DEMYSTIFYING DISPLAY OPTIMIZATION IN PERFORMANCE CAMPAIGNS BY DAVE HIMROD & KEVIN SINHA

WHITE PAPER

THE SECRET SAUCE Algorithmic optimization engines are often cited as the “secret sauce” for networks and ad technology platforms. But have you ever wondered what ingredients make up the sauce? For performance campaigns (those that link advertising to a user’s purchase, a sign-up, or some other desired action), sophisticated algorithms leverage real-time data into the buying process through an “optimization engine.” These engines enhance a media planner’s natural instincts by automatically optimizing bids based on historical results. This white paper explores the core ingredients that optimization engines use to make display buying more effective. We will cover the mechanics of optimization algorithms that determine bid values for the best performing impressions and offer some interesting tips that can boost the return on investment (ROI) of your campaigns.

1

[email protected] ©2012 AppNexus, Inc. All Rights Reserved.

www.appnexus.com

WHY OPTIMIZATION MATTERS The competition today in ad technology is tight; even a 1% lift in profits can be the difference between winning and losing a client. Historically, media buyers have used manual optimization to adjust targeting and bidding on individual campaigns to get that incremental lift. However, as available online content grows exponentially, pure manual optimization is both inefficient and ineffective. Algorithmic performance optimization automates this process, creating a scalable and repeatable way to identify high-performing inventory and to bid at a price that increases the likelihood of making a profit.

UNDERSTANDING ALGORITHMS purchase the best-performing impressions

Optimization Algorithm:

for their

A defined way to mathematically

An optimization engine is composed of various algorithms that help media buyers campaigns

by

mathematically

determining the most appropriate price to bid for each opportunity.

determine the best price to bid for an impression.

A common misconception is that the Internet is made up of “better performing” and “remnant” inventory. While there may indeed be premium and poor content, in most performancebuying scenarios the value of the inventory is

INVENTORY

determined by a combination of the content

The Internet is a big place. To help media

of the site and a specific campaign’s creative

buyers find the best performing ad space

and targeting.

for a particular digital campaign, it can be helpful to classify the digital ad inventory in

Optimization algorithms evaluate all the factors

the Internet according to characteristics that

that help establish this value. Platform-specific

influence the type of users they attract. The

nuances aside, we’ll discuss a few elements

classification allows an optimization engine

that all optimization engines must consider in

to make assumptions up front that speed the

order to most accurately predict the value of

optimization process and to make campaign-

an impression: inventory, performance goals,

inventory performance associations in an

frequency, and learnings.

efficient way. 2

[email protected] ©2012 AppNexus, Inc. All Rights Reserved.

www.appnexus.com

Characteristics of inventory classifications may include:

Quick Tip:

• Publisher

Manually narrowing down

• URL

inventory options for your

• Geography (the location of the user’s IP address)

campaign based on your expertise is a simple way to

To

be

useful,

inventory classifications

must strike a balance between granularity

influence the optimization

and volume. The balance ensures that

process. For example, if a

there is enough data for each slice of

media buyer is targeting

inventory evaluated to accurately predict

movie-goers, then the

its performance. For example, one slice of inventory may look like this:

campaign should specifically

• CNN

target websites categorized

• www.cnn.com/tech

as “entertainment” before

• CNN UK-based users

optimization takes over.

PERFORMANCE GOALS The heart of an optimization engine is determining how much to bid for a particular impression, based on a campaign’s goals.

Translating Goals Into Spend Direct response advertisers typically pay for a particular user response via a cost-per-click (CPC) or cost-per-action (CPA). However, in a real-time environment, impressions are most commonly transacted between advertisers, networks and publishers on a cost-per-thousand impressions (CPM) basis. Since CPC/CPA is measured differently than CPM, a translation needs to occur. At this point, optimization plays a factor. Optimization algorithms translate these CPA/CPC goals into estimated CPM bids based on a variety of factors, including historical user response rates for a particular type of impression. If historical rates indicate that the response rate is low, then a lower bid will be submitted. Conversely, if the response rate is high, then a higher bid will be submitted.

3

[email protected] ©2012 AppNexus, Inc. All Rights Reserved.

www.appnexus.com

Calculating Valuation So how does that translation work? Using the likelihood of a user’s response to an impression, we convert the advertiser’s goal of how much to pay for a conversion (CPA/ CPC) into a CPM bid. Here is a simplified equation:

CPA

*

Likelihood of Response

*

1000

=

CPM Bid

$3

*

0.01%

*

1000

=

$0.30

If the advertiser is willing to spend $3 per conversion and the likelihood of a response on the impression is 0.01%. Since advertising inventory is priced in cost-per-thousand impressions, the CPM bid should be $0.30. If the probability of a response is higher, the bid is increased.

Don’t shy away from high CPMs. As a media buyer, you may naturally be averse to buying impressions with a high CPM via real-time bidding. But don’t forget that it isn’t the cost of the impression that matters; it’s the cost of the acquisition. High CPM impressions are worth every penny if they result in converting users for a lower CPA/CPC.

FREQUENCY An optimization engine also needs to consider the individual user; for example how frequently or how recently a particular user has seen an ad. Think of frequency as an application of the law of diminishing returns. If a user sees the same ad 20 times in a one-hour browsing session, he/she is likely to become desensitized to the message and less likely to interact with subsequent ads. An optimization algorithm looks at how performance degrades with higher frequency and uses this data to modify a bid.

4

[email protected] ©2012 AppNexus, Inc. All Rights Reserved.

www.appnexus.com

LETTING THE ALGORITHM LEARNN

determine if you like them, the algorithm bids

Optimization only works when there is

inventory becomes the historical data used to

historical data for a campaign’s response

determine future bids.

on different slices of inventory to see how they perform, and the conversion rate on that

rate on a variety of inventory. But what if a campaign is new and there is no historical

A complex optimization algorithm might

data? In that case a campaign will have to

provide you with a jump-start in learning based

“learn” what works and what doesn’t.

on past data aggregated from other campaigns. In the restaurant example, you might be fairly

Let’s use a real-world analogy to illustrate

sure you like Thai food, and so you begin your

how learning works. Suppose you are a

search with Thai restaurants.

foodie who just moved to New York City, and your goal is to find restaurants that you like.

The learn phase can be a double-edged sword.

Recommendation sites and your friends can

On one hand, it should be brief, so that the

help you narrow the list, but ultimately this is

advertiser doesn’t continue to spend money on

a question of your personal taste.

impressions that don’t perform. Conversely, the learn phase needs to be long enough to collect

Randomly selecting restaurants until you

adequate data to run an efficient campaign.

accumulate positive experiences would take a



very long time, and you might go broke trying new places before you find ones you really like. So how do you determine where to eat?

Quick Tip: One thing to consider whether you’re

If you were an optimization algorithm, you would categorize New York City restaurants

running a new campaign or evaluating

by a number of factors: cuisine, neighborhood,

ad technology partners is determining

price, hours, and so on, just as you might

how to leverage your historical data.

categorize inventory by the factors we described above. From your experience with

Chances are you’ve collected a

these types of eateries in the past, you identify

treasure chest of valuable data that

the factors with the greatest likelihood of a

can help decrease your learn time.

successful dining experience for you, and select restaurants with these attributes for you to try going forward.

Look for ways to incorporate that data into your current work stream to turbo-charge your learning.

Likewise, the learn phase in display optimization is essentially a process of gathering data and experiences. Similar to trying restaurants to 5

[email protected] ©2012 AppNexus, Inc. All Rights Reserved.

www.appnexus.com

GETTING OPTIMIZED A slice of inventory is considered optimized for a particular campaign when the algorithm has enough data and recognizes a significant percentage of users performing well for that campaign-inventory combination. The optimized inventory will now receive higher bids and more ad spend from that campaign. However, a campaign doesn’t stop learning. Just like our New York City foodie who has discovered 5 favorite restaurants but lives in a city where new restaurants are popping up all the time, the optimization engine is designed to constantly test new inventory and improve results.

TIPS TO IMPROVE YOUR OPTIMIZATION Now that you know the building blocks of a display optimization engine, you’re equipped to get the most out of your technology solution. Next time you’re evaluating a technology partner or reviewing your latest ad campaign, remember the key principles of successfully optimized campaigns:

√

Data is king: Leverage historical learnings of similar campaigns. Ensure your ad technology provider is porting learnings from your past campaigns and getting you optimized as fast as possible.

√

Don’t forget to learn: Plan on using 5% to 20% of your budget in data gathering to help

√

Inventory: Narrow down your inventory options in advance when possible. Make sure you

increase the ROI of a performance campaign.

have the tools (both automated and manual) to eliminate inventory sources that are not relevant to your campaign.

We’ve only just skimmed the surface of optimization capabilities to give you a basic orientation. A great optimization engine is always developing on top of this central framework. Stay tuned to AppNexus www.techblog.appnexus.com for the latest optimization innovations.

6

[email protected] ©2012 AppNexus, Inc. All Rights Reserved.

www.appnexus.com