August 23, 2007


By Lai Kok Fung, BuzzCity CEO

John Wanamaker, a US merchant in the late 1800s, famously once said “Half the money I spend on advertising is wasted; the trouble is, I don’t know which half.”

We are much closer today to solving Wanamaker’s dilemma. Internet metrics and computer algorithms make it possible to target ads to users who are likely to act on them.

Providing the parameters to computers to analyse the data and filter out the white noise has been a passion of mine since grad school. Back then, I wrote a PhD dissertation on “adaptive learning” with the somewhat clunky title “Deformable contours : modeling, extraction, detection and classification”.


For example, I would feed a database of images into a computer and directed it to locate the ones shaped as keys. The first lesson was that you need to start with some assumptions and a model, otherwise it takes too long for the computer to find what you’re looking for. Second, you have to realise that your initial model is not perfect. You need it to adapt based on experience.

The picture above was taken from my thesis. In clockwise direction, it shows the original "noisy" image, the intermediate processing step called the "edge magnitude", the final and initial boundaries.

My research applies directly to internet advertising. I’ve even had to dust off some of my old reference books, particularly “Adapative Filter Theory” by Simon Haykin.

Let me show you an abstract example and then move to the real world.


Suppose there are two types of people who visit a website -- circles and squares (which are you?) -- and two types of advertisers. Advertiser A wants to reach circles; Advertiser B is looking for squares. Each advertiser has a $2 budget and is willing to pay $1 per valid lead.

Since the audience is evenly split between circles and squares, there is a 50% chance the website publisher will deliver the right audience. It has to provide four leads to each advertiser, effectively at $0.50 per lead. After selling the entire audience, the publisher earns $4.

With perfect targeting, though, the publisher can deliver two circles to Advertiser A, two squares to Advertiser B to earn $4 AND it will still have four leads left to sell, giving it the potential to maximise revenue at $8 – twice as much as it was earning before. So the advertisers are happy because they are reaching the intended audience, the publisher is happy because it is making more money and users are happy because the ads are more relevant to them.

The basis of matching advertisements to consumers on the internet is probably common knowledge now to many readers – put sporting shoes ads on sports sites, movie ads on entertainment news sites and new car ads just about everywhere.

In reality, though, ad placement is a lot more complicated.

There may be hundreds of companies placing ads that target sports fans. Which ads though are likely to receive the most click-throughs and subsequently the most online purchases?

Information about a publisher’s content and an ad’s content provide an initial guess. But to determine an ad’s relevancy to users, placement algorithms take into account more variables, such as

• the country where a user is coming from (based on IP addresses)
• user demographics (normally obtained from registration forms)
• user interests (inferred from surfing history, types of ads clicked, search terms used), and
• type of mobile phone.

We also constantly observe how a class of users behave to certain ads. For example, if we see that music ads work well on one page of a publisher’s site, but gaming ads work better on another, we adjust the placement accordingly. Our experience shows us that modifications based on incoming data can easily double or triple ad revenue for a publisher.

With nearly one million clicks per day, our engineers have tons of data – more than I ever could have dreamed about when I was working on my dissertation – that can be used to optimise ad performance. And then there’s the anecdotal feedback as well. When a new mobile site is launched – or if a new ad campaign is launched – publishers are quick to call me if the ads are not producing.

The buzz surrounding online advertising is reflected in recent acquisitions. AOL bought TACODA, a company that delivers targeted rich media and video ads based on consumers’ internet surfing habits, for US$275 million. And Google bought DoubleClick for more than US$3 billion, despite some controversy surrounding DoubleClick’s use of sypware to track and record which advertisements users view while browsing. Within two to three years, online ad spending will surpass newspaper advertising in the US.

Advertising on the mobile internet becomes more exciting because (1) there’s more data to work with (mobile device, telecom carriers, even location) and (2) we’re reaching a new demographic in developing countries and among blue collar workers.

Advertisers can specify their targeting parameters and control how much they spend by opting for CPM (cost per thousand impressions) or CPC (cost per click). Then the algorithms kick in to optimise revenue for publishers. And since myGamma is one of the largest mobile communities in the world, BuzzCity can provide sophisticated click analytics based on user demographics. Users win out as well since they are more likely to see ads that meet their needs and interests. At the same time, though, the industry needs to exercise caution to avoid infringing on consumers' privacy.

As a computer and modeling geek, it's a thrill for me to be part of this new wave. We are continually learning and adapting.