2. Who am I?
Company
ADVERTISING
e-X combines world-class data and technology to help brands
Platform INTELLIGENCE define, model and reach their audiences online.
PLATFORM
Our audience engagement divisions apply advertising
intelligence to deliver high-impact, high-engagement
Divisions campaigns across display, video and mobile to more than 450
million unique monthly users worldwide.
Advertising Intelligence 2
3. Have you not heard of us?
1,255
873
800
702
495
482
481 Global monthly unique visitors
(2012 average, comScore)
361 (In millions)
300
244
Advertising Intelligence 3
4. Big data: Why you can't beat the
model in online audience targeting
Doug Conely
Senior Director, Global Data & Targeting
www.exponential.com 4
6. And there’s all sorts of it
DEMOGRAPHIC SEARCH SOCIAL GRAPH BEHAVIORAL
Advertising Intelligence 6
7. Regardless of which data set you use it into action
But it’s all meaningless unless you turn
DATA INFORMATION INSIGHT ACTION
Advertising Intelligence 7
8. From data to information
http://www.techradar.com/reviews/gaming/games-consoles/xbox-360-703247/review
http://www.pcmag.com/article2/0,2817,2365824,00.asp
http://reviews.cnet.co.uk/games-consoles/nintendo-wii-review-49285257/
http://www.amazon.co.uk/Nintendo-Wii-Console-Includes-Sports/dp/B0007UATDG
http://www.pocket-lint.com/review/4447/sony-ps3-slim-console-review
http://www.play.com/Games/PlayStation3/4-/16440111/Sony-PlayStation-3-PS3-Slim-
Console-With-320GB-HDD/Product.html
Technology Consumer Electronics Audio, TV & Video Nintendo DS
Hardware Camcorders Nintendo Wii
Information Tech Cameras Sony PS2
Software Car Electronics Sony PS3
Home Comms Sony PSP
Mobile Comms Console & Handheld PS Vita
Video Games Platforms Computer X Box 360
Advertising Intelligence 8
9. Information to insight: what a computer tells us
Luxury Car Brand Conversion: Lift vs Network Reach
2,000
1,500
1,000
Conversion Lift Index (Base = 100)
700
500
400
300
200
150
150
1 1.5 2 3 4 5 7 10 15
% Network Reach
Advertising Intelligence 9
10. We need humans to extract stories that:
Luxury Car Brand Conversion: Lift vs Network Reach
2,000
1,500 VALIDATE
1,000
SURPRISE
Conversion Lift Index (Base = 100)
700
500 CHALLENGE
400
300
200
150
150
1 1.5 2 3 4 5 7 10 15
% Network Reach
Advertising Intelligence 10
11. For example:
VALIDATE SURPRISE CHALLENGE
Car customers really Hotel shoppers are DVD rental service
do have strong car interested in dining out, enthusiasts are not movie
in-market and car nightlife, music artists goers, but rather families
interests who are currently on tour and hardcore gamers
Advertising Intelligence 11
12. From insight to action
Marketing strategy
Media buying & planning
Creative
AUDIENCE DISCOVERY ACTION
Advertising Intelligence 12
13. Insight to action: in digital we can buy what we learn
Luxury Car Brand Conversion: Lift vs Network Reach
2,000
1,500
1,000
Conversion Lift Index (Base = 100)
700
500
400
300
200
150
150
1 1.5 2 3 4 5 7 10 15
% Network Reach
Advertising Intelligence 13
14. But it turns out… you can’t beat the model
Brand X Beer Drinker
Propensity to visit Brand X
Exponential Network Reach
Advertising Intelligence 14
15. Questions
Doug Conely I Senior Director, Global Data & Targeting
doug.conely@exponential.com
www.exponential.com 15
Editor's Notes
In 2011 IBM released a paper with the oft repeated comment that “ “Everyday, we create 2.5 quintillion bytes of data–so much that 90% of the data in the world today has been created in the last two years alone.” When people engage online they create huge amounts of data. Exponential has the same big data issue as many other businesses. We collect 80 billion events a month across 450 million users worldwide and organise that into 50,000 categories. That’s equivalent to seeing more than 5600 Olympic Stadia of people more than 170 times a month each. Day to day online business dwarfs the data potential of the biggest events. [ “The amount of data is meaningless”] All big data presentations start with stats about the vast scale of data now collected. But we would argue that the amount of data being collected today is now so vast as to be incomprehensible to the human mind. 5 years ago it was about how big your data warehouse was and how fast you could process data. Today, it really isn’t about how big it is but what you do with it.
Companies are making bets on the kinds of data that yield the most value. In truth, this is driven by where the company came from (search, traditional media, social media, display advertising) more than anything else, but all yield big data. Which of these kinds of data yields the most value is open to all sorts of argument. Our punt is on ‘behavioural’ – or interest-based – that is defining people’s interest by understanding the pages they have visited on the web. The point is all that – with scale – all can yield big data and that, if the data is big enough, all can act as a valuable proxy to people’s interests and, therefore, the kind of advertising that is likely to work for them. [PB: Doug, need examples of companies in each bucket…]
So we prefer to think about how you turn data from information then to insight and finally to action. This is not a new notion, we might even have stolen from Frank Zappa, but there’s still too much noise about data collection and not enough focus on data actionability.
For many types of data, the internet can be thought of as a categorisation problem. Until, it’s categorised – ordered – it can’t be examined. The way we turn data into information is… [explain page-level contextualization briefly]. [Explain briefly how other sorts of data are categorized?] [PB: Doug – need the screenshots for this]
So now we can examine the data but it is here that we learn you cannot do without humans. Our machines tell us the online interests and intentions that represent a brand’s most likely customers – the dots are different online ‘behaviours’ – those showing the highest ‘lift’ are the behaviours that most indicate a propensity to buy. But the output of that analysis is the above. There’s little we do can do with this, especially since the chart – by definition - looks much the same for every brand. So the information has now been made relevant is STILL isn’t insight.
That’s where art needs to be layered over the science, that’s where we need humans to make sense of the information; to extract stories that: Validate: Sometimes analysis of the information yields findings that that simply validate our assumptions. Surprise: Sometimes the analysis tells us things that, in retrospect seem like common sense, but – because of the way we think and measure – haven’t been acted upon before. Sometimes it challenges your assumptions; it tells us things we’d never have thought of and gets us to rethink who we are targeting and why.
Validate: Sometimes analysis of the information yields findings that that simply validate our assumptions. Like our car example - people that book test drives and order car brochures demonstrate auto in-market/interest behaviors across the web. BUT, it’s end of funnel – that doesn’t help further upstream. Surprise: Hotels booking site digital team thought they should be targeting people looking at hotels – but learned it’s too late by then. They should be targeting people looking at nights out, shows and concerts; people that need a hotel for the night. Sometimes it challenges your assumptions. Like the DVD rental group who thought their customers were movie lovers but actually turn out to be people with young children or hardcore gamers – those that can’t or won’t go out for entertainment.
So what can we do with these insights? Well, firstly, these aren’t insights for optimising your display advertising. They are insights into the interest and intentions that define your customers – they are insights for your marketing strategy – who are we targeting and why? It informs your entire media buying and planning strategy, not just on, but offline. And they inform the kind of creative you should be using for your audiences – again, on and offline.
But, none of these places can the links be forged and acted upon programmatically. Even in online, this has been the problem with many of the planning tools in place - they delivered great insight but then still required you to find and buy a proxy to the audience you just identified. Fortunately digital advertising has become sophisticated enough that audiences can now be passed from planning application to media buying application, in our case on a single platform. The idea is that marketers should be able to buy what they learn.
But, we can go further that manually selecting the audience segments the combination of art and science identifies for us to buy media. We have found that for the most ‘audience efficiency’, ‘you can’t beat the model’. In our case we expose that audience model as an explicit trade off between the lift, or quality of fit, of the model to a brand’s target audience – the y-axis – and the reach you can achieve – the x-axis. The larger the model, the worse it gets, but at least it’s transparent. What we find now is that the traditional approach of manually selecting audiences based on their nearest analogue to your survey data, or even based on your audience insights discovery, is never as efficient as the model. In the chart above, the beer brand in question used its brand site visitors as a proxy to its customers. They could compare the lift and reach of their traditional, preferred audience targets – e.g. Men, Sports and Arts & Entertainment – to that suggested by a model. In all cases, their preferred audiences sacrificed significant lift for any given reach or vice versa. For a campaign with brand objectives this represents a missed opportunity and real wastage. So we’ve turned Big Data into a simple, scalable, repeatable action: “you can’t beat the model”