• Stop being a LURKER - join our dealer community and get involved. Sign up and start a conversation.

Attribution: Do you give the last touchpoint all of the credit for a sale?

And if I were a dealer I'm going to do what with 80 touchpoints exactly?! Actionable recommendations are what? Of course you two are going to agree, you're pimping the same product. LOL!


Man I sure would be pissed if I was paying cars and autotrader a bunch of money...



Ben, thanks for taking the time to post this. It leads me to my second set of questions.

1. We already know that almost 100% of people coming into the dealership visit websites first. What is this telling us we didn't already know?

2. What decisions large or small do you think you can make with this?

1. We already know that almost 100% of people coming into the dealership visit websites first. What is this telling us we didn't already know?

Answer: Ya to be clear, this is just one customer example, we have a lot of examples that have a lot of 3rd party interactions in them with little to no paid search/organic or direct website visits. My sister in law for example exclusively used Edmunds to buy her used Subaru Forester. (Before you ask I don't have her click path)

2. What decisions large or small do you think you can make with this?

Answer: The point I was really trying to drive home is its really about the entourage effect marketing has. Looking at singular points in the 80 things we saw wouldn't tell the story correctly. IF we went by first touch (organic in this case) we'd ignore 79 things. IF we go by last touch (paid search) we ignore 79 things. On that same note, we can't just look at the data from one singular customer point of view either, when we view the data we are really looking for patterns that emerge amongst all click-paths collected, for example if all this dealers click-paths looked identical to 'Shawn' then we would come to the same conclusion you did about 3rd party auto, but as a one off customer experience this wouldn't be actionable (unless maybe Shawn has an identical twin ha).
 
Shawn's Event Trail from Clarivoy's raw data:

Organic Search (Google) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Organic Search (Google) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Cross Family Visit > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Facebook Campaign Visit > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Paid Search (Google) > Website Visit (Direct) > Paid Search (Google) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Website Visit (Direct) > Paid Search (Google) > Purchased: 2015 Infiniti Q40

I see good validation and cancel-worthy validation of varying marketing channels with months of this data. I also see a case to prove to GMs and dealer principals that my digital marketing net needs to be wide. Assuming Clarivoy, or something similar, isn't charging as much as some of the marketing channels I might be canceling I'd want to look at this data at least quarterly.

I do see actionable data on the overall aggregate and maybe there is something in these details that can't be shown to us on DealerRefresh.
 
  • Like
Reactions: BenHadley
I see good validation and cancel-worthy validation of varying marketing channels with months of this data. I also see a case to prove to GMs and dealer principals that my digital marketing net needs to be wide. Assuming Clarivoy, or something similar, isn't charging as much as some of the marketing channels I might be canceling I'd want to look at this data at least quarterly.

I do see actionable data on the overall aggregate and maybe there is something in these details that can't be shown to us on DealerRefresh.
You might see it, but your average dealer doesn't. They don't have the time for it and expect (there's that word again) for it to be done for them by the attribution group. It happens time and time again.
 
You might see it, but your average dealer doesn't. They don't have the time for it and expect (there's that word again) for it to be done for them by the attribution group. It happens time and time again.

That's all in how one designs the UI. A successful technology company has to have a simple UI in order to bridge the chasm from early adopters to the early majority.

Moore-Contract-automation-lifecycle-adoption-curve.png
 
There is a framework to make decisions using Markov chains and studying the removal affect. We built this last year in our effort to make sense of what Ben is talking about.

Markov-1.png


Every customer journey (sequence of channels/touchpoints) is represented as a chain in a Markov graph where each node is a state (channel/touchpoint) and each arrowed link between nodes represents the probability of transition between states.

By computing the model and estimating transition probabilities we can attribute every channel/touchpoint. The probability of conversion of the complete model is 33.3% (0.667 * 0.5 * 1 * 0.5 + 0.333 * 1 * 0.5)

The last step is to estimate the impact of each channel/touchpoint on our final conversion. Here, we apply the principle of Removal Effect for analysis. The core of Removal Effect is to measure how many conversions (or percentage of conversions) would be lost if we remove one of those channels. In our simple example, suppose we take out C1.

Probability of conversion after removing C1 is 16.7% (0.333 * 1 * 0.5) or a 50% reduction
 
There is a framework to make decisions using Markov chains and studying the removal affect. We built this last year in our effort to make sense of what Ben is talking about.

Markov-1.png


Every customer journey (sequence of channels/touchpoints) is represented as a chain in a Markov graph where each node is a state (channel/touchpoint) and each arrowed link between nodes represents the probability of transition between states.

By computing the model and estimating transition probabilities we can attribute every channel/touchpoint. The probability of conversion of the complete model is 33.3% (0.667 * 0.5 * 1 * 0.5 + 0.333 * 1 * 0.5)

The last step is to estimate the impact of each channel/touchpoint on our final conversion. Here, we apply the principle of Removal Effect for analysis. The core of Removal Effect is to measure how many conversions (or percentage of conversions) would be lost if we remove one of those channels. In our simple example, suppose we take out C1.

Probability of conversion after removing C1 is 16.7% (0.333 * 1 * 0.5) or a 50% reduction
Nicely done, are the outcomes shown within your UI? If they are, that is fantastic and you are light years ahead of the game. I've never seen these kinds of options. I saw your universal benchmark options too, are those in your attribution tool?
 
Last edited:
There is a framework to make decisions using Markov chains and studying the removal affect. We built this last year in our effort to make sense of what Ben is talking about.

Markov-1.png


Every customer journey (sequence of channels/touchpoints) is represented as a chain in a Markov graph where each node is a state (channel/touchpoint) and each arrowed link between nodes represents the probability of transition between states.

By computing the model and estimating transition probabilities we can attribute every channel/touchpoint. The probability of conversion of the complete model is 33.3% (0.667 * 0.5 * 1 * 0.5 + 0.333 * 1 * 0.5)

The last step is to estimate the impact of each channel/touchpoint on our final conversion. Here, we apply the principle of Removal Effect for analysis. The core of Removal Effect is to measure how many conversions (or percentage of conversions) would be lost if we remove one of those channels. In our simple example, suppose we take out C1.

Probability of conversion after removing C1 is 16.7% (0.333 * 1 * 0.5) or a 50% reduction


giphy.gif
 
Nicely done, are the outcomes shown within your UI? If they are, that is fantastic and you are light years ahead of the game. I've never seen these kinds of options. I saws your universal benchmark options too, are those in your attribution tool?

I will tell our story on Friday next week and trust me when I say I am not promoting an attribution tool, just sharing what we learned trying to build one. We went deep down the rabbit hole working on this problem and learned a lot. We discovered some blockades from making everything useful so a dealership can sell and service more cars, which I feel from a technology side are not getting any better.
 
By computing the model and estimating transition probabilities we can attribute every channel/touchpoint.

Hey @jon.berna, will you explain this phrase to me? Not asking to be a jerk and this isn't a "gotcha" attempt, I'm here to learn.

It seems to me that confidence in the complete model relies entirely upon confidence in the accuracy of the estimations of transition probabilities. Is that correct? In other words, if the estimate of transition probability from C1 to C2 is inaccurate, the entirety of the equation is flawed. This would then be exacerbated by the potential error rate for each estimation in series as it compounds the error and variability of the equation.

1. Is that a fair assessment of the analysis above? What am I missing?
2. If so, what steps did you take, tests did you perform, behavior did you observe etc. to assure that your estimations were within an acceptable statistical error rate?
 
  • Like
Reactions: Alexander Lau