Rethinking Keyword Attribution: A Big-Picture Approach

Current attribution models often provide an inadequate prediction of future success. Here's a new way of looking at keyword attribution that may just change all that.

Attribution modeling is one of the best tools we have to understand what's working - and what isn't - in any given marketing campaign. Sometimes, it's the only data-driven way to settle inter-departmental debates about which channel or type of terms, brand vs. non-brand, warrants the biggest investment. And choosing the "right" attribution model can be an inter-departmental fight in itself. Since budget decisions can depend upon the models used, "whoever owns your attribution model owns your budget."

Nonetheless, the most common models being applied rely upon arbitrary weighting depending upon the order of which an ad is clicked (or viewed), yet we know that the customer journey is not linear. This will result in a model that is an inadequate predictor of future success. Bing Ads researchers are testing a novel approach to keyword attribution and believe to have uncovered another, perhaps better, way.

Many Options, No Easy Answers

But before we get to that, let's look at the traditional approaches. Each of the models currently in use - last ad, first ad, linear, position, time decay, and even custom models - can often fall short. Why? Because none of those models provides a causal view of what buyers are actually searching for that eventually leads to conversion. And yet we continue to rely on them. This issue isn't limited to search novices - even third-party search experts place their trust (and dollars) in non-predictive models with the traditional approaches. In fact, the chief marketing officer (CMO) of one of our biggest advertisers once recently told me he had a simple solution to a complex problem: last-click. Unfortunately, it just isn't that simple.

Is Arbitrary Data Driving Your Strategy?

Traditional attribution models - including last-click - record ad impressions and clicks before conversion. Factor in a fixed weight function, and you get your dollar value. Each approach has its own perceived benefits, but it's fair to say that the one thing all current attribution models have in common is that they are subjective and will not give you the optimal attribution.

The problem is this: no matter how you look at it, this is all a more or less an arbitrary function. You cannot prove causality - and ultimately, value - without experimentation. And keywords are just too granular for practical experimentation with current technology.

Thinking More Holistically

If you cannot prove causality, what could be a better, alternative approach? Rather than tracing each query and applying an arbitrary fixed weight, what if you could analyze each query in a conversation funnel and its relation to the converting query? What if, in this analysis, you could also include ads that were NOT clicked, and queries you DIDN'T bid for?

Here's a new approach: Assign a weight based on a number of attributes, such as query relatedness, frequency, and user diversity. Here's a look at how Bing Ads is calculating a weight (or assist value):

1. Establish the converted term and then trace back prior search queries.
2. Calculate the relative probability of search queries appearing in the conversion path (relative to the probability of appearing in any search path).
3. Calculate degree of "relatedness" - defined as to how closely an assist keyword and converted keyword are semantically connected to one another.
4. Calculate the user diversity - defined as how many searchers are searching for the same assisting keyword and then converting in the funnel.
5. Calculate the frequency of the pattern and pair the converted term with prior search queries.

This novel approach is a better way of assigning attribution weights and uncovers typical patterns of related queries, helping advertisers better understand their buyers' entire decision process. Perhaps even more important, it provides a reliable way to discover new keywords, words that are more likely to drive success for future campaigns.

How Does It Work?

The new attribution model was applied to the movie and entertainment category to see which search queries helped to drive conversions. The graphic below shows the converted term (in blue) identified in the center, along with the different search paths users took to end on that converted term. The larger the bubble, the larger the attribution weight and "assist" power it had on the converted term. Based on our analysis, the term "hindi movies 2013 full movie" played a significant role in driving conversions for the term "new hindi movies."

As you can see, there are number of searches leading up to a conversion, so it's important for advertisers to understand the whole search path to properly optimize their campaigns. While this keyword attribution approach is still in a limited pilot phase (with even more enhancements in the works), researchers at Bing Ads believe it has significant promise, offering advertisers a more methodical alternative to current models. We'd love to hear your thoughts on the approach, so leave your comments below.