U.S. patent application number 14/201514 was filed with the patent office on 2014-09-04 for bid optimization in search engine marketing.
This patent application is currently assigned to Yahoo! Inc.. The applicant listed for this patent is Yahoo! Inc.. Invention is credited to Pavel Berkhin, Usama M. Fayyad, Scott Gaffney, Bassel Ojjeh, Rajesh Girish Parekh, Andrew Tomkins.
Application Number | 20140249914 14/201514 |
Document ID | / |
Family ID | 39499382 |
Filed Date | 2014-09-04 |
United States Patent
Application |
20140249914 |
Kind Code |
A1 |
Berkhin; Pavel ; et
al. |
September 4, 2014 |
BID OPTIMIZATION IN SEARCH ENGINE MARKETING
Abstract
Methods and apparatus are described for optimally allocating an
online advertising budget for a search engine marketing (SEM)
campaign among a fixed set of keywords.
Inventors: |
Berkhin; Pavel; (Sunnyvale,
CA) ; Fayyad; Usama M.; (Sunnyvale, CA) ;
Gaffney; Scott; (Menlo Park, CA) ; Ojjeh; Bassel;
(Palo Alto, CA) ; Parekh; Rajesh Girish; (Mountain
View, CA) ; Tomkins; Andrew; (San Jose, CA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Yahoo! Inc. |
Sunnyvale |
CA |
US |
|
|
Assignee: |
Yahoo! Inc.
Sunnyvale
CA
|
Family ID: |
39499382 |
Appl. No.: |
14/201514 |
Filed: |
March 7, 2014 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
11609782 |
Dec 12, 2006 |
8712832 |
|
|
14201514 |
|
|
|
|
Current U.S.
Class: |
705/14.48 |
Current CPC
Class: |
G06Q 10/06 20130101;
G06Q 30/0244 20130101; G06Q 30/0275 20130101; G06Q 30/02 20130101;
G06Q 30/0249 20130101; G06Q 10/04 20130101 |
Class at
Publication: |
705/14.48 |
International
Class: |
G06Q 30/02 20060101
G06Q030/02 |
Claims
1. (canceled)
2. A computer-implemented method of allocating an advertising
budget among a set of keywords, each keyword having a bid, a bid
intensity, and a utility associated therewith, the method
comprising: selecting ones of the set of keywords based, at least
in part, upon the utilities of the keywords; raising the bid
intensities associated with the selected ones of the set of
keywords such that the advertising budget is not exceeded; and
raising the bids associated with first ones of the selected
keywords based, at least in part, upon whether the bid intensities
associated with the selected keywords reach maximum values.
3. The computer-implemented method of claim 2, wherein raising the
bids associated with the first ones of the selected keywords is
performed such that the advertising budget is not exceeded.
4. The computer-implemented method of claim 2, further comprising:
reducing the intensities of second ones of the set of keywords in
conjunction with raising the bids associated with the first
selected keywords to ensure that the advertising budget is not
exceeded.
5. The computer-implemented method of claim 2, further comprising:
reducing the bids associated with one or more of the first ones of
the selected keywords based, at least in part, upon whether raising
the bids associated with the first ones of the selected keywords
resulted in an increase in revenue.
6. The computer-implemented method of claim 2, wherein selecting
ones of the set of keywords comprises selecting keywords having the
highest utilities among the set of keywords.
7. The computer-implemented method of claim 2, further comprising:
ranking the set of keywords according to the associated utilities;
wherein selecting ones of the set of keywords includes selecting
keywords from the set of keywords based, at least in part, upon the
ranking.
8. The computer-implemented method of claim 2, further comprising:
ranking the selected keywords according to the associated
utilities; and identifying the first ones of the selected keywords
based, at least in part, upon the ranking.
9. The computer-implemented method of claim 2, wherein raising the
bids associated with the first selected keywords comprises raising
the bids associated with the first selected keywords until the
utilities associated with the first selected keywords are
substantially in equilibrium.
10. The computer-implemented method of claim 2, further comprising
generating a budget allocation based, at least in part, upon
statistics derived from raising the bids and bid intensities.
11. The computer-implemented method of claim 10, further
comprising: communicating the budget allocation to an entity acting
on behalf of an advertiser.
12. The computer-implemented method of claim 2, wherein the utility
of each keyword is derived based, at least in part, upon a
corresponding conversion rate and the associated bid.
13. A system for allocating an advertising budget among a set of
keywords, each keyword having a bid, a bid intensity, and a utility
associated therewith, the system comprising a processor and
non-transient memory with program logic for execution on the
processor, the program logic comprising: a keyword selection module
configured to select ones of the set of keywords based, at least in
part, upon the utilities of the keywords; an intensity raising
module configured to raise the bid intensities associated with the
selected ones of the set of keywords such that the advertising
budget is not exceeded; and a bid raising module configured to
raise the bids associated with first ones of the selected keywords
based, at least in part, upon whether the bid intensities
associated with the selected keywords reach maximum values.
14. The system of claim 13, the program logic further comprising:
an intensity reduction module configured to reduce the intensities
of second ones of the set of keywords in conjunction with raising
the bids associated with the first selected keywords to ensure that
the advertising budget is not exceeded.
15. The system of claim 13, the program logic further comprising: a
bid reduction module configured to reduce the bids associated with
one or more of the first ones of the selected keywords based, at
least in part, upon whether raising the bids associated with the
first ones of the selected keywords resulted in an increase in
revenue.
16. The system of claim 13, the program logic further comprising: a
budget generation module configured to generate a budget allocation
based, at least in part, upon statistics derived from raising the
bids and bid intensities.
17. The system of claim 13, wherein the utility of each keyword is
derived based, at least in part, upon a corresponding conversion
rate and the associated bid.
18. A system for allocating an advertising budget among a set of
keywords, each keyword having a bid, a bid intensity, and a utility
associated therewith, the system comprising: means for selecting
ones of the set of keywords based, at least in part, upon the
utilities of the keywords; means for raising the bid intensities
associated with the selected ones of the set of keywords such that
the advertising budget is not exceeded; and means for raising the
bids associated with first ones of the selected keywords based, at
least in part, upon whether the bid intensities associated with the
selected keywords reach maximum values.
19. The system of claim 18, further comprising: means for reducing
the intensities of second ones of the set of keywords in
conjunction with raising the bids associated with the first
selected keywords to ensure that the advertising budget is not
exceeded.
20. The system of claim 18, further comprising: means for reducing
the bids associated with one or more of the first ones of the
selected keywords based, at least in part, upon whether raising the
bids associated with the first ones of the selected keywords
resulted in an increase in revenue.
21. The system of claim 18, wherein the utility of each keyword is
derived based, at least in part, upon a corresponding conversion
rate and the associated bid.
Description
RELATED APPLICATIONS
[0001] This application is a continuation of U.S. patent
application Ser. No. 11/609,782, entitled "Bid Optimization in
Search Engine Marketing," by Berkhin et al, filed on Dec. 12, 2006,
which is incorporated herein by reference in its entirety for all
purposes.
BACKGROUND OF THE INVENTION
[0002] The present invention relates to search engine marketing
and, in particular to techniques for optimizing bids for a set of
keywords associated with an online marketing campaign.
[0003] Search Engine Marketing (SEM) constitutes a significant
portion of the online advertising market. Among other things, SEM
involves selecting relevant keywords and initiating and maintaining
advertising campaigns targeting search engine users entering the
keywords in their search queries. The search engine selects
appropriate links to relevant content, and also places sponsored
links on the search results page. In order to be included among the
sponsored links, an advertiser pays an amount of money to the
search engine operator for every user click, i.e., the so called
keyword bid.
[0004] Unfortunately, there currently are no systematic techniques
for efficiently allocating an advertising budget over a suitably
constrained set of keywords. Typically, such allocations are done
without any real empirical basis. Instead, advertisers use a
variety of approaches from exercising a "gut feeling," to simply
bidding equally on every keyword. The shortcomings of such
approaches are manifest.
[0005] More effective techniques for allocating an advertising
budget over a set of keywords are therefore desirable.
SUMMARY OF THE INVENTION
[0006] According to various embodiments, methods and apparatus are
provided for allocating an advertising budget among a fixed set of
keywords. Each keyword has a bid, a bid intensity, and a utility
associated therewith. The bid intensities associated with selected
ones of the keywords are raised such that the advertising budget is
not exceeded. The selected keywords have the highest utilities
among the fixed set of keywords. When the bid intensities
associated with the selected keywords reach maximum values, the
bids associated with first ones of the selected keywords are raised
such that the advertising budget is not exceeded. The first
selected keywords have the highest utilities among the selected
keywords.
[0007] According to a specific embodiment, the bid intensities
associated with second ones of the keywords are lowered in
conjunction with raising the bids associated with the first
selected keywords to ensure that the advertising budget is not
exceeded.
[0008] According to another specific embodiment, the bid associated
with each keyword is initially set at a minimum bid which
guarantees appearance of a link for the associated keyword among
sponsored search links associated with search results.
[0009] According to yet another specific embodiment, the
intensities associated with the keywords are initially set to a
uniform intensity such that the advertising budget is not exceeded
by the minimum bids.
[0010] According to a further specific embodiment, statistics are
accumulated for a subset of the keywords representing a conversion
rate for each.
[0011] According to yet a further specific embodiment, the utility
for each of the subset of keywords is derived with reference to the
corresponding conversion rate and the associated bid.
[0012] According to a still further specific embodiment, the
selected keywords are ranked according to the associated
utilities.
[0013] According to another embodiment, a portion of the
advertising budget is reserved for further evaluation of the
utilities associated with a subset of the keywords not included
among the first selected keywords.
[0014] According to still another embodiment, it is determined
whether raising the bids associated with the first selected
keywords satisfies an optimization condition such that a change in
revenue associated with raising the bids is greater than zero.
[0015] According to an additional embodiment, the bids associated
with the first selected keywords are raised until the utilities
associated with the first selected keywords are substantially in
equilibrium.
[0016] According to a further additional embodiment, a budget
allocation, is communicated to an entity acting on behalf of an
advertiser. The budget allocation is derived with reference to
statistics derived from raising the bids and bid intensities.
[0017] A further understanding of the nature and advantages of the
present invention may be realized by reference to the remaining
portions of the specification and the drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
[0018] FIG. 1 is a flowchart illustrating operation of a specific
embodiment of the present invention.
[0019] FIG. 2 is a simplified diagram of a network environment in
which specific embodiments of the present invention may be
implemented.
DETAILED DESCRIPTION OF SPECIFIC EMBODIMENTS
[0020] Reference will now be made in detail to specific embodiments
of the invention including the best modes contemplated by the
inventors for carrying out the invention. Examples of these
specific embodiments are illustrated in the accompanying drawings.
While the invention is described in conjunction with these specific
embodiments, it will be understood that it is not intended to limit
the invention to the described embodiments. On the contrary, it is
intended to cover alternatives, modifications, and equivalents as
may be included within the spirit and scope of the invention as
defined by the appended claims. In the following description,
specific details are set forth in order to provide a thorough
understanding of the present invention. The present invention may
be practiced without some or all of these specific details. In
addition, well known features may not have been described in detail
to avoid unnecessarily obscuring the invention.
[0021] Embodiments of the present invention address the following
problem: Given a budget, how to properly allocate the budget among
available keywords. According to specific embodiments, given a
fixed set of keywords, there is an optimal bidding strategy based
on a concept referred to herein as "keyword utility" that maximizes
both revenue and return on investment (ROI) for a fixed budget.
According to such embodiments, the optimal allocation is reached by
raising bids and intensities for the strongest keywords until you
get to a stationary point of (approximately) equal utilities under
the given budget constraint. It should be noted that, though in
reality search engine operators usually charge advertisers slightly
less than their (maximum) bids, this factor is disregarded in the
following examples for the sake of simplicity. However, embodiments
are contemplated in which this factor is taken into account.
[0022] It should also be noted that the set of keywords to which
the techniques of the present invention are applied may be derived
in a wide variety of ways. For example, keywords may be selected
according to marketing research or by observing bidding behavior of
similarly situated advertisers. According to one class of
embodiments, a keyword set may be derived using the techniques
described in U.S. patent application Ser. No. 11/444,996 for
KEYWORD SET AND TARGET AUDIENCE PROFILE GENERALIZATION TECHNIQUES
filed on May 31, 2006 (Attorney Docket No. YAH1P016), the entire
disclosure of which is incorporated herein by reference for all
purposes.
DEFINITIONS AND NOTATIONS
[0023] The following concepts and notations will be used herein to
describe specific embodiments of the invention. As used herein, the
term "keyword" relates to individual words as well as phrases, and
is used interchangeably with the term "query." An advertiser
defines a set W={w} of keywords related to its business. The
advertiser then bids on these keywords in, for example, the SEM
context to place sponsored search links whenever users enter a
keyword from the set (e.g., in a search query) into a search
engine.
[0024] W a set of k keywords which may be enumerated, W={w.sub.1,
w.sub.2, . . . , w.sub.k}.
[0025] f(w) is the probability of keyword w as a search query among
all queries, i.e., query frequency; f.sub.i=f(w.sub.i).
[0026] b(w) is a bid or bid price for keyword w as set by an
advertiser; b.sub.i=b(w.sub.i). The bidding results in a particular
position or rank j=j(w, b) for w among other sponsored search
results for a keyword w.
[0027] C.sup.SS (w) is a click through rate (CTR) for a keyword w;
C.sub.i.sup.SS=C.sup.SS). Click through rate depends on the
position j among other sponsored links. Positions with smaller,
i.e., more favorable, j result in a higher CTR. So, the click
through rate C.sup.SS (w) monotonically depends on the bid price
b(w). According to a specific embodiment, the click through rate is
assumed to be the number of clicks divided by the number of
impressions. The term "impression" refers to the presentation of
the sponsored link in a sponsored link list associated with search
results (as opposed to the click which refers to actual selection
of the link by the user).
[0028] K(w) is the conversion rate, i.e., the fraction of clicks
which result in a desirable event; K.sub.i=K(w.sub.i). The higher
the relevance of the advertisement to a keyword w, the higher is
the value of K(w). The conversion rate typically depends on
relevancy to the keyword and the effectiveness of the landing page
and not, in general, on the position of the link. Thus, this
parameter is different in this regard from CTR which strongly
depends on position.
[0029] g is a revenue associated with a single conversion.
[0030] Reasonable practical assumptions may be made about
introduced quantities. For example, query frequencies are known for
frequent queries, and there are reasonable ways to construct a
viable proxy for infrequent queries. For example, we may set f(w)
equal to some small default value for a particular category of
infrequent queries. In another example, the dependence of b(w) on
other advertisers bids (due to a process known as "bubbling") may
be disregarded.
[0031] In addition, it can be argued that the conversion rate K(w)
does have some dependence on link position. That is, there is some
evidence that users who select sponsored search links placed
directly above "organic" search results are less experienced users
than those who select sponsored search links placed to the right of
the organic search results list. It has been assumed that these
inexperienced users may think that the sponsored links are part of
the organic results, and that therefore, the conversion rate for
such clicks is lower. Nevertheless, it is a useful simplifying
assumption to assume no dependence on position.
[0032] Problem Formulation
[0033] To illustrate the scope of the problem, we first assume that
an advertiser is interested in maximizing only one thing--return on
investment (ROI). Stating this assumption another way, the
advertiser's focus is on the price per conversion rather than the
number of conversions. In this scenario, the advertiser should bid
the smallest bid price (i.e., the least expensive sponsored search
link position) for the single "best" keyword w. The definition of
"best" in this context is discussed below. Because the best keyword
is used, the ROI for this approach will be the highest.
Unfortunately, because the low bid price results in an unfavorable
positioning for the sponsored search link, the number of
conversions, and consequently total revenue for the campaign, will
be extremely low. Thus, this model and its equivalents (e.g.,
cheapest price-per-acquisition models) are not viable.
[0034] To illustrate the other extreme, we instead assume that the
advertiser is only interested in maximizing revenues, or
equivalently the number of clicks and conversions. The clear way to
achieve this is to bid the highest bids that would secure the
highest position (e.g., the first position above the organic search
results) for every keyword w in Was soon as the price per
acquisition is below the revenue per conversion. Under such a
bidding strategy, the number of conversions, and consequently
revenues for the campaign, is maximized. Unfortunately, the
advertising budget grows uncontrollably. That is, the ROI of such a
strategy would be unacceptably low from the perspective of most
advertisers.
[0035] The foregoing examples illustrate that an advertiser cannot
construct a campaign which results in both the best possible ROI
and the highest possible revenue. However, it is possible, in
accordance with embodiments of the present invention, to achieve
the best ROI and the highest revenue (or reach) for a fixed
advertising budget. This can be understood from the simple
formula:
R O I = Revenue Budget . ##EQU00001##
When the denominator of this expression is fixed, the maximization
of ROI and of Revenue becomes an equivalent problem.
[0036] Optimal Bid Allocation
[0037] According to specific embodiments of the invention,
techniques are provided for developing an optimal bid allocation or
distribution in which, for a set of keywords w.sub.i, rational bids
b.sub.i are determined.
[0038] Imagine a user initiates a search using the query w, and a
query search results page (let us assume the first page) is
returned having one sponsored search link impression. In our
notation, the average revenue per one w search is given by:
Revenue per one w=C.sup.SS(w)K(w)g (1)
and the average bidding cost per one w search is given by:
Cost per one w=C.sup.SS(w)b(w). (2)
[0039] Both quantities should be multiplied by query frequency f(w)
if we want w-revenue and w-cost from single average user search. It
then follows that ROI per one w search is given by:
R O I ( w ) = C SS ( w ) K ( w ) g C SS ( w ) b ( w ) = K ( w ) g b
( w ) . ( 3 ) ##EQU00002##
[0040] We will refer to ROI(w) as the keyword utility. From (3) we
see that ROI actually does not depend on click through rate. In
addition, the conversion revenue g is independent of a keyword w
and can therefore be omitted if particular w optimization is
involved. Thus, the quantity
R O I ' ( w ) = K ( w ) b ( w ) ( 4 ) ##EQU00003##
can be interpreted as the reach per amount of money spent, thereby
establishing the equivalence of revenues and conversion volume. Let
us now sort all keywords so that
ROI(w.sub.i).gtoreq.ROI(w.sub.2).gtoreq. . . .
.gtoreq.ROI(w.sub.k).
[0041] Now to formalize the concept of a budget. As used herein,
budget B refers to the amount of money spent per one average
search. It is combined from different bid costs over a variety of
keywords w that happen during the search along with their
probabilities:
B=.SIGMA..sub.w.epsilon.WC.sup.SS(w)b(w)f(w)u(w) (B)
In this formula u(w) is the intensity of the w-bid such that
0.ltoreq.u(w).ltoreq.1. Bid intensity refers to a fraction of time
to bid on the corresponding keyword. For example, an advertiser may
choose not to bid on a particular w at all (i.e., u(w)=0), to bid
all the time, i.e., on every occurrence of w (i.e., u(w)=1), or to
do something in between. A crude approach to controlling intensity
would be to put a bound on the number of impressions. However, a
wide variety of approaches may be used, and the particular
mechanism by which intensity is controlled is not relevant to the
scope of the invention.
[0042] If the advertiser bids for the cheapest position for a most
profitable keyword w.sub.1 and budget is so small that
B<C.sup.SSb.sub.1f.sub.1, then no other strategy can be better
(note that in this case the intensity
u(w)=C.sub.1.sup.SSb.sub.1f.sub.1/B). Thus, in this scenario, the
best approach for the advertiser is to exhaust the budget on
w.sub.i up to the fullest possible intensity.
[0043] In reality, such a strategy would result in very few clicks.
Therefore, to better illustrate the invention, a more realistic
budget will be assumed such that B>C.sub.1.sup.SSb.sub.1f.
Assuming an initial bid on w.sub.1 having full intensity (i.e.,
u.sub.1=1), what should be done with the remainder of the budget B?
There are three alternatives: (1) to increase bid b.sub.1, (2) to
blend the bid on w.sub.1 with bids on other keywords, or (3) to do
both.
[0044] Recall that an advertiser always can bid more or less for
specific keywords. Higher bids results in more a favorable position
j in the sponsored search links and therefore a higher click
through rate. This, in turn, results in increased revenue. However,
higher bids also result in a decrease in ROI given the higher cost
per acquisition. Instead, embodiments of the invention enable
strategies that, given a fixed budget B, result in the highest
revenue. As discussed above, we know that such approaches will also
have the maximum ROI among all strategies that have the identical
budget B.
[0045] An example will be illustrative. For simplicity assume we
have only two keywords and that a current allocation between the
two keywords fits a budget B. We will also assume initially that
u.sub.1=1. Using our notation, this may be represented as:
C.sub.1.sup.SSb.sub.1f.sub.1+C.sub.2.sup.SSb.sub.2f.sub.2u.sub.2=B.
[0046] As we know, the ROI for this budget distribution is given
by:
R O I = g C 1 SS K 1 f 1 + C 2 SS K 2 f 2 u 2 C 1 SS b 1 f 1 + C 2
SS b 2 f 2 u 2 = g ( C 1 SS K 1 f 1 + C 2 SS K 2 f u 2 ) / B ,
##EQU00004##
where the numerator times g is simply a corresponding revenue.
[0047] The goal in this example is to find the condition under
which bid b.sub.1 should be increased. And because B is fixed, an
increase in b.sub.1 means a simultaneous decrease in intensity
u.sub.2 associated with the bidding on w.sub.2 in order to preserve
budget balance. Assume that the smallest (empirically determined)
feasible increase of the current bid b.sub.1 is given by
b.sub.1,new=b.sub.1+.DELTA.b.sub.1. This results in .DELTA.C.sub.1
increase of click through rate (assuming that the increase in
b.sub.1 was sufficient to get to the next better position). This,
in turn, will increase the budget by .delta., necessitating a
reduction of an equal amount which may be achieved by decreasing
the intensity of the bid on w.sub.2 from u.sub.2 to
u.sub.2,new=u.sub.2-.DELTA.u.sub.2 such that:
.delta.=(C.sub.1.sup.SS+.DELTA.C.sub.1.sup.SS)(b.sub.1+.DELTA.b.sub.1)f.-
sub.1-C.sub.1.sup.SSb.sub.1f.sub.1
.delta.=C.sub.2.sup.SSb.sub.2f.sub.2.DELTA.u.sub.2 (5)
[0048] From this we get two equations, the first representing
redistributing the budget to keep it constant, and the second
representing new revenue (we use symbol R for revenues):
(C.sub.1.sup.SS+.DELTA.C.sub.1.sup.SS)+(b.sub.1+.DELTA.b.sub.1)f.sub.1+C-
.sub.2.sup.SSb.sub.2f.sub.2(u.sub.2-.DELTA.u.sub.2)B
R.sup.new=(C.sub.1.sup.SS+.DELTA.C.sub.1.sup.SS)K.sub.1f.sub.1+C.sub.2.s-
up.SSK.sub.2f.sub.2(u.sub.2-.DELTA.u.sub.2)=R+.DELTA.R, (6)
where
.DELTA.R=.DELTA.C.sub.1.sup.SSK.sub.1f.sub.1-C.sub.2.sup.SSK.sub.2f.sub.-
2.DELTA.u.sub.2. (7)
[0049] From (5) we find that:
f 2 .DELTA. u 2 = .delta. C 2 SS b 2 ( 8 ) ##EQU00005##
and
.delta.=.DELTA.C.sub.1.sup.SS(b.sub.1+.DELTA.b.sub.1)f.sub.1+C.sub.1.sup-
.SS.DELTA.f.sub.1=.DELTA.C.sub.1.sup.SSb.sub.1.sup.newf.sub.1+C.sub.1.sup.-
SS.DELTA.b.sub.1f.sub.1
or equivalently:
.DELTA. C 1 SS = .delta. b 1 new f 1 - C 1 SS .DELTA. b 1 b 1 new (
9 ) ##EQU00006##
[0050] Notice that (for small .DELTA.b.sub.1>0), the positive
sign of .delta. defines the sign of .DELTA.C.sub.1.sup.SS; this is
why monotonic dependency of the click through rate on bid is
important. Substituting (8) and (9) into (7), we see that the
impact on revenue R for a small budget reallocation is:
.DELTA. R = .DELTA. C 1 SS K 1 f 1 - C 2 SS K 2 f 2 .DELTA. u 2 = (
.delta. b 1 new f 1 - C 1 SS .DELTA. b 1 b 1 new ) K 1 f 1 - C 2 SS
K 2 ( .delta. C 2 SS b 2 ) = .delta. ( K 1 b 1 new - K 2 b 2 ) - C
1 SS K 1 .DELTA. b 1 f 1 b 1 new ( 10 ) ##EQU00007##
[0051] Now we will rationalize this formula. Our assumption was
that
K 1 b 1 > K 2 b 2 . ##EQU00008##
When .DELTA.b.sub.1 is small, the second term containing
.DELTA.b.sub.1 is small, while the difference
K 1 b 1 new - K 2 b 2 ##EQU00009##
is still positive. We may therefore conclude that a small increase
in bid price for the keyword with the larger ROI index offset by a
corresponding reduction in the intensity associated with the
keyword having the lower ROI results in a positive revenue impact
.DELTA.R>0. Therefore, an optimal bidding strategy corresponds
to an equilibrium budget allocation such that:
K 1 b 1 .apprxeq. K 2 b 2 .apprxeq. .apprxeq. K k b k . ( 11 )
##EQU00010##
[0052] Since in reality a bid price cannot be changed continuously,
when .DELTA.b.sub.1/b.sub.1.sup.new is not small, formula (10)
provides a condition when the increase in bid b.sub.1 is justified.
That is, an increase should be attempted if:
.delta. ( K 1 b 1 new - K 2 b 2 ) - C 1 SS K 1 .DELTA. b 1 f 1 b 1
new > 0 ( 12 ) ##EQU00011##
The condition for justifying a decrease may be similarly derived.
It should be noted that, while the foregoing example was described
with reference to the case involving two keywords, the technique
described may be readily generalized to cases involving any number
of keywords, as well as to sets of keywords.
[0053] According to a specific embodiment, a simple and efficient
algorithm is provided to derive the optimal allocation under the
model described above which does not require explicit computation
of equation (12). First, observe that an advertiser should never
raise a bid on a keyword until the intensity for the current bid
has reached 1, i.e., until the advertiser is bidding on each
occurrence of the keyword. Then observe that the impact of setting
the intensity of the increased bid for that keyword to a non-zero
value is that for some fraction of occurrences of the keyword, the
sponsored link will be shown at a more favorable position (with
higher expected revenue and lower ROI). That is, the advertiser
will bid on the same overall number of impressions for the keyword,
but some of the impressions will generate a higher click through
rate, at the same conversion rate, but at higher cost per click.
Imagine, for example, that the user bids on 200 impressions per
hour of the keyword, which generates two clicks, and the user pays
a penny for each click. The user has additional budget to apply,
and increases the intensity of the next higher bid for the keyword,
i.e., the bid required to place a sponsored search link at the next
higher rank. As a result, the user now bids low on 100
impressions/hr, and higher on the other 100 impressions/hr. The
first batch of impressions again generates one click, at cost of
one penny. But the next batch now generate three clicks, at a cost
of two pennies per click.
[0054] In general, each impression has an expected number of
conversions, x, and an expected cost, y. By raising the intensity
of the higher bid, certain of the impressions (and in fact all of
the impressions once the intensity of the higher bid reaches l)
will generate ax conversions, at an expected cost by per
conversion. Thus, each impression processed at the higher cost will
result in an additional (a-l)x conversions at an additional cost of
(b-l)y per conversion. A virtual keyword with these properties may
be created which the user can bid on independently of the keyword
under consideration. The optimal solution to the model above may
then be attained by greedily selecting real and virtual keywords
according to this scheme, without ever considering raising a
bid--any bid raises are modeled simply by the purchase of virtual
keywords.
[0055] Once the intensity for a virtual keyword reaches one, a new
virtual keyword may be introduced to capture the increase in both
conversions and cost that result from moving from rank j-1 to rank
j-2. Or if desired, a virtual keyword may be introduced for each
keyword and each rank. The algorithm will perform identically
whether the virtual keywords are introduced up front or in a lazy
manner as the greedy selection proceeds. Further, this greedy
algorithm allows either a discrete or a continuous version of the
incremental bidding on a keyword.
[0056] We will now describe with reference to FIG. 1 an example of
a process for allocating an advertising budget among keywords
according to a specific embodiment of the invention. Notice that
our definition of budget B is that it is equal to the cost per
average search which can be computed from a daily advertising
budget by dividing the daily budget by the daily number of
searches. Notice also that for many tail queries a minimum bid
results in a single sponsored search link. Therefore, for such
queries we do not necessarily need to waste time on experiments
beyond accumulating aggregate statistics. This does not mean that
the basic strategy for such queries is different, but simply that
application of the strategy may not result in any change except a
potential change in intensity.
[0057] Initially, each bid b.sub.i is set to the minimum value that
guarantees appearance of a link among the sponsored search links on
the first (or a sufficiently high enough) search page for each
keyword w.sub.i in a fixed set of keywords W(102). Uniform
intensities u.sub.i=const are set to guarantee that we stay within
the budget (104).
[0058] Statistics C.sup.SS(w) and K(w) are then accumulated (106).
Since keywords are different in terms of their relevance and in
terms of other advertiser competition, after a while, differences
will be identified among sample conversion rates K.sub.i that is a
ratio of the number of conversions over the number of clicks. Since
the confidence level is initially low, a conservative low-bound
estimate is used.
[0059] When sufficient statistics are available, the keywords are
sorted in order of decreasing utility ROI.sub.i=gK.sub.i/b.sub.i
defined by (3) such that ROI(w.sub.1).gtoreq.ROI (w.sub.2).gtoreq.
. . . .gtoreq.ROI(w.sub.k) (108). Keywords having high utilities is
where most of the budget is then focused (110). It should be noted
that, initially, only a few ROIs will be positive while the ROIs
associated with the tail keywords will be zero.
[0060] For the top few keywords, i.e., i=1, 2, 3, . . . , I, the
intensities are then gradually increased (112), moving the u.sub.i
as close to 1 as possible with the constraint being to keep total
spending on these top few keywords under pB, where p defines a
fraction of the budget (e.g., 0.9) used in exploitation. According
to a specific embodiment, a portion of budget B (i.e., 1-p)B is
reserved for other keywords (e.g., those in the tail or additional
keywords) for the purpose of monitoring their statistics for
possible future use. For these keywords the intensities remain
low.
[0061] If the high-intensity, least-bid distribution on the few
high-utility keywords consumes the full exploitation budget pB
(114), this is considered a desirable outcome, and the advertising
campaign may continue to be monitored without taking any immediate
action (116). In reality, the majority of advertisers are likely to
actively look for new keywords or attempt higher bids on existing
ones in order to get more traffic.
[0062] If the high-intensity, least-bid distribution reaches full
intensity, i.e., u.sub.i=1, and the available exploitation budget
pB is not reached (114), more of the budget may be introduced into
the campaign, i.e., the bids b.sub.i may be increased. According to
a specific embodiment, the bids b.sub.i are incrementally increased
for the keywords with the highest utilities (118). Each time this
is done, it is then determined whether revenues actually increase
(120), e.g., as described above with reference to equation (12). If
an actual increase is not realized, the bid(s) is (are) returned to
previous level(s).
[0063] When the exploitation budget pB is fully consumed (122) bids
can no longer be increased without offsetting the increases by
lowering intensities on other bids. That is, one or more of the
bids for specific high utility keywords may continue to be
increased (124), but each such increase is then offset by a
decrement in the intensity of bids on lower utility keywords (126).
The specific amount of such decrements may be derived from equation
(8) for the model case of two keywords. As will be understood with
reference to the foregoing description, this keeps the budget
equation (B) in balance.
[0064] After one or more bids of the high utility keywords are
increased and offset by decrements of the intensities associated
with lower utility keywords, the optimization condition represented
by equation (12) is monitored. If the condition is violated (128),
the increase is reversed (130).
[0065] Eventually, the increases to the bids for the high-utility,
full-intensity (u.sub.i=1) keywords brings this keywords into a
substantial equilibrium (132) represented above by equation
(11),
K 1 b 1 .apprxeq. K 2 b 2 .apprxeq. .apprxeq. K k b k .
##EQU00012##
After this, the campaign may be monitored (116), and/or more or
different keywords may be introduced (134) and the process
repeated.
[0066] The lower utility keywords may be maintained with low
intensities and be evaluated using the exploration portion of the
budget, e.g., (1-p)B, in order to keep eye on their statistics.
According to a specific embodiment, a relatively high proportion of
the budget may initially be allocated to exploration with the
portion allocated for exploitation being increased over time, i.e.,
p=p(t) is an increasing function of time t, p'(t)>0. In
addition, it is contemplated that keyword sampling, including
low-utility and low-frequency keywords, may continue
indefinitely.
[0067] It should be noted that undesirably high bids can happen for
at least two reasons. First, the budget B may be so large that to
entirely consume it requires bids to be so high that profitability
deteriorates to zero, i.e., the so-called keyword inventory
problem. Second, the keywords themselves may be so irrelevant (in
terms of a lack of conversions K.sub.i) that even reasonable bids
violate the profitability requirement. In general, it is desirable
to constrain keyword utilities gK.sub.i/b.sub.i from going below
one or the advertiser will lose money. Put another way, increases
in bids have a limit defined by:
gk.sub.i/b.sub.i>1 (13)
[0068] It should be noted, that this condition should not be
violated during various phases of the above described
technique.
[0069] An interesting consequence of the approach described above
is that an advertiser may choose to ignore click through rates when
allocating its advertising budget among keywords. However, click
through rates may yet be important in at least one regard. That is,
advertisers are definitely interested in developing more creative
and effective landing pages for individual keywords or groups of
keywords that affect both conversion rates and click through rates.
Therefore, according to specific embodiments of the invention,
where campaign management resources are limited, prioritization of
keywords with respect to landing page improvement may be done based
on the index: C.sup.SS(w)K(w)f(w). The larger the index, the
greater the urgency to review its landing page.
[0070] Embodiments of the present invention may be employed to
facilitate allocation of an advertising budget over keywords for an
online advertising campaign in any of a wide variety of computing
contexts. For example, as illustrated in FIG. 6, implementations
are contemplated in which the relevant population of users interact
with a diverse network environment via any type of computer (e.g.,
desktop, laptop, tablet, etc.) 602, media computing platforms 603
(e.g., cable and satellite set top boxes and digital video
recorders), handheld computing devices (e.g., PDAs) 604, cell
phones 606, or any other type of computing or communication
platform.
[0071] And according to various embodiments, user data processed in
accordance with the invention may be collected using a wide variety
of techniques. For example, collection of data representing a
user's interaction with a search engine interface and the
associated sponsored links, landing pages, and web sites may be
accomplished using any of a variety of well known mechanisms for
recording a user's online behavior. Once collected, the user data
may be processed in order to facilitate budget allocation according
to the invention in a centralized manner. This is represented in
FIG. 6 by server 608 and data store 610 which, as will be
understood, may correspond to multiple distributed devices and data
stores. The budget allocation process may be performed by
representatives of individual advertisers, by representatives of
search providers (e.g., Yahoo! Inc.), or by representatives of
third party advertising services. In the latter two cases,
recommendations may then be made to advertisers about how to
allocate their advertising budgets, or the campaigns could be
initiated and run on their behalves.
[0072] The various aspects of the invention may also be practiced
in a wide variety of network environments (represented by network
612) including, for example, TCP/IP-based networks,
telecommunications networks, wireless networks, etc. In addition,
the computer program instructions with which embodiments of the
invention are implemented may be stored in any type of
computer-readable media, and may be executed according to a variety
of computing models including a client/server model, a peer-to-peer
model, on a stand-alone computing device, or according to a
distributed computing model in which various of the functionalities
described herein may be effected or employed at different
locations.
[0073] While the invention has been particularly shown and
described with reference to specific embodiments thereof, it will
be understood by those skilled in the art that changes in the form
and details of the disclosed embodiments may be made without
departing from the spirit or scope of the invention. For example,
some of the statistics mentioned above may not be available in all
circumstances. However, it will be understood that a variety of
alternative metrics and substitutes may be employed without
departing from the scope of the invention. For example, conversion
events may be registered by a variety of techniques (e.g.,
so-called beacons), be reported by customers directly, or, in
certain cases, be identified with click through rates.
[0074] In addition, although various advantages, aspects, and
objects of the present invention have been discussed herein with
reference to various embodiments, it will be understood that the
scope of the invention should not be limited by reference to such
advantages, aspects, and objects. Rather, the scope of the
invention should be determined with reference to the appended
claims.
* * * * *