded
Restocking Fee
Guaranteed
•
Low Noise
PCB Board
CAS
4 Layer Module Board
• 64x4 DRAM Type
•
Latency
Chips
•
•
2.5
•
Brands
PC3200
$15]
Memory up
to
-
Cool
Down
the
Memory up
-
Cool
to
35%
Down
40%
website designed to induce consumers to upgrade to a higher quality
module.
45
memory
Tufshop
Price: $53.81
Price (with Selected Options): $90.36
Super Buys
Make processor upto 30%
i
faster or your
maximum efficiency. You must have
(Highly Recommended)
Memory Upgrade
-
Certified intel
f~ Memory Upgrade
-
Certified
^f^7
J\
motherboard to run with
awesome
this
value package.
Approved specs Memory [+$23.11]
AMD Approved
Memory [+$17.35]
specs
I
XiJ*"\ Bonus Buys
Consider taking advantage of these special offers. Compare and
save. Purchase everything from one location and save on shipping
-
[""
Cable Upgrades
Rounded IDE and Floppy Cables (Complete Set) [+$11.91]
-
P* Essential Equipment
Sony Floppy Disk Drive [+$16.84]
-
P" Bonus Buy
-
10 pack of hand thumbscrews for Case [+$4.95]
f" Bonus Buy
-
12-Pc Computer Tool
Bonus Buy
-
RatPadzGS Ultimate Mousepad/Gaming Surface [ + $11.97]
f" Bonus Buy
-
CD-DVD Media
~ Thermal Management
'^§§§3^
^""*^
—
-
CAS
-
2
[+$16.98]
Cleaning Kit [+$4.93]
80mm
Dynatron
Related Options
Please take advantage
Memory Upgrade
Kit
of
Case Fan
[
+ $12.87]
these special offers.
Upgrade (Offers Performance Increase & Helps
in
I
1
Overclocking) [+$18.25]
p-
Memory Upgrade
CAS 2.5 Upgrade (Improves Performance over Cas3 & Helps
Games) [+$6.35]
-
with Applications and
*"
Pretest
|
|
'
Have us
\t
RMAs
C
No
<•
Pretest
in
merchandise before we ship to avoid costly
the future and Maximize your time
test your
Pretesting
's25P5333
"^
^~~r
-
Standard Pretest (Avoid costly RMAs) [+$6.97]
Memory Performance
Options to
make your hardware &
C
No Memory Performance Enhancements
,-
Memory Upgrade
6 Layer
-
PCB For
Stablity of
applications fly
Memory
-
more
layer
-
More
Better Design [+$8.37]
S*fe»» Enhancers
Options to
make your hardware &
C
No System Performance Enhancers
C
Memory Upgrade
C
Memory
Cooling
-
(•
Memory
Cooling
-
C
Memory Upgrade
applications
Thermaltake Memory Cooling
-
Kit (Active)
Copper Passive Memory Cooling
Kit
Aluminium Passive Memory Cooling
-
Thermaltake Memory Cooling
fly
Kit
[
+ $19.99]
[ + $11,15]
Kit
[+$9.91]
(Passive) [+$14.99]
Hi
:
Variable
Mean Stdev
Min
128MB PC100 Memory Modules
Max
683 website-day observations
LowestPrice
Rangel-12
62.98
33.31
21.00
6.76
2.52
1.00
13.5
PLow
PMid
PHi
log(l + PLowRank)
QLow
QMid
QHi
66.88
34.51
21.00
123.49
90.72
115.27
40.09
35.49
149.49
46.36
48.50
185.50
1.86
0.53
0.69
3.26
12.80
17.03
2.01
3.31
1.86
3.45
120.85
163
25
47
128MB PC133 Memory Modules
608 website-day observations
LowestPrice
Rangel-12
131.00
71.02
37.02
21.00
5.92
2.74
2.00
15.00
PLow
73.68
36.69
21.00
PMid
98.50
41.84
22.10
PHi
128.46
47.56
48.50
log(l + PLowRank)
1.77
0.64
0.69
QLow
10.08
12.13
QMid
2.14
4.91
QHi
4.79
3.93
256MB PC 100 Memory Modules
131.49
153.45
189.50
3.40
99
100
35
575 website-day observations
LowestPrice
Rangel-12
130.30
28.95
143.44
58.17
15.78
32.46
6.39
47.67
PLow
67.15
PMid
205.29
85.36
77.49
PHi
250.09
93.14
98.49
log(l + PLowRank)
1.90
0.55
0.69
QLow
2.87
4.61
QMid
1.00
1.88
QHi
0.87
1.63
256MB PC 133 Memory Modules
215.00
75.80
283.49
372.89
417.18
2.91
33
18
16
575 website-day observations
LowestPrice
Rangel-12
PLow
PMid
PHi
log(l + PLowRank)
QLow
QMid
QHi
Table
1:
Summary
143.90
71.97
25.49
155.75
14.01
32.46
6.60
269.00
67.00
291.45
345.45
392.21
2.73
92.26
43.00
78.49
104.37
1.97
0.49
0.69
5.29
10.24
136
1.10
1.93
12
3.61
3.86
19
213.14
249.95
77.95
91.91
statistics for
47
memory module data
Max
Mean
Stdev
Min
Website B
392 clock speed-day observations
Variable
Data
PLow
PHi
log(l + PLowRank)
QLow
QHi
for
180.80
LowestPrice
Rangel-12
43.84
43.90
105.25
302.45
219.68
142.39
341.45
2.08
0.42
0.88
2.64
0.88
1.25
0.57
0.92
173.08
42.36
97.00
298.33
11.16
5.62
3.75
36.69
8
6
Website C
392 clock speed-day observations
Data
for
PLow
PMid
PHi
log(l + PLowRank)
QLow
QMid
QHi
180.83
214.71
LowestPrice
Rangel-12
Table
2:
Summary
103.59
136.49
302.25
336.25
230.69
44.17
43.96
44.54
150.64
365.47
1.99
0.49
0.69
2.64
1.07
1.54
0.83
1.29
9
0.52
0.85
4
173.08
42.36
97.00
298.33
11.16
5.62
3.75
36.69
statistics for
Athlon processor data
Dep. rax.: quantities
of each quality level
Low q Mid? Mighty
Independent
Variables
log(l
AMD
12
+ PLowRank)
-1.48 T
-0.76*
(15.1)
(5.8)
(3.2)
0.43
-6.86*
2.72*
\og{PMid)
-0.41*
(0.5)
(5.9)
(2.0)
-0.06
2.68
-5.10*
(0.1)
-0.26*
(1.8)
(3.9)
SiteB
-0.31*
-0.59*
,:;.:,)
(2.9)
(5.7)
Weekend
-0.51*
-0.95*
-0.72*
(5.9)
\og{PHi)
log(LowestPrice)
Number
(8.6)
(8.1)
-1.08
1.10
0.90
(1.6)
(1.3)
(1.1)
683
683
683
of Obs.
Absolute value of ^-statistics in parentheses.
-48
denotes significance at the 10
% level.
Table
3:
Demand
for
*
denotes significance at the
128MB PC100 Memory Modules
48
5%
level,
f
128MB PC133 Modules
Low
Mid
Hi
128MB PC100 Modules
Mid
Hi
Low
PLow
PMid
PHi
-25.0*
0.4
-0.1
-12.9*
-7.0*
-6.9*
2.7*
2.7
-5.1*
PLow
PMid
PHi
Table
4:
-16.3*
-5.5*
7.9
-3.9
-4.1*
-3.4
2.4
-2.6
-36.8*
PLow
PMid
PHi
memory modules:
Price elasticities for
-4.1*
-11.5*
0.2
-0.1
-1.5
-0.1
-5.5*
-4.5*
256MB PC133 Modules
Low
Mid
Hi
256MB PC100 Modules
Mid
Low
Hi
PLow
PMid
PHi
-33.0*
-41.4*
-28.4*
-11.5
-2.7
4.5
3.6
-1.4
-5.5
-1.0
three qualities in each of four product
classes
Dep. var.: quantities
of each quality level
Low q Mid q High q
Independent
Variables
log(l
+ PLow Rank)
log(PMid)
\og(PHi)
Weekend
log(LowestPrice)
TimeinDays
-1.14*
0.10
(0.3)
(6.1)
(4.7)
-0.34
-4.81
4.14
(0.1)
(0.8)
(0.4)
-2.70
-2.98
-21.48
(0.7)
(0.5)
(1.5)
-0.47*
-0.26
-0.81*
(3.0)
(2.4)
(1.2)
-1.47
2.86
6.77
(0.5)
(1.0)
(1.5)
-0.02*
-0.02*
-0.01
(2.3)
(2.3)
(0.7)
0.46
-0.25
-1.14*
LowHasFan
(1.8)
(0.9)
(2.3)
Speed650
-2.88*
-2.97*
-5.74*
(2.5)
(2.9)
(3.0)
Speed700
-1.66*
-1.67*
-3.07*
(1.9)
(2.2)
(2.1)
Speed750
-0.82
-1.19*
-2.34*
(1.5)
(2.5)
(2.5)
392
392
392
Number
Absolute value of
-0.86*
of Obs.
i-statistics in parentheses. *
Table
5:
Demand
for
denotes significance at the
Athlon CPU's at
49
site
C
1%
level.
Product Category
Dependent
Variable
128MB
PC100
Memory
128MB
PC133
Memory
256MB
256MB
Athlon
Athlon
PC100
PC133
Memory
CPUs
Site B
SiteC
Memory
CPUs
Low
-1.47
-1.41
-1.79
-2.29
-1.28
-1.14
Quality
(0.10)
(0.09)
(0.36)
(0.36)
(0.27)
(0.19)
Sales
Medium
-0.76
-0.49
-0.79
-1.57
-0.86
Quality
(0.13)
(0.12)
(0.29)
(0.79)
(0.18)
Sales
-0.97
(0.31)
High
-0.41
-0.18
-0.66
-0.64
0.10
Quality
(0.13)
(0.06)
(0.25)
(0.35)
(0.30)
Sales
Standard errors
Table
6:
in parentheses.
Effect of low quality Pricewatch rank
on
sales of each quality level: estimates
from
six datasets.
Independent
Variables
log(l
+ P Low Rank)
log(P)
Number
of Obs.
Dep. variable:
Medium and
Dummy
for choice of site
Medium
A
quality
High quality
-0.38*
-0.68*
-0.22*
(5.12)
-3.01*
(4.57)
-0.87
(3.03)
-5.03*
(3.18)
(0.67)
(3.53)
10845
4078
6757
high
table presents estimates of logit models. The dependent variable is a dummy for
whether a consumer chose to buy from site A (as opposed to site B). The first column
pools the observations from all eight specific products: medium- and high-quality; 128MB
and 256MB; PC100 and PC133. The second and third are separate regressions for the
medium- and high-quality subsamples. Absolute value of z-statistics in parentheses. *
denotes significance at the 5% level. The regressions also included category dummies and
a linear time trend.
The
Table
7:
Evidence of incomplete consumer learning: effects of product prices and low-quality
on where consumers buy medium- and high-quality memory modules.
price rankings
50
Dependent
variables: Quantity of
memory modules
Low
Mid
-1.48*
-0.45
-0.22
-1.89*
-0.34
0.79
(7.8)
(1.5)
(0.6)
(5.0)
(0.2)
(1.7)
-1.06
-7.78*
-0.47
11.93
-9.07
-2.49
(0.7)
(3.8)
(0.1)
(1.9)
(1.1)
(0.5)
3.58
2.84
-3.35
5.54
4.26
-5.30*
Variables
log(l
+ PLowRank)
\og{PMid)
log(PHi)
Low
Hi
(1.9)
(1.0)
(0.9)
(1.0)
(0.5)
(1-7)
SiteB
-0.28*
-0.49*
-0.61*
-0.40
-0.48
-1.10*
(2.9)
(3.2)
(3.3)
(1.9)
(1.2)
(4.3)
Weekend
-0.48*
-0.97*
-0.67*
-0.57*
-0.98*
-0.75*
(7.9)
(8.0)
(5.1)
(2.0)
(6.9)
(5.2)
log(LowestPrice)
-1.85*
0.71
0.80
-5.01*
1.02
1.21
(2.6)
(0.7)
(0.8)
(2.5)
(0.3)
(0.7)
683
683
683
683
683
683
Number
of Obs.
Absolute value of
Table
Cost instruments
Mid
Hi
Other speed instruments
Independent
128MB PC100
of each quality level
8:
^-statistics in parentheses.
*
denotes significance at the
Instrumental variables estimates of
5%
level.
PC100 128MB Memory Demand Model
Product Category
MB
Memory
PC100 PC133
Quality
Low
Memory
PC133
4.2%
16.3%
24.3%
12.6%
2.9%
19.9%
24.9%
15.8%
-3.4%
-4.4%
27.2%
7.6%
-2.4%
15.6%
26.9%
11.5%
9.5%
2.8%
9.3%
14.7%
6.3%
$5.06
$10.25
$17.63
$22.41
$5.62
$12.40
-0.7%
17.0%
Mid
Hi
Overall
Average
in $
The
table presents revenue-weighted
sites
A, B, and
C
in each of six
Table
9:
MB
PC100
128
Mean
256
CPU/fan combos
SiteB
mean percentage markups
product categories.
for
products sold by web-
percentage markup in six product classes
51
SiteC
Predicted
site
B
Predicted profits
c = 78
Base costs
quantities
Rank
Price
1
68
69
70
72
74
74
74
75
77
77
78
78
2
3
4
5
6
7
8
9
10
11
12
QLow
QMid
QHi
SiteB
54.4
5.3
2.8
-5
29.9
3.9
2.3
19.5
3.1
2.1
14.1
2.6
1.9
10.7
2.3
1.8
8.5
2.0
1.6
7.0
1.8
1.6
92
122
145
154
141
130
5.9
1.7
1.5
127
5.0
1.5
1.4
12!)
4.4
1.4
1.4
121
3.9
1.3
1.3
3.4
1.3
1.3
118
112
c
=
68
A
SiteB
SiteB
147
229
246
264
267
244
225
218
217
204
198
-343
275
259
239
235
227
202
184
175
188
S'l
Site
-95
-2
55
85
85
s.|
88
96
92
92
172
160
154
145
table presents predicted sales and profits that website B would have earned if it had
occupied each position on the Pricewatch list pictured in Figure 1 throughout the day on
The
October
12, 2000.
Table
10:
Predicted profits as a function of rank: October
52
12,
2000
^
•
<-\
-1
r -
MIT LIBRARIES
3 9080 02618 0114