OptimizingSEMCampaigns

SEM Campaigns on Auto Pilot
Vijay Ramachandran and Seetharamakrishna M
Commerce Central, Yahoo Small Business
Abstract
We describe a set of algorithms to create and optimize automated SEM campaigns for catalogue driven E-commerce
merchants. We first describe methods to find the best landing page for a particular search term considering various
static page and visitor-specific variables, by treating this as a ranking problem. We also describe a way to determine
the initial bids, and the algorithm to tune these bids. We show that this method enables e-commerce merchants to
effectively advertise important products in their catalogue with a limited budget.
1. Introduction
The Internet has become a major advertising medium,
with billions of dollars at stake[1]. Search engine mar-
keting (SEM) is a form of Internet marketing that in-
volves the promotion of websites by increasing their
visibility in search engine results pages (SERPs)
through advertising. In an SEM marketplace, an adver-
tiser runs a campaign with the search engine provider to
have their content advertised as ‘sponsored links’ in
response to search queries. Advertisers need to decide
which keywords describe the intent of an internet
searcher to buy a product they’re selling, and bid for the
opportunity to show an ad in response to that search
query. Upon the user clicking on a particular link, that
advertiser pays an amount to the search engine for re-
ferring that traffic. Hence, SEM is an action oriented,
Pay Per Click (PPC) marketplace. E-commerce adver-
tisers typically advertise to sell the products in their
catalogue, i.e., their SEM campaigns are measured by
how effectively they maximize the money spent on PPC
ads, and sales made in the short term from the visitors
those ads refer.
SEM tends to be very competitive, with many advertis-
ers contending for qualified users [23]. An e-commerce
merchant typically carries many products in their cata-
logue, and therefore has to choose which products to
sell given a limited budget. This implies that they ini-
tially need to choose from a wide range of potential
keywords to bid on. Further, they carry many similar
products in their catalogue, and a single search phrase
can accurately match any of them. SEM marketplaces
expect the advertiser to fix the web page
associated with every keyword. Hence, another problem
faced by e-commerce merchants is to decide which
page to associate with every keyword they bid on. Once
a campaign is live, typically, market dynamics and user
behaviour cause assumptions about validity of a key-
word, amount of bid, etc. to change. Hence, these cam-
paigns need to be actively managed by changing bids
on keywords, dropping or removing keywords, etc.
Advertisers therefore face many difficult problems in
running a profitable campaign - deciding which key-
words to bid on, deciding the best landing page for each
keyword, choosing how much to bid on these keywords
and how and when to vary the bids. SEM management
tools comprise of Keyword Suggestion tools, Bid Man-
agement Tools, Competitive Analysis Tools, and Affili-
ate Marketing Tools [25]. They work well for mer-
chants with very large budgets. On the other hand,
many agencies provide SEM as a service, with humans
managing all aspects of the campaign. This works well
for advertisers with fewer pages to advertise.
Commerce Central provides various services to e-
commerce merchants, including live site-visitor analyt-
ics, and various advertising services. In this paper, we
describe how we solve the various tasks associated with
running an SEM campaign for ecommerce merchants.
First, we describe how to obtain a candidate set of key-
words to bid on, from keyword suggestion tools. We
next describe a method to score web pages from a mer-
chant’s site using various site-visitor metrics. This is
used to rank these pages, which is used in different
ways. We attempt to solve the problem of finding the
best landing page for any keyword by modeling the
expected Return on Investment (RoI) on a (keyword,
landing page) combination, and ranking the landing
pages on this expected RoI. We then describe an algo-
rithm to decide which keywords to bid on initially, and
how much to bid on each keyword, in a scenario where
the expected spend on all available keywords is greater
than the available campaign budget. We touch briefly
on the algorithm to tune these bids, which was devel-
oped at Lexity prior to their acquisition by Yahoo. Fi-
nally, we present our results.
2. Background
To help with managing various aspects of an SEM
campaign, they are organized hierarchically[13] into

1. Campaigns. The budget for the entire cam-
paign, as well as limits per day, are set at this
level. Targeting parameters, such as geo and
language, and duration of the campaign, are
also set at this level.
2. Ad Groups and Ads. Keywords that an adver-
tiser bids on, and ads shown to the customer,
are associated with ad groups. Ads are shown
to users as part of the SERP, and they must
conform to standards set by the search engine,
such as number of lines, total number of char-
acters, permissible language, etc. The Cost Per
Click (CPC) bids can be set at this level, and
apply to all keywords under this Ad Group.
Search Engines advice that the keywords and
the ads under the Ad Group be related closely,
from a semantic perspective, so that they can
be managed together.
3. Keywords and associated Landing Page. These
are the search terms that the advertiser bids on.
The match type can be chosen as exact, phrase,
or broad (which will expand the set of words
to those determined by the search engine to be
semantically similar). The advertiser can
choose to have different bids per keyword, ra-
ther than at an Ad Group level, giving them
finer grained control over their campaign.
Each keyword must be associated to exactly
one Landing Page, which is the web page on
the advertiser’s site which they intend the visi-
tor to visit upon clicking on the ad. SEM cam-
paigns typically have well defined conversion
objectives - either something is intended to be
sold, or contact information gleaned from the
user. Therefore, the choice of landing page is
very important for the campaign. Search en-
gines typically use the semantic relatedness of
the keyword and the landing page to boost
bids, as it is shown that relevant landing pages
lead to a better user experience and a virtuous
cycle for the SEM ecosystem.
In addition to the above, there are various factors which
affect the success of an SEM campaign.
1. Conversion events. These can be tracked using
web visitor tracking systems. This is an im-
portant measure of the effectiveness of the
SEM campaign.
2. The SEM Auction, and Ad Rank. The search
engine conducts an auction for every chance to
show an ad, with all qualifying advertisers (as
defined by the applicability of a bid keyword
to the user query) [15, 16]. In addition to the
bid for that keyword, the search engines typi-
cally use an Ad Rank to boost the bid of that
keyword. This is calculated using various
keyword specific metrics (such as closeness of
match, CTR, landing page quality, etc.) and
advertiser specific metrics (account wide CTR,
etc.). Having a high ad rank allows an adver-
tiser to get clicks at lower bids.
3. Quality Score. This is a metric which is an in-
dicator of how well the (keyword, landing
page, ad) tuple are contributing to the Ad
Rank. This is meant to be an indicative score,
with a strong correlation to the Ad Rank [14].
Ensuring that the keyword is semantically re-
lated well to the chosen landing page, and sim-
ilarly, having a semantically related ad and
keyword, are good ways to have a good quali-
ty score. To get a better ROI (Conver-
sions/Clicks) within a limited budget greatly
depends on quality score[2]. Good quality
score is driven by highly relevant keywords,
good quality Ads and better landing pages.
Good quality score can lead to lower pric-
es(bids) and better ad positions [3]. First Page
Bid Estimate (FPBE) for any keyword is the
estimated bid to show an ad for this keyword
on the first page of the search results. This
naturally is less for those keywords with a high
Quality Score.
Commerce Central provides various services, as “apps”,
to e-commerce merchants [8]. Many different e-
commerce platforms, such as Yahoo Shopping,
Shopify, and Big Commerce, are supported. This is
done by defining a normalized data schema to describe
Products, and contain fields such as “Title”, “Descrip-
tion”, “Price”, etc. This data is extracted from specific
e-commerce platforms using available API, and stored
in Commerce Central’s Merchant DB [9]. One of the
most popular Commerce Central apps is Live Web In-
sights [8], which uses javascript beacons to enable store
owners to track their site traffic in real time. This app
gathers a rich set of data, such as time on page, number
of unique visitors, number of visits by each visitor,
number of products bought, and value of each purchase
transaction. This is stored in Commerce Central’s Pixel
DB. We leverage data from both Merchant DB and
Pixel DB in powering our algorithms.
For our automated SEM solution to work well, it has to
provide merchants with Return on Investment (RoI)
comparable to that if a human was running the cam-
paign, with a pre-defined budget. This boils down to
picking a close-to-optimal set of keyword/landing page
pairs, ensuring that the entire budget is spent, and en-
suring that enough sales happen to users acquired
through the SEM advertising channel.

3. System Description
Our system attempts to solve the problem of creating an
optimal SEM campaign for e-commerce merchants by
breaking it into a series of smaller problems. First, we
attempt to find the right set of keywords for a store.
Next, we attempt to find the best landing page for these
keywords, and order our bids by the most “lucrative”
keywords. Finally, we use campaign in-flight metrics to
learn about the actual performance of keywords, and
adjust our bids accordingly. This diagram explains the
various components in the system, and what data flows
between them.
The flow starts, once the customer signs up for an App
like SEM from ‘App FE’ which is stored in ‘Proxy
DB’via Proxy API. ‘App Runner’ which runs as sched-
uled job, finds out that the new store is signed up, is
responsible for Adwords account management, Cam-
paign Creation, Keyword bidding etc. ‘App runner’
gets keywords from ‘Trinity’ which is responsible for
shortlisting product urls from ‘Merchant DB’, fetching
keywords[6] and storing them in ‘Keyword DB’. Chan-
nel Manager is responsible for fetching the store data
like product details, product urls and order details for a
store from the hosted platforms like BigCommerce,
WooCommerce, Magento etc.
These are some of the more interesting problems we
tackled while building this system.
1.1. Determining a set of relevant candi-
date keywords
To get good quality traffic via SEM, it is important to
pick the right set of keywords to bid on. These key-
words should be relevant to the product/web page being
advertised, and must ideally indicate intent to buy the
product/service. All search engines who provide SEM
as a channel also provide tools which suggest keywords
for a given url, or keywords which are similar to a set
of input keywords. This is typically powered by histori-
cal search data, which link searches to clicks on the url.
This is also typically enhanced by content analysis of
the site. Further, related keywords are suggested by
these tools, by using semantically similar terms from
both search sessions as well as unrelated sessions. Fi-
nally, there are a plethora of keyword tools which offer
these suggestions. We decided to use the Google key-
word suggest tool for now, as we found that it provides
sufficient number of keyword suggestions for the sam-
ple stores we tried. The input is a set of advertiser
URLs. In response, a set of keywords suggested for
each URL. Each keyword also has aggregate statistics
such as average number of searches, and average Cost
Per Click (CPC). In addition, Google also provides the
set of product or service category that this keyword is
applicable for.
1.2. Scoring and ranking store URLs
based on user activity
An e-commerce store typically has lots of products, and
a limited advertising budget. It has also been shown that
the budget, and the spread of bids amongst keywords
within this budget, should not be too thin, as it might
starve potential high performing keywords. Therefore,
it is important to limit the number of products to adver-
tise via SEM. We attempt to solve this problem by
ranking the URLs in the store on the propensity that a
visit to that page will result in an order for the store.
Ideally, we should have tried to learn a model that pre-
dicts a buying transaction based on various site visit
variables. Due to time pressure in releasing the product,
we instead came up with a heuristic function which
rewards beneficial actions, such as time on page, and
penalises actions indicating that the web page is bad,
such as bounce rate. We describe the metrics we com-
puted based on data from the Pixel DB, and how they
are used to contribute to the URL score, below. In addi-
tion to the raw ratios, we also compute the normalized
value of this ratio for each page against all pages on the
site.
First, for every metric under consideration for a page
Ui, we normalize that page metric against the page met-
rics of all pages in the system, so as to dampen the ef-
fect. i.e.,

∀ Mi for page Ui, we compute the normalized metric
NMi = Mi / Σ M ∀ pages i
The metrics Mi we calculate are Pageviews (P), Num-
ber of Unique Visitors (V), Bounce rate (B).
Next, we calculate Visitor Conversion rate (VC). This
represents the number of visitors who visited a product
page as compared to the number of times a product on
that page was purchased.
● For a url (Ui):: VCi = Number of purchases
made on Ui / Number of visitors of Ui and
● Conversion rate fraction, PVCi = VCi / Σ VC ∀
pages i
Next, we attempt to calculate a Page View Weighted
Conversion Rate in an attempt to merge the page view
and conversion metrics as: PCWi = 20 * PVCi + Pi
The intuition in combining them is that if a standard
conversion rate is 5% (an upper bound)[19], then every
conversion is as valuable as 1/0.05 times a page view.
Next, we combine the beneficial statistics, conversion
rate and pageviews, with the harmful statistic, Bounce
rate, as follows: Score for Ui SCRi = PCWi - Bi
1.3. Selecting the right landing page for
a keyword
For every keyword K, there are potential set of N
URLs, U1...UN which can serve as the landing page for
that keyword. Our task is to select that Ui which max-
imizes the RoI for K. We calculate this as follows:
Let PMi be the profit margin for the product being sold
on page Ui. Let C be the cost for one click on K. Let the
expected number of clicks be E. Let CRi be the conver-
sion rate for URL Ui. Then, estimated total profit for
selecting Ui as the landing page is calculated as
EPi = E * (CRi * PMi - C)
and the selected landing page Uj is then simply
max(U1,...,Un)
However, certain complications exist. The expected
number of clicks, E, is dependent on the Ad Rank[11]
of the pair (K, Ui). Typically, we only know the price,
Pi, for the product associated with Ui, and not PMi.
Lastly, we might not know CRi reliably for every page.
Hence, we make the following modifications in calcu-
lating the score:
● Ei = V * SMi, where V is the keyword search
volume, and SMi is the semantic match score
of (K, Ui). We calculate SMi using the normal-
ized match score of (K, Ui) in a search index.
[a]
● We use Pi instead of PMi
● We use SCRi instead of CRi
1.4. Campaign set up and initial bids
A typical e-commerce advertiser has a limited budget,
and many products that they have available to sell. The
number of keywords that they can bid on is typically
very large, typically an order of magnitude or two high-
er than the number of products they have to sell. Hence,
it is important for the campaign to bid on the more lu-
crative keywords, and ensure that there is sufficient
budget to pay for these keywords. Let B be the total
budget available, CPCi be the bid for Keyword Ki as
reported by the search engine, Ti the search volume
reported for Ki, and SCRi the score for the selected
landing page Ui for Ki. Let the initial bid for Ki be bi.
Then
● Expected traffic Ei for Ki is Ti * SMi (from
equation [a])
● Expected spend on Ki is bi * Ei
Hence, if we order Ki by decreasing value of SCRi, we
can then bid on these keywords such that
Σ bj <= B, where bj is the bid on those Kj ordered by
decreasing value of SCRj
Since Ei is our estimated traffic, based on how we think
that search engines might calculate Ad Rank, we posit
that this is probably an overestimate of the actual traf-
fic. Hence, we reduce Ei by a damping factor d. Exper-
imentally, we determined that a good value for d is 0.1
1.5. Grouping keywords into Ad Groups,
and showing the right ads
The quality of ads and their relevance to a user's search
are very important factors in determining how well an
ad ranks[11] in an search ad auction, which in turn af-
fects how effective the advertiser’s campaign is. In this
paper [16], we describe a system to come up with high
quality ads for many products and merchants in a scala-
ble manner. The essential trade-off made is create ads
for related groups of products. This is done by catego-
rizing products into a taxonomy. Then, we come up
with taxonomy specific ads, and show which text to use
to come up with high quality, relevant ads.
1.6. Tuning of bids in-flight
The aim of the in-flight bid tuning system is to increase
the bids of keywords that are yielding good traffic and
conversion, decreasing or stopping bids on those that
are not, and adding new keywords if budget allows.
Determining the correct bid is a complex optimization
problem. This algorithm to change bids was developed
by Lexity prior to the acquisition by Yahoo, and con-
sists of applying various rules based on in-flight met-
rics. For instance, bidding is stopped if quality score[2]
is less than 3 or if the bid to show on the first page is
greater than $2. In another example, the bid is raised if
the keyword’s FPBE is less than $2.
4. Case Study/Experience
Calculation of various metrics as described in Section
3.2 was a classic case of aggregating data from log

files. We used hadoop and the computation grid to cal-
culate these heuristics from the raw counts.
Many of the solutions are based on heuristics, and we
had to iterate a few times to come up with the correct
formulae. We tested the algorithms, without actually
creating and running the campaigns, on approximately
50 Million user events from 10 stores, which in aggre-
gate had 100000 products. Some of the practical prob-
lems we faced while managing the SEM campaign for
the ecommerce stores:
1. API limits: Since we relied on external tools
to suggest keywords, accessed via API, there
were limits on how many URLs we could get
keywords for. To ensure that we get sugges-
tions for the most lucrative URLs, we ran the
URL scoring algorithm for all URLs, and used
a subset from amongst the highest ranked
URLs. We experimentally determined that
10% of the top ranked URLs was a good com-
promise.
2. Lack of keyword data: The Google API doc-
ument says that it can return upto 800 keyword
suggestions for every URL. However, for one
store(http://everestgear.com) with 5000+
products, we noticed that the same 800 key-
words were suggested for all URLs we tried!
Hence, the system fall back to using th ‘prod-
uct title’ for keywords and also bing keyword
research tool[7] to increase the keyword cor-
pus.
3. Bad Merchant Data: There is no direct way
to get the store data like product details, prod-
uct urls and order details etc for the app signed
up stores. So we fetch the data from the hosted
platforms like WooCommerce, Magento,
BigCommerce etc which the store is hosted on.
We noticed many errors, such as missing
product pages, test pages with incorrect prod-
uct prices, etc. We needed multiple heuristic
filters to tackle such data.
4. Sparse Visitor Data: In general we noticed
the visitor activity on individual pages is
sparse. Since this affects the ranking of urls,
we attempted to mitigate by normalizing all
scores to a store level.
5. Experiments
We launched SEM App as a controlled beta release for
a month, during which two ecommerce stores
(http://everestgear.com, http://caravantshirts.com.au)
opted for a free trial. We presented below the CTRs and
CPCs of the best, and worst performing keywords, in
terms of CTR.
Results for Caravan T-Shirts
(http://caravantshirts.com.au), an Australian T-shirts
store.
Top and bottom five keywords with the average
CPC, CTR and Quality Score[14].
We observed that keywords with higher CTR needs less
cost(avg CPC) when compared to keywords with lower
CTR. We posit that this is because of their uniformly
high quality score, but we don’t have statistically sig-
nificant data yet.

We also observe that the keywords we ranked highly in
section [3.4] formed the majority of the keywords get-
ting clicks, showing that our algorithm of finding the
more important (keyword, landing page) pairs seems to
work.
Detailed results can be found here.
Results for Everest Gear (http://everestgear.com),
larger store selling mountaineering gear:
Top and bottom five keywords with the average
CPC, CTR and Quality Score[14]
We observed that keywords with higher CTR needs less
cost(avg CPC) when compared to keywords with lower
CTR.
In summary, we can say that our system works well in
creating relevant campaign with generally high quality
score metrics. It also seems to pick the rank the right set
of keywords to bid on. However, the poor conversion
rates are a cause for concern. For Everest gear, we no-
ticed that the keywords suggested by the SEM provid-
ers were of low quality, and very generic in nature (as
opposed to being specific to the products).
6. Related work
Traditionally, SEM campaigns concentrated on opti-
mizing bids once the campaign has started [4, 19]. The-
se don’t deal with the situation where one has to choose
amongst the choice of many products and associated
keywords to bid on, because of limited budgets. Our
solution aims to solve this by using marketplace pro-
vided data to choose the more lucrative candidate pages
to bid on.
Besides the keyword suggestion tools provided by Bing
and Google, here are many tools such as [21]. In addi-
tion, Yahoo’s content analysis platform extracts phrases
and entities from a document [22]. We decided to use
the tools provided by the search engines, as we use this
data in various other algorithms throughout the system.
7. Conclusions
We have described a system which aims to set up and
run an efficient SEM campaign for e-commerce mer-
chants with large product catalogues. We have shown
that using keyword data from SEM tools, in conjunc-
tion with visitor data from our analytics tool, is a pow-
erful way to shortlist products which are more likely to
convert, and provide a better RoI for the campaign. We
notice that our current solution worked acceptably well
for a smaller store, but did not have the desired impact
for a larger store.
We have used heuristics in many places, and are at-
tempting to create a more formal model in these in-
stances. We are currently working with the labs team to
estimate the value of a page, by learning the probability
that a visit to the page will lead to a buy. We are refin-
ing are keyword selection and landing page matching
algorithm. Finally, we are refining our bid tuning algo-
rithm to better leverage data made available by Google
and Bing since it was originally conceived.
References
Due to space issues, we updated all the references here:
https://docs.google.com/a/yahoo-
inc.com/document/d/1iq_ZdzXIYy_XbNaDbBPO4yqJ
2Lm3iU3_a7Q3AgSiK_U/edit#
Store Keywords
bid on
Keywords
with CTR
> 0
Top 100
ranked
keywords
with
clicks
Percentage
of top
ranked
keywords
with clicks
Everest
Gear
500 98 74 75%
Caravan
T-Shirts
874 109 85 77%

OptimizingSEMCampaigns

More Related Content

What's hot (20)

Viewers also liked (6)

Similar to OptimizingSEMCampaigns (20)

OptimizingSEMCampaigns