Attribution is any method of understanding the customer touchpoints before making a purchase, and deciding which touchpoints were influential. We’ve had attribution tools in AdWords in a long time in the form of Search Funnels, but the new introduction brings us attribution modeling.
Attribution modeling has become a fix-all term over the last few years. Companies offering it haven’t been able to accurately articulate exactly what you’ll do with it, and lots of companies have thought about it without taking the plunge.
At the moment it’s half a solution. It gives you more information and more insight, but can offer little to no direction into what to do with it.
Attribution modeling is super valuable and the various tools that offer it do a lot. But, like all things, you need to have a plan first.
What’s New in AdWords?
Modeling tools have been in Google Analytics for a while, so hopefully you’ve all had a chance to look at them a bit. The AdWords data includes only AdWords touchpoints, so it’s a bit more limited. But let’s look at that data in the context of a user’s path to conversion.
Many users will interact with your ads (and site) more than once on their journey to converting. Not all of those touchpoints had any impact on their likelihood of making a purchase with you.
The goal of attribution modeling is to choose different weighting options that might reflect how a person interacts with your brand. By choosing one that fits better with the behavior of your users you can see which of your touchpoints were costing you a lot of money that was better spent elsewhere.
1. Last Click
Last click attribution is the standard. This model has already existed for a long time and will continue to do so. It gives the entire value of the conversion to the final click. Simple.
This is appropriate in a lot of cases. You might be in a market where people decide very quickly and easily, or make impulse purchases. You may be in a scenario where people research and then look for suppliers separately. If being present in the first stage wasn’t important, then you want to know which keywords (and ads) were present as final-stage influencers.
The downside is that you lose all credit for keywords that generated the person’s interest in your brand in the first place.
2. First Click
If you do credit all value to the final click then you might find your total conversion volumes dropping as you start turning down (or off) keywords that seemingly led to no conversions.
By using first click attribution instead you can see that the other way. Which keywords helped users learn about my brand in the first place? How did they originally find me, and let me make sure I give those keywords the credit they deserve.
It’s compelling for some markets, particularly those where purchases might be delayed. If I sell a non-essential product that people might wait until payday to purchase (a new jacket, maybe?), then I could expect them to revisit my site several times in between finding the jacket (and making the decision) and finally purchasing.
Since the first click was the decision maker I would want to give that the credit.
But what if the first click was important, but not the decision maker? What if the user maybe wanted to compare prices?
This is confusing, because position-based attribution is actually the catch-all term for all of these kinds of attribution where the position in the funnel is the determining factor. But I’m going to gloss over that a bit and talk about the position-based model in AdWords.
In this model you will basically credit both the first and the last interactions, and downweight all the interactions in between. The result is that you see which keywords drove the awareness of your brand and which keywords they used to make the final decision. If you had missed either of those interactions (e.g., you weren’t present on those searches) you would have missed the sale.
This makes sense when you talk about users who compare prices. This is a pretty clear and obvious research cycle:
“skiing holidays” > “skiing holidays in france” > “french ski break in march” > “france skiing march 10″
If the first click helped the user decide to go to France for their holiday, then I need to give that some credit. But if the final search was a price comparison, then if I’d missed that then I might have missed out on their booking.
But do we really want to downweight all intervening interactions so much?
4. Time Decay
If a research cycle is particularly long then which companies the user saw at the beginning might be fading from their memory.
Having had a presence early on could still build some credibility, but the later searches were the ones that really decided them.
Time decay models give more weight to more recent interactions, and reduce the credit given to earlier interactions.
It makes sense if the user is taking a long time to research and we want to be on shortlists but not over-prioritizing searches that aren’t really very focused, but we still don’t know which of our touchpoints was the decision maker. We might have removed it!
The most obvious of attribution models, linear attribution gives equal weight to every touchpoint.
It sounds good that we never lose the effect of any of our keywords. We don’t know which ones we could have safely removed and still made the sale, so give them all something.
But how helpful is that? In this case we know definitelythat we’re giving credit where it isn’t due. Our best influencing keywords will look poorer and our worst influencing keywords will look better. So are we able to make good decisions from this?
What’s Wrong With Attribution Models?
I don’t like:
Choosing a Model Without Good Reason
Choosing a model is a case of luck. Every user is different. Every user’s path to converting is different. Every user’s influential touchpoint is different.
Whenever we act on the information from an attribution model we are always either: upweighting certain touchpoints that we think influenced people or downweighting those we think didn’t.
So unless every user was influenced by the same touchpoints we are absolutely making our brand less visible for some users at key moments. What will attract more users like User A will definitely attract fewer users like User B.
That’s bad. It might play out well overall and there might be more As than Bs. But we don’t know that in advance and we can’t split test it. Time-series testing is bad. Without a split-test you reallydon’t know.
Before you can choose a model and act on it you absolutely must be confident that you’re reflecting the most common paths your users are taking.
Position-Based Models in General
In the bigger picture terminology, position-based models refer to attribution models that credit interactions based on their position in the path. Every model in AdWords fits into this category.
But in reality attribution is a million times more complicated than that.
How much credit should your brand keywords get? Brand keywords are generally navigational (e.g., “take me to the page” rather than “show me people who can sell me X”). With that in mind, was brand ever an influencer for you?
So really you’d want to choose a model that downweights brand wherever it is in the path!
How about if the same touchpoint is used twice in a row? If they interacted with your brand then repeated the same search to come to your site that sounds like they are just trying to get back to you via the method they found you before. So should you downweight any repeat searches? If you do and you’re not present the second time will you lose that person? Maybe. Sometimes.
What are your unique selling points vs your competitors? If you know you’re the cheapest in the market, then you really want to be present for any searches that are obviously price-sensitive, right? That’s where you’re going to be more persuasive than they are. If you offer the best quality, then any searches including the term “best” would be worth upweighting in your model, because they probably reflect the more influential touchpoints.
What Should You Do?
This is the tricky bit.
Because AdWords attribution models only include paid search touchpoints, there’s a part of me that says: nothing. Just go look at the data in Google Analytics instead.
But even if you do that, it isn’t clear whether any actions you take as a result are genuinely safe.
Look at your attribution models in the context of your multi-channel funnels, your paths to conversion, etc. Use it as a tool to see which keywords did really well if you cut brand out of the equation. Or only credit X keywords if they’re first click.
Most importantly, think before you act.
Make sure you can logically explain the reason behind anything you can see before you act on it, so you can understand why it would or wouldn’t behave the way you expect it to. And remember that if you apply model A to the data then act, it will look better under model A. Not model B.
Don’t expect optimizing for first click data to make your last click stats improve. They will not.