Archive for the 'Predictabuy' Category

Real-time Context Targeting

Targeting local advertising using the real-time context of local searchers is an effective alternative that overcomes the limitations inherent in simply advertising a local business as a potential alternative based on the category and location of the search.  In this post, I’ll describe what real-time context is, how it can be discovered and how models can be trained to target local advertising using this approach.

This post is part of the Improving Local Ad Performance series.  In The Challenge of Fine Segmentation in Local Advertising I discussed the challenges inherent in relying solely on a ‘like for like’ targeting approach in local search.  In a like for like targeting, local advertising is displayed by matching the category and location of the advertising business to the category and location of a search request.  In today’s post, we are going to discuss using Real-time Context Targeting as a complementary alternative to this approach.

What is Real-Time Context?

Let’s use the example of someone searching for a Taxi to call.  They might be looking up a number for a particular Taxi, looking for any old Taxi or even looking for the nearest available Taxi.  The following diagram illustrates various situations where a user might be looking for a Taxi:

The Real Time Context of Someone Looking for a Taxi

The Real Time Context of Someone Looking for a Taxi

The actual context of the user is hidden from us, but let’s assume for the moment we actually know it (or could at least take an educated guess at it).  In that case, we could advertise based on that context.  So, for example, if they are call a Taxi because they’ve lost their keys, we could advertise the Locksmith they will also need!

Why Not Just Advertise Taxis?

There are several reasons to consider advertising something other than just Taxi services:

  1. Other types of advertisers (a Locksmith for example) might be willing to pay you far more for a lead.  So, even if your response rate ended up lower than for a Taxi ad you might make more money because you’ve delivered more valuable leads.
  2. Some users may be loyal to a particular brand of Taxi or using a service that provides the nearest available Taxi.  As such, they aren’t really open to substitution – but may be more more open to something that meets another need in their context.
  3. You might not have Taxi ads in that locality.

If the advertising is relevant to the real-time context of the caller, they’ll appreciate and perhaps act upon them.

How Do You Determine Context?

Sounds great – but how do we discover the hidden context of our searcher?  We have to infer the hidden context from characteristics (or attributes) of  the search and current events.  Examples include:

  • time and date based attributes (time of day, day of week, weekday/weekend, holiday, seasons, etc.);
  • place attributes (requested location and current location of the searcher if known); and
  • real-time events such as weather, sporting events and cultural events.

In fact, by examining past search history (in a completely anonymous way that protects the privacy of individuals) and the historical event record we can apply machine learning algorithms to build models that recommend the best ads based on current events and the attributes of a local search.  An example of how we do this at Predictabuy is shown below:

Real-time context targeting of local advertising

Real-time context targeting of local advertising

What’s in the Event Stream?

The explosion of social sites and mobile usage exemplified by Twitter and Facebook provides a rich and evolving set of events that a context-driven local targeting engine can exploit.  As such, this approach will just get richer and more effective over time.  If you have a suggestion for events you think will be important in local advertising leave a comment and let me know.  I’d love to hear from you.

Learn More

For a free consultation on how you can use Real-time Context Targeting in your business contact me.

This post is part of an ongoing series on Improving Local Ad Performance.  Upcoming posts will cover the use of consumer behavior and preferences in the effective targeting of local advertisements.  To ensure you don’t miss any of the discussion:

Advertisements

Microsoft Researchers Increase CTR 670% Using Behavioral Targeting

Behavioral targeting has been around for a while in various commercial products.  But now, for the first time, researchers at Microsoft Research Asia have completed an important empirical study demonstrating that it works.  In fact, it works extremely well increasing the click-through rate (CTR) up to 670% with the potential to achieve improvements in excess of 1000%.  Wow.  And kudos to Microsoft for undertaking this as an academic research program and making the results available to all.

They cover a lot of important ground in an academically rigirous way.  There are three key conlusions.

The Basic Assumption Holds

The basic assumption of behavioral targeting is that all the people who click on an ad are more similar to each other than they are to all those people who clicked on other ads.  If you can’t prove this assumption, it’s back to the lab to whip up something new.

But, not to worry, the researchers found that the people who clicked on the same ad are up to 90 times more similar to each other than users who clicked on another add.  Whew!  I guess that’s good news for anyone who has been touting the merits of behavioral targeting.  It’s also intuitively satisfying.  Still, it’s great to have it proven by research.

CTR Can Be Increased by Up to 670%

They used click-through rates as their measure of performance (because it is a readily available measure).  They implemented behavioral targeting by segmenting the users with various strategies and then compared what the CTR would have been with and without the segmentation strategy:

Through studying ads CTR before and after user segmentation for ads delivery, we observe that ads CTR can be improved by as much as 670% over all the ads we collected.

and:

In addition, we notice that if we can further design more advanced BT strategies, such as novel user representation approaches and novel user segmentation algorithms, ads CTR can be further improved beyond 1,000%.

Short Term Search Behavior Gives the Best Results

Finally, the researchers examined several different approaches to implementing behavioral targeting:

Through comparing different user representation strategies for BT, we draw the conclusion that the user search behavior, i.e. user search queries, can perform several times better than user browsing behavior, i.e., user clicked pages. Moreover, only tracking the short term user behaviors are more effective than tracking the long term user behaviors, for targeted ads delivery.

What it Means: A Mobile, Local Perspective

This study was done using logs from users searching, browsing and clicking on the web.  Local and mobile bring additional nuances to the equation.

This report is exciting for us here at Predictabuy because it confirms a lot of our own research which is specifically aimed at understanding user behavior in a mobile, local context.  Our research shows:

  1. short term behavior is also a stronger predictor than long term behavior in a mobile, local context;
  2. situational factors such as location, time of day, day of week and weather are very useful in user segmentation; and
  3. advertising performance benefits from dividing users in to  more segments than the number used in the Microsoft study.

via Greg Linden.

Read the full paper for yourself:”How much can Behavioral Targeting Help Online Advertising?” (PDF)

Tracked calls are user generated content

How we think about things matters.  I enjoy finding and creating new ways of looking at things because they generate new insights which can eventually lead to disruptive change.  In fact, this blog is about exploring perspectives on mobile, local search and advertising.

And the seed for new perspectives comes from conversation.  Last week I got a new perspective on call tracking as the result of a twitter conversation (with @sebprovencher).

You can think of tracked calls (or clicks or any user action really) as a form of user generated content. In fact, especially for calls resulting from local searches, it’s an extremely valuable form of user generated content:

  1. EVERY user provides this feedback EVERY time they make a call (assuming you are tracking the calls of course); and
  2. It’s a highly structured signal – the user has taken a very explicit action to contact a business.

And, in the case of calls, you can further strengthen the signal by also keeping track of how long the caller stayed on the phone.

Utilizing User Generated Content

To be sure, this implicit user generated content is different from content like ratings and reviews.  You probably don’t want to display something like “300 people have called this person’s ad” (though, thinking about it that might actually be kind of interesting).

However, through appropriate analysis this data can be used to provide users with recommendations based on their situation and preferences.  For example, by looking at this data you can learn what people are looking for on Friday night versus Wednesday morning.  This is the sort of analysis we’re doing at Predictabuy.

And here’s another interesting thing: combining this implicit data with coventional ratings results in better recommendations than you can get by looking at the ratings alone.

What is predictabuy?

Predictabuy is a start-up providing software and services to local information publishers.

We provide advertising recommendations that improve ad relevance resulting in increased advertising revenue for  publishers.  Our recommendations are based on where a user is, their current situation (such as time, date and weather) and their personal preferences.  We learn how to make recommendations by analyzing past and current local search history.

Predictabuy recommendations are able to improve on traditional targeting methods by appropriately taking  into account many more factors and dynamically generating recommendations based on real-time conditions.  This slices a local publishers inventory more finely and with greater precision allowing more accurate selection of the most valuable advertising and higher participation rates.

By building models of individual and aggregate consumer behavior from a publisher’s search logs Predictabuy lets both publishers and consumers realize the value inherent in this data.