Conversion Pixel Categories for Improved Performance

Here at AppNexus, we've long believed that we will achieve better outcomes for marketers if we complement our advanced machine learning with customers' extensive business knowledge. When we started developing the AppNexus Programmable Platform a couple of years ago, this belief was core to how we designed optimization. We believe that our customers should have the flexibility to customize our algorithms and manually program their own buying rules, however that shouldn't be necessary in order to achieve results. When it comes to out-of-the-box optimization, our algorithms should just work. This approach should hold true for all types of outcomes that a marketer is driving towards, whether it is an impression, a view, a click or any other conversion event.

As with all machine learning, though, success is dependent on good data -- not only the volume of data, but also its integrity, meaning the maintenance, accuracy, and consistency of data over its entire life cycle. Take an algorithm that predicts clicks, for example. Clicks occur frequently and, thanks to standards shared across the industry, can be measured with good precision. Likewise, the definition of a click doesn't change from marketing campaign to marketing campaign. A click is a click -- someone hovers their cursor or finger over an ad, and then they press a button or screen.

Accuracy: ✓

Volume: ✓

Consistency: ✓

(As for maintenance, let's assume that's a given for any reliable tech company.)

However, if we consider an algorithm that predicts conversions, the data isn't as straightforward. Because marketers' core businesses are so different, their definitions of "conversion" are plentiful and diverse. Conversion can mean anything from a page visit, to a product hover, to an in-store purchase, to a newsletter signup. The definition can change not only from marketer to marketer, but also campaign to campaign.

Consistency: ✗

Further complicating the situation, a trader may have measurement set up on all, some, or none of these events. Perhaps the marketer ran a campaign last year and already placed a pixel on the billing information page, but has not yet placed a pixel on the order confirmation page. In that case, a trader would likely optimize to the only event she can track -- driving users toward the billing workflow -- even though placing an order is the event that the marketer actually cares about. In other words, the "conversion" that the trader is optimizing to isn't actually the true conversion.

Accuracy: ✗

Finally, depending on the event being measured, the volumes of different conversions can be wildly different. Motivating users to enter their email address in exchange for a 10% discount is much easier than motivating users to sign up for a credit card, for example. As a result, the conversion rates for two such events are likely to be very different. Predicting email signups will be easier because the event occurs more frequently and, as such, has more data associated with it. Predicting credit card signups, on the other hand, will be harder because the event occurs rarely and therefore has less meaningful data associated.

Volume: ✗

As the AppNexus team analyzed this problem a bit more, we realized that without knowing the type of conversion that a marketer is driving toward, our algorithms could only be so successful. Without that metadata, the only things we know about a conversion are: its value as defined by the trader, the pixel ID with which it's associated, and maybe the URL on which it fired. We don't know the type of event a trader is optimizing to, how frequently that type of event occurs, whether completing that type means a consumer is more likely to complete another type, how many steps need to be taken before completing the event, and so on.

Taking optimization to the next level

We realized that if we started collecting and storing more information about conversions, our prediction would have the potential to be much more robust. For example, we could:

  • Predict the probability of a specific type of conversion, rather than predicting generic "conversions."
  • Predict the probability of one type of conversion given another.
  • Optimize to higher funnel events, such as landing page visits, when lower funnel events, such as checkouts, are scarce.
  • Optimize to something that we haven't even thought of yet!

That's why we're introducing event categories as a new feature within conversion pixels.


Starting in late July, traders can help us take optimization to the next level by categorizing conversion pixels with one of seven events:

  • Visit landing page
  • View an item
  • Add to cart
  • Initiate checkout
  • Add payment info
  • Purchase
  • Generate lead

Selecting an event category not only enhances our Data Science team's "sandbox", but also introduces new possibilities for CPA optimization. The event types you provide won't affect our optimization algorithms immediately, but by providing our optimization team valuable data for prediction, they enable us to start testing and incorporating it.

If you don't see a good fit for your type of conversion, you may also leave the field blank. If that's the case, our team would love to learn more about your optimization scenario here.

Finally, event type categories are just the beginning: over the coming months, you can expect other optimization improvements to be released regularly. Sign up to get updated for new product releases here, and if you have questions or other ideas for optimization, please drop us a line.

Megan Arend

Product Manager for advertiser optimization.

Subscribe to AppNexus Product Blog

Get the latest posts delivered right to your inbox.

or subscribe via RSS with Feedly!