Culture is a stronger determinant of success with data than anything else. Including data.
[People + Process + Structure] > [Data + Technology]
It seems hard to believe. Yet, it is so fantastically true. At least for now. At least until AGI takes over.
Why is this formula material?
The first part of the equation, for better or for worse, improves in an evolutionary manner. The second part of the equation most frequently improves in a revolutionary manner.
The challenge for Senior Leaders is that revolutions seem a lot more attractive and hence they charge full speed ahead. This results in frustration, derailed careers and a massive amount of money flushed down sad places.
Revolutions in our context, almost always fail. Evolution works. Hence, it is dangerous to overlook the super critical importance of P+P+S.
You want to win big with data, with marketing, with transformative digital yada yada and blah blah, evolve. Do so at the fastest pace you can put in place for transformation of the left-side of the above equation, and use the same pace to evolve the right-side of the above equation.
This will ensure that the people, process and structure will be smart enough to take advantage of the smart and wizbang tech.
Maybe this metaphor will help make this real.
You can't give a toddler a Harley Davidson motor cycle. The moment your start the motorcycle, the toddler is going to start crying. It is not the mistake of the toddler, she is just a toddler after all. It is not the mistake of the Harley, it is a very cool motorcycle. The mistake is yours.
The toddler needs something to steady her, something she can push, something to exercise her legs to make them stronger. At some point, she would love a Harley (as her father that might freak me out, but I digress).
This post is organized into the following structure:
Digital Attribution's Ladder of Awesomeness
Wait, Wait, What the Heck is Attribution?
Digital Attribution's Ladder: Step Details
Step 1: Optimal Metrics
Step 2: Macro and Micro-Outcomes
Step 3: Assisted Conversions
Step 4: Standard Attribution Models
Step 5: Custom Attribution Modeling
Step 6: Data-driven Attribution Modeling
Step 7: Pan-Existence Modeling
Step 8: Nonline Controlled Experiments
Step 9: Advanced Controlled Experiments
Closing Thoughts
I'll say this again at the very end… As a Marketer or an Analyst, there is nothing you'll attempt that will be more complex and challenging than what you are about to read in this post. The spectrum of upgrades you have to make to your tools and data along with your people, process and structure, are likely to be unmatched.
That is why this is so much fun. I have a huge smile on my face as I'm typing this sentence, I get so excited about this stuff. If you follow the advice outlined, the most likely outcome is an increase in the slope of your career's graph as it heads up and to the right! :)
Let's go.
When it comes to your Digital Marketing and Digital Analytics practice, I've advocated slow and steady evolution.
The problem is sometimes you might not know what that path looks like, what the steps are. To address that, on this blog I've shared something I call the ladders of awesomeness – my view of what the entire evolutionary path looks like.
As an example, here's the Digital Marketing ladder of awesomeness:
Very cool, right?
It is not easy to linearize it all, the world is rarely that clean. But, you have an overall structure that can guide your strategy.
My recommendation… Partake in honest self-reflection, let that help you identify where on this ladder today, then, rather than shooting for the moon, figure out how to get to the next step. In taking that step, you should not just implement cool technology and do cool marketing, you should also invest in growing the skills, experience of your people and invest in putting scalable structures and processes to take advantage of this next cool thing. Win that, then go to the next step. Win that, then… well, you get it.
Cool technology plus savvy people to take advantage of the new possibilities plus processes to execute at scale set in the best-fit structure equals winning big.
My second ladder of awesomeness was very exciting as well. It lays out an evolutionary path for the key performance indicators you should use to drive digital sophistication inside your company. You'll find it here: Digital Metrics Ladder of Awesomeness .
It tells you not to go after Customer Lifetime Value right away. That is a insufficiently prudent use of Earth's oxygen. The metrics ladder lays out a path that will get you there, step by step while ensure your org is coming along with you.
Digital Attribution's Ladder of Awesomeness.
The other day, I had the amazing privilege of delivering a keynote with my point of view on attribution. The CMO expressed a desire for the audience to learn about advanced attribution strategies.
It is a topic I love and adore, but it is also a topic way more complicated than anyone is willing to admit.
Rather than simply give them all the advanced attribution modeling techniques, I took the opportunity to create a ladder of awesomeness for digital attribution. I did not want them to make the mistake of trying to achieve revolution at the end of the keynote, rather I wanted to give them a path to achieving a global maxima. One step at a time.
Here's the ladder I drew at the end of my keynote summarizing my worldview…
The overall execution I recommended was the same as in the case of my other two ladders of awesomeness :
1. Figure out what step you are at.
2. Check to make sure that your organization (people, process, structure) has maxed out the benefits of that step.
3. When confident that people, process, structures are helping you max out the complete value of that current step, go to the next step. Don't jump two steps! Just one step forward.
4. Buy new technology, if needed, invest in implementing it and using it, start to focus on getting your people, process, structure to evolve to take max advantage of this next step.
It is a mistake to believe that each step is the same "size" / requires the same effort or skills.
To illustrate this, in the space I had available on the slide I was projecting, I shared some sense of effort/skills/time that might be required to take one step up…
You can see that the initial elements are pretty small, then things get complicated, but it is not an even distribution. If you do Step 4, Step 5 might actually take less time. It is also clear that things get insanely hard as you get towards the end. Insanely hard is putting it mildly.
I am sure you are very curious, what each of these elements entail!
It is very hard to capture an entire keynote, and a life-time of bruises that the wisdom above reflects, in a simple blog post. The keynote contained solutions for each step, it would take too long. Let me give you a brief sense for each element, that should give you enough to explore in a much more focused manner.
But, before that…
Wait, Wait, What the Heck is Attribution?
: )
I'm sure it is clear to most of you, but for some of our new peers let me quickly explain, and then we'll explore all the elements in the digital attribution ladder of awesomeness.
Here's the simplest way to think about it. Most of us make decisions about the effectiveness of our digital marketing initiatives, owned, earned or paid, as if the real world looks like this…
Irritatingly we believe this because Google Analytics, Adobe, IBM and all other digital analytics tools tell us to believe that. They base all computations in their standard reports on an awfully silly thing called last-click.
Why do I say irritating?
Because the above picture actually looks like this…
Suddenly most of your standard Adobe and Analytics reports are more than lying to you about the effectiveness of your marketing investments.
The art and science of allocating optimal amounts of credit to each marketing channel, based on the activity it created, is called attribution analysis. The end goal is to recommend an optimal mix for your marketing budget.
Take a look at the first row above. Attribution analysis will help you understand how to value Social Network AND the Direct channel AND Organic Search.
Smarter attribution of the outcome, smarter marketing decisions.
Digital Attribution's Ladder: Step Details.
Getting back to our story.
My core recommendation is that rather than jumping directly to attribution modeling or media mix modeling, that you build a strong, step-by-step, foundation of people, process structure along with data/tools sophistication. Let's look at each step in the evolutionary journey.
If your company's dashboard is full of Visits, Time on Site, Impressions, % Exits, basic activity metrics then your company is not ready for attribution anything. You would think if you throw in Conversion Rate in there and you are ready. Nyet.
The most primitive thing you can do to have a very strong people, process, structure foundation is to pick great metrics to measure. Tough metrics. Smart metrics. Metrics that actually tell you if the business is doing well.
There are many ways to pick really good metrics. For example, checkout my list here: Best Web Metrics / KPIs for a Small, Medium or Large Sized Business.
Or, if you have a savvy digital strategy powered by my ultra-awesome See-Think-Do-Care business framework, you can use my recommendations in the framework to judge how optimal your current metrics strategy is…
Using these metrics, vs. the basic activity metrics like Visits and Time, is hard, taking advantage of them requires smarter people. Additionally, actioning the powerful insights you get from the above list requires smart processes and smart structure.
See what I mean when I say optimal metrics create the cultural and thinking sophistication required to do harder things? If you don't have this. Don't move forward.
Step 2: Macro and Micro-Outcomes.
A typical macro-outcome is an ecommerce order, a lead submitted for a B2B company, a new profile opened by a visitor to a content site, a donation on a non-profit website. So on, and so forth.
Most of you already measure the heck out of this. (If you don't, go back one step.)
Only a handful of you measure micro-outcomes.
Micro-outcomes for an ecommerce website would include store lookups, coupon downloads, new accounts, reports their users can download, email signups, reviews submitted, product amplification, videos watched, charitable efforts, blog subscribers, community celebrations, etc. etc. And, all of these are for just one brand's website. They make a few things like tooth-paste which are sold online, but the primary channel of distribution is offline stores. It is impressive to think that that aforementioned list are all the things they do online! We bring immensely smart nonline decision making for this client by optimizing for their macro-outcome (orders) and all these micro-outcomes.
Can you see how savvy the company's people, process and structure would have to become to allow optimization of a portfolio of outcomes, rather than just one (conversion rate)?
It is hard to do this. It is hard to compute the economic value of all these outcomes. It is hard to optimize for the entire portfolio.
That is how you get ready to do sophisticated things like attribution modeling.
Can you smell attribution? Close, but one more step before we get to it. First, let's get your org ready to use the metric that truly is the precursor for sophisticated attribution modeling.
In Google Analytics go to Conversions tab, then Multi-Channel Funnels and finally click on Assisted Conversions.
I love this report.
It is your org's introduction to moving beyond the awful last-click conversion obsession. In this report you'll see a more complete view of your marketing performance…
You are going to have a lot of arguments about which department (and people!) should get more credit, how to value the budget now that you have these Assisted Conversion numbers, why did Display go from $121 to $6 (!), so on and so forth.
As you resolve these issues, and start to take action by changing how much budget you spend on the channels above, you are collecting the elements required to be successful with online and offline attribution modeling.
You jump directly to attribution anything, a cold, hard wall is waiting for you to run into it.
Step 4: Standard Attribution Models.
Congratulations, it took you 18 months, :), but you are ready to do attribution modeling.
It is very easy to start. In Google Analytics, including the free version, go to Conversions, then Attribution, and then Model Comparison Tool.
You'll see Last-Interaction listed already. Next to it you'll see vs. select model. Click.
You'll see seven default models listed. Most of these models are for esoteric needs, or are flat out wrong. Take the First-Interaction model as an example. Choosing this model is like you giving all the credit to your first girl-friend for you marrying your wife. The definition of insanity.
There is just one model that passes all the smell tests, Time Decay. It provides reduced credit to marketing touch-points that are future back in the customer journey. Simple.
Use Time Decay for your first step into attribution modeling.
The red and green arrows to your right are helping guide your decisions related to the shifts in budget that you should consider in order to optimize your marketing and advertising to get the best possible results from your budget.
At this point, you'll be delighted that you listened to me and did Step 3 resulting in increased savvy in your people, process and structure. If you'd skipped that, at this stage all you would have is a clever report that has zero impact on your company!
Even if you did this as Step 4, you'll still require incremental investment in getting your org to understand the data above, you'll have to invent a new cultural norm of taking the red and green arrows above and creating tests from the recommendations, putting the tests into market and create a feedback loop of lessons that your org structure can learn from and improve future strategy.
It is a lot of work. Totally worth it because of the impact on cost-savings and increased profit.
Step 5: Custom Attribution Modeling.
Having completed all that hard work, and now that the org is making incrementally smarter decisions, you are ready to take advantage of your unique knowledge you have about your business, your customer behavior, and your strategy.
Custom attribution models allow you to take a base layer of smarts from Google Analytics, and add in yours.
For a client I've spent a lot of time with, here's the custom attribution model…
The reason for the choices above is business knowledge, customer behavior and business strategy. As you make the seven choices required above, you'll lean on those three elements – and lots of conversations with key business leaders.
The actionable steps you'll take from application of your model will be similar to the ones outlined in the step above.
Customer attribution modeling is incrementally better than standard attribution modeling. In doing this step successfully, you are strengthening leadership connections, and more buy-in from multiple departments (finance, sales, support etc.). It is not hard to imagine how critical that is to achieving success with data.
Step 6: Data-driven Attribution Modeling.
One of the painful things you'll run into while creating your custom attribution model is that persistent pain in the rear-end… Opinions.
Person x will say no, Avinash is wrong, we should not favor Clicks, my ads have no clicks, they only have impressions, change Avinash's model to over value Interaction Type Impressions . This person is wrong, and I am right. :) But, sadly you can't pull me out of your pocket so that I can tell them how wrong they are!
I kid only slightly. You are going to run into a lot of this. And, for some of these opinions you'll never have definitive data to prove the opinion right or wrong.
Where humans fail you, let Machine Learning come to the rescue.
Google Analytics looks at all of *your* data, all of the click-paths of your actual visitors, how each marketing channel delivers value to you (based on a success criteria *you* define) and helps create an attribution model that reflects your reality. This attribution model is called a data-driven attribution model. Opinions can now go live in a very dark place, while Machine Learning illuminates the world.
You click on the Model Explorer in the Attribution folder to see your data-drive model…
As an Analyst, I have to admit I get a special sense of pride when I see the shades of blue above. There is no way that a human can get to this level of insight, at this scale, or so frequently (your model is refreshed all the time with new data/behavior). I should probably be scared that these machines are making me redundant. For now, I am simply amazed.
You can see why Google wants you to pay for this feature (among many other great things in GA 360). It is smart, it is computationally intensive, and a competitive advantage for you.
Your data-driven model eliminates opinions/feelings/politics from the process of getting to the best model for you, and it is exquisitely yours.
Actions you'll take, changes you'll drive to your marketing budgets, will follow the patterns set in Stage 4 and 5 (which is why it is still important to go through the pain and build the right foundational P-P-S upgrades).
Bonus : If you want to learn a lot more about each attribution model in Analytics, pros-cons, how to use them efficiently, this post has a deep dive you'll love: Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models.
The next three steps in the ladder of awesomeness are complex and advanced. They apply to perhaps only the largest companies on the planet. Let me cover them here briefly, just so that you'll have a sense for them if you are in a large company or on your way to becoming one!
Step 7: Pan-Existence Modeling.
Unless you do something extraordinarily unique in your analytics execution, almost everything you do above won't be tied to a single person's behavior. It will be tied to cookies, it might be fractured by devices, browsers, and other things that make tracking a single individual difficult. It goes without saying that some of this might also be due to compliance to local laws (which I deeply stress you should read up and be familiar with for your local legal entity).
It is absolutely imperative to stress that even with all the limitations I've just mentioned, you are still better off taking the journey outlined in the above six steps when compared to being stuck in the awfulness that is last-click reporting.
Understanding how an individual human behaves does take you to a whole new level. Imagine data-driven attribution modeling that understands one individual's behavior across mobile and desktop! Now throw in the ability to tie that behavior to their activity inside your store or call-center (sales or support). #mindblowing
Imagine this stitched together view, across devices…
[More on the above report here: Develop a Smarter Understanding of Your Audiences]
What do you need to get to pan-existence attribution modeling?
Some way of identifying, incentivizing and tracking a person. The technology to do it exists. In the Google Analytics world, implement the Universal Analytics User-ID Override feature.
You can extend this to also tie to a CRM record you have for the person off a loyalty card or whatever global identifier your company has developed to identify its customers uniquely.
Everything in your analytics tool of choice becomes more powerful instantly, including attribution modeling.
This is hard to do, ask for help, there are tons of authorized consultants who can speed up your time to market/victory.
Step 8: Nonline Controlled Experiments.
I'm sure the above step planted in your head the thought of how you can attribute the online campaigns the impact that happens offline (remember 80%+ commerce in the US continues to be offline, even with Amazon becoming amazonish!).
There are two strategies to keep in mind.
First. You can attribute credit more directly (without Step 7) by leveraging the power of controlled experiments.
Here's a simple example. You already know that that non-brand PPC campaigns drive a ton of last-click and assisted conversion. But, you also know that online campaigns drive offline impact. But, how can you prove it?
Run a controlled experiment.
In this example we ran a four week test across a total of 11 test markets (covering 128 stores) and 39 control markets (covering 621 stores). A little picture for you to show distribution and the design for experiments savvy that you'll bring…
At the end of the test we proved that every $1 spent on non-brand paid search marketing drove $15 in store sales.
If you do this well, you can be even smarter. In this case we were able to identify that the $15 in sales was at a 22% contribution margin (an unheard of accomplishment in retail). Oh, and we were not done. We could also identify that the sales lift for product category X was 3.5% and for product category Y it was 2.31%.
Impressed?
Think of how incredibly powerful this can be if it is a part of your standard operating procedure on the web. Attribution of the effectiveness of your online advertising to drive multichannel results.
You can leverage the smarts of controlled experiments without any of the seven steps above, but it is easy to see how much measurement, analysis, marketing, people, process and structure savvy you need to pull this off, hence it is Step 8.
If you actually complete Step 7 successfully, you can take your controlled experiments up several notches, including tying online behavior to longer-term outcomes tied to a single human. Then, you can go back and customize your overall marketing portfolio to micro-segments of individuals with shared attributes. This is very much in the holy grail region.
Step 9: Advanced Controlled Experiments.
We eschewed attribution modeling above. We go back to modeling, but of a different type.
The most common implementation in Step 9 is media-mix modeling (or as some like saying, marketing mix modeling). Boiled down to its most essential it is the creation of a multivariate equation that when solved through the application of some delicious statistical regression, helps identify the optimal mix of your marketing portfolio.
Almost always, media-mix models include all your marketing – TV, radio, digital, etc. This allows them to be the go to source for CMOs choosing to drive unified strategic conversations.
There is plenty of art involved in creating media-mix models, and in the hands of an organization, or Agency, without the optimal people, process and structure, the results are no less garbagy then other opinion based strategies.
I believe that the best way to eliminate biases (or more usually opinions), I recommend a heavy use of complex controlled experiments, varying multiple elements (unlike just one above), as the optimal source of inputs required by the multivariate equation.
My biggest complaint about media-mix models, even the most sophisticated ones, is that, if you are executing See-Think-Care intent strategies, the thing digital is really, really, really good at (and traditional media mostly completely incapable of), media-mix models have a very hard time identifying value from those super valuable activities. They are biased towards short-term commercial results. Basically Do intent strategies.
Hence, a biblical belief that media-mix models are the word of God when it comes to optimal marketing investment is incredibly flawed. You will undervalue See-Think-Care business strategies, which in turn will mean you will not use digital to do what it is exquisitely qualified to do (ex: with See and Care help you build owned audiences!).
And, all because some Agencies and Companies believe in judging a fish by its ability to climb a tree . CMOs and Analysts at these Agencies are actively plotting against allowing the company's marketing to evolve to where the present is, and where the future will be.
With that little concern expressed, hopefully squarely lodged into your mind, I still recommend media-mix modeling powered by inputs from controlled experiments. The reason is simple. We all have to make money for our companies. And, media-mix modeling is an incredibly valuable tool in that quest. Just remember, it simply solves for the now and not the next or the long.
Bonus: If you would like a more complex view into the three strategic ways to frame the attribution opportunity, check out this post: Multi-Channel Attribution: Definitions, Models and a Reality Check .
That's your evolutionary ladder when it comes to solving one of the most complicated challenge you are likely to face as a Marketer or an Analyst. The spectrum of upgrades you have to make to your tools and data along with your people, process and structure, are likely to be unmatched by any other challenge in front of you.
That is what makes this so much fun, so satisfying as a career choice and so rewarding from a compensation perspective. There is literally no harder thing you can do. I hope that, when offered, you'll choose to accept the ring. :)
Good luck!
As always, it is your turn now.
Does your company/agency's macro approach to achieving the optimal marketing portfolio reflect a revolution or an evolution? If you've completed all the steps, which step was the hardest? What is the most difficult facet of attribution modeling to explain to your senior executives? Does your company prioritize evolution of people, process, structure as it drives new contracts and expenditure on tools/data? If you had to give our readers one advice from your attribution journey, what would you say?
Please share your feedback, critique, praise, wisdom and best practices via comments.
Thank you.
Via 
Thank you for another great piece of wisdom Avinash. I couldn't agree more, the struggle is real and almost all clients I've wokred with were they have wanted to bite of more ladder steps than they were ready to chew never really left their current step.
Something I noticed that you didn't mention but which really helps you get your business (or client) through steps 4-6 on the ladder is to upload real and relevant cost data for as many channels (or campaigns) as possible. Looking at red and green arrows showing slight changes in the number of conversions (as in your image) usualy results in "That's cool, … well let's get back to what we were doing" whilst visible changes to CPA way more often results in "Well could we do more of this and less of that instead?"
In a nutshell, proper cost data helps make the attribution model reports way more actionable, disregarding of whether one is using standard models or data driven models. This actionability in turn makes it a lot easier to get the organisation to accept and internalise the steps.
Henrik: YES! Yes, yes, yes. More times yes. :)
I am embarrassed that in this long arc, I totally forgot to stress the value Cost Data. In our Master Certification Course on Web Analytics at Market Motive, I have an entire 10 min section on the value of Cost Data when it comes to attribution.
I highly recommend this for all paid efforts. Typically these will include Display, Email, YouTube (assuming PPC will flow automatically if your accounts are linked). This would allow us to see numbers in rows where you see a dash at the moment (first picture in Step 4 in the post).
We see similar shifts that are currently visible for the last two rows, and that, as you mention, can drive valuable conversations and action.
Thank you.
Avinash.
PS: Here's an example of how the table looks after we apply cost data for our email campaigns:
This might be my new favourite post on Occam's Razor!
I am a little embarrassed to admit that I had not considered the many implications of analytics data on people and organizations. I have often given a Harley to babies. Reflecting back I can see now what a gap in my strategy was. Thank you Avinash.
Data-driven attribution models and controlled experiments were new for me. I have only theoretical knowledge on them. I have a new to-do list to get myself and my large company up to snuff on these strategies.
Amazing honest approach towards setting the tone and pace of expectations vis-a-vis for digital marketing spends. Love the easy to understand diagrams and feeling excited to use these as a guide.
Thank you Avinash.
Hello Avinash,
thank you for another great article. Can you share a little bit more details about the experiment about the online influence to offline sales?
Thank's
Miroslav: As you can sense, I've shared real data for a client with all the real results. That does mean that I can't share any more detail due to their permission.
But, using design of experiments strategy to set up these tests is not insanely complicated. It is good though to work with an experienced consultant who has experience setting up experiments and getting the few data pieces needed to be stitched together.
You might also find this to be of value: About offline conversion tracking
All the best!
Avinash.
Hello Avinash,
Thanks for a great article!
We struggle with this topic a lot in my company, because we have a catalog and other offline marketing programs. If everything were all offline or all online, life would be a lot simpler :) though perhaps less interesting. Just to complicate things further, we have a very long buying cycle (expensive product, considered purchase), and people use multiple visits to the web site via multiple means of getting there. (I always tell people that my company has the most interesting analytical and customer acquisition challenges around.)
We DO track micro-conversions, which has been very helpful. Any tips on how to incorporate offline into this type of model? Thanks!
Claire: These things get us paid tons and tons of bucks because they are complicated. It is ok.
Attribution modeling is all about the ability to understand multiple people visiting multiple times and delivering multiple outcomes. And, one of the things you can track is offline conversions.
Here's an article on how to do that in Adwords: Adwords offline conversion tracking
You can also use the Universal Analytics feature-set to send data into Google Analytics. A bit about that is here Analytics Data Import But, you will likely need an authorized consultant to help you out as it is just a bit complicated. You'll find some at http://www.bit.ly/gaac
Avinash.
Hi Avinash,
I am beginning the process of implementing this path to marketing attribution awesomeneess. Thank you!
Can you give me some advice on working with a 14-day free trial before purchase? We want to look at multi-channel attribution models pre-trial but $ is not received on trial. We are considering placing a micro-conversion here with an expected value but we know that different channels convert to paid at different rates – we would be optimizing with the faulty assumption that each channel's trials are worth equal value.
Another option is looking at assisted conversion and multi-attribution models at payment but now we have introduced all the channels that were used for visits after trial started.
Any ideas?
Love your blog! This article came at the perfect time :)
Bobby: My primary advice is to hire a GACP (www.bit.ly/gaac) as they will spend time understand your local situation and variables at play and give you optimal advice. Honestly if these things could be answered with a blog post comment reply, life would be easier! :)
Let me try to add some value.
You might be making this more complicated than it needs to be. If you take your average conversion rate and the average order size, you are likely to get really close to understanding how much value each channel is adding. Yes, the CPA might be not exactly perfect, but remember you have the raw conversions number that is uninfluenced.
At the end of the day, you don't really care about the conversion rate of one channel (that is last-click thinking), you and I are interesting in the role each channel plays in the conversion path (see the image for data-driven attribution). Hence you might not need the complexity you express.
Finally, remember that you have two segmentation options when you are doing attribution modeling. You have Conversion Segments (just click on that button in GA, see the sweetness that shows up), both default and you can create your own. Secondly, you also have Conversion – you can pick the conversion you want to look at. This might be of value.
Avinash.
Hello Avinash ,
Great builds and real fascinating stuff .
At the onset we should all thank a gentleman called Mr. Alvin Toffler who "originally" propounded the concept of User Generated Content and the rest as they say was history . The point is that it is all about insight . He foresaw the power of real time human interactions that would change the technology / marketing / communications landscape for ever .
Thus the debate on methodologies /tools / processes / consumer planning best practices / research best practices versus compelling consumer insight that would change the game . My suggestions are the following
1. Platform integration is critical – bringing all the "data" together – both structured & unstructured and I mean from a variety of sources .
2. Make the data work for you and NOT the other way around – meaning using multiple tools across multiple platforms perform specific tasks DISPROPORTIONATELY . Simply put some web analytics tools perform certain tasks brilliantly . Use only this tool set to analyze a set of data sets it can HANDLE best
3. Run analysis across platforms and tool sets handling "piecemeal" data sets
4. Integrate all the extracted "juices" together : metrics ,trends , segments , behavior , clusters , products , competition , disruption et al
5. Align all these variables together – meaning those naturally co-related and statistically significant . So we should obtain clusters of different variables mentioned in # 4
6.When the rubber hits the road – Test , test , test , hit the market , talk to consumers across spaces ( digital , offline ) – does the all clusters make sense ? can some of the clusters be used to drive disruption ? Can those drive compelling innovation ? Are the clusters making any sense at all " Or are they a plain cock & bull story ? If yes , then go back to the drawing board ! Or we take a different route . Good news for all clusters NOT making sense – consumers always do not know what they want – they need to be reminded or even better persuaded in case of a new innovation …..thus disruption & changing the game
7. If the clusters are making sense , then it is time to go way deep . The gold rush has just begun . It is just one cluster or two clusters that is making sense – meaning are we about to hit one gold mine or two gold mines ? Or is it a diamond mine ?
8. We are close to generating that one or two COMPELLING / PROFOUND insight
9. Turning the insight into services & products – Mind to market
These are my thoughts . Hope I am making some sense …..
Feels fantastic to be in a platform like this . Simply Awesome ….
Cheers ,
Sushen
Hi Avinash,
Thanks for this article. This is awesome!
Best wishes from Auckland.
Dj
Excellent article. Many thanks Avinash for sharing all these valuable insight.
Love the blog.
What are your thoughts on the value of Time Decay as the "only" baseline model for longer sales cycles? You use the analogy of giving credit for your marriage to your first girlfriend. I don't love the analogy on a number of levels.
Another (equally biased/flawed) analogy might be: using Time Decay is like giving most of the credit for your sale to your conversion funnel (instead of the methods by which you attracted the traffic to your site in the first place, or all the awesome things your site did to GET people into your conversion funnel).
Point being: if a sales cycle is typically 30-90 days, and a brand isn't widely recognizable, and it's a larger sale in general… we are not talking impulse buys of t-shirts or newsletter sign-ups here… my instincts are to weigh the initial brand introduction much heavier than you seem to want to. Let's say we aren't at the data-driven attribution modeling stage yet (it's coming… but probably 12 months away). Can you show me why my thinking is still flawed?
Thanks,
RFF
RFF: When I do attribution modeling, the two models I'll play with are Time Decay and Data-Driven. They provide the best contrasts, best openings for arguments. Of course Data-Driven is the right one. :) If I don't have access to Data-Driven, I'll use Time Decay and Custom Modeling.
I'll humbly disagree with your analogy for Time Decay. Here's a simple way to think about it: If your second or third or fourth touch point was so great, why is it that the seventh one resulted in a conversion?
So. They deserve some credit, certainly, and they are getting it in Time Decay. But, in a scenarios where we have incomplete data, what you know for sure is that the seventh resulting in closing the deal.
All that said, you don't need to buy anything I'm saying or anyone else. You are right to point out that you have instincts based on your experience and your data analysis. Build a custom model based on your instincts, one based on mine (!), and experiment. Let data help your instincts be validated – the blessing of the web!
I am so glad you are thinking about this in a critical way.
Avinash.
Thanks for your typically thoughtful reply! Given me a good initial direction to head in.
Hello Avinash,
How do you segment the sessions by See / Think / Do / Care in order to apply the proper metrics on those 4 segments?
Thank you.
Tryphon: You can segment them by any dimension you are interested in.
For example… One way to solve for Think intent by a customer is to deliver optimal content (as they have weak commercial intent). Hence, you can segment Page Depth by Landing Page. Or, with Think intent you want to accomplish micro-outcomes so you can segment Traffic Sources by Per Visit Goal Value.
So on and so forth.
Avinash.
Many thanks for this great article!
From our experience marketers get overwhelmed when having to look at macro and micro outcomes as well as assisted conversions. We implemented a report which would display all the assists a channel or campaign delivered, because we thought it necessary. We ended up ditching that report because marketing found it interesting to look at but couldn't derive any actionable insights.
As you already pointed out, Step 4 + 5 involves a lot of opinions and discussions about which standard attribution model should be applied, e.g. why a position based model should be better than time-decay model. For example countering your argument against first-click, I could argue that it was your first girlfriend which prevented you from becoming a monk and not marrying at all (we really had similar discussions).
Custom models aren't really that a big improvement in terms of leaving behind opinions. Different marketing departments might bring in different political interests, i.e. they want to make their channels look good. Data from different data sources, with different data quality to customize a rule-based model will be brought to the discussion. This can lead to an overly complex set of rules (if an organic search click happened x hours after a branded search click, give all the credit to the branded search click). Such a custom model takes into account the different goals of every department, but might become very hard to maintain.
Our approach is helping clients "jump" to data driven models as quickly as possible and to prevent them from spending too much time in the previous steps and respective discussions. Unfortunately not every digital advertiser is capable of spending what GA 360 costs. So that's where we try to help small to medium ad spenders as well as startups to apply data driven attribution models on the basis of machine learning algorithms. We are by no means as comprehensive as GA 360 and we are not targeting the same customers.
As Henrik Lundqvist mentioned in one of the previous comments, including real costs data indeed is key in making the insights actionable (and painful). Simply put: If I as marketer sees, that she waists USD 10.000 per day in a specific campaign, the motivation to change something is a lot higher than to see, that this campaign is overvalued by X percent in terms of attributed conversions. That's why we aim at integrating with all larger ad platforms via APIs to pull in the cost data automatically.
Full disclosure: I am Co-Founder of adtriba.com, a marketing attribution startup. Your articles played a big role in deciding to start this business and allow many more businesses to bring more transparency into digital marketing.
Janos: You are right, as this blog post outlines, there are shortcomings in each step, including data-driven, until you get to the very top of the ladder of awesomeness.
The challenge we are trying to solve is to ensure, with each step, that the people, process and structure can take advantage of what data and tools can offer in that step.
Else, the challenge is we can, say, sell a tool/data to a company for $50,000, they have no ability to use it, our long term return is whatever profit you can make from that one time sale. If on the other hand we also make the org smarter, figure out where they are on the ladder and move them up – while moving up their people, process, structure – we can make the $50,000, :), and then all the $$$ in consulting and tools savvy. :) Of course, the client also makes long-term gains from our efforts.
Adtriba sounds exciting, all the best!
Avinash.
Avinash,
Your take on attribution has been great help. We don't necessarily have a long sales cycle, but we do have a lot of overlap between ad accounts. Unfortunately we are a lead gen site too.
We are mapping subsequent conversions back to the User_id of the user, but are worthless for attribution modelling because when mapped back they are being done as direct / none and at the time of mapping and not the time of conversion.
Have you any experience of mapping back conversions from 3rd party sites to the correct source of traffic initially?
Help would be much appreciated
Steve: I am afraid this is a question that will require a little digging into and identifying what is going on at your end. The best strategy is to work with an authorized consultant who can help you address these issues very precisely. Here's a list: http://www.bit.ly/gaac
Using the lead submission as the initial criteria to assess value of each marketing channel is a very smart strategy. Then being able to tie back an offline conversion to online is of course also very clever. Running the attribution model with this new data as the success is a bit tricky, but I don't know how in doing that you lose the source channel. As I mentioned, a consultant should be able to help.
Avinash.
Hi Avinash, thank you for this article! You are right about this being a very difficult problem to solve.
I am wondering if you can share your opinions on the following:
I have embarked on the process of finding an attribution model that will get buy-in from all stakeholders and be used across my company. However, I am finding, like you mentioned, it is a time consuming process to say the least. I feel like if and when I am able to reach consensus
It is quite possible the technology/metrics/marketing tactics we are using and measuring are so different that a lot of the effort put in to developing the best model will be, well, obsolete.
So my questions for you are:
1. Is there a point of diminishing returns on the ladder, where step x is 'good enough' and any incremental benefit achieved from climbing more steps will not be outweighed by the time required to get there?
2. Have you seen many companies successfully get to step 9? And if so, is there a commonality among them? Do they have execs leading the attribution charge? Or a team completely devoted to developing the best model? Or is it something else?
Any insight would be much appreciated. Thanks :)
Andrea: First, you are very smart to worry about the point of diminishing returns.
There is one, it is very hard to know when you have reached it. Perhaps, when you have fewer fights about who deserves how much credit – fights that are purely driven by politics and self-preservation.
The ideal point of recognition would of course be the reduction marketing gains you are getting from driving changes suggested by the current attribution model. Honestly, this does not happen often enough. People!
I have indeed worked with companies that are on step 9 (else everything I am saying would be theory, and there is not a brand attribute of this blog!). They tend to be larger (because step 9 costs money). They have a culture where the CMO herself/himself directly gives many craps about maximizing for the Nonline Marketing Mix. They tend of the companies where controlled experimentation and modeling are a core part of the DNA for marketing spend that is material. And, as you can imagine, they do have a dedicated team to Nonline Analytics (I put the stress there and not on having an "attribution team," which is necessary but not sufficient).
One quick bit of advice for you. Don't try to convince *all* stakeholders. This is not a democracy, everyone does not get one vote. Align the most senior person who will talk to you, align with the one or two people who own most of the marketing budget. After that, drag everyone else along. I say drag out of love. Once you start showing that you can drive smarter decisions for the two people with the biggest budgets, the rest will come along (even if some will do so reluctantly).
I can't help it, one last, last thing. Make sure you tell everyone that with the attribution journey your goal is this: Be Less Wrong Every Day.
Maybe that will give you inspiration and tell everyone else not to get so hung up on things.
Avinash.
Thank you for the insight!
I also can't help it, one last thing. What do you see as the primary root problem attribution models are trying to solve?
I have found that on the surface it looks like an attribution model is the solution to optimize marketing spend but in actuality it seems more driven by the need to inform managerial discussions like resource allocation and performance reviews with data. Is this something you see as well? And if it is, do you think an attribution model is the best solution to that problem?
Thanks again!
Andrea
Andrea: You are trying to have an optimal mix when it comes to your marketing spend. That simply is the goal.
A part of that process is yes informing managerial discussions of all types. This is the output of knowing what mix of media works to drive the best value for the business.
Remember though, doing attribution modeling is just one step in your ladder of awesomeness. Keep going up and to the right.
Avinash.
Thank you, thank you, thank you!
Hi Avinash, Have been following your blog since last 6 years, and they are really insightful.
This particular post is my favorite one now.
There have been no structure around Digital Marketing implementation so far! This is the first one I've come across so far :)
Thanks!
I really love the Digital Marketing Ladder of Awesomeness.
This overall structure can really help me with my Internet Marketing strategy.
Thank you for sharing this.
i little typo: "Additinoally"
cheers, i will go back to read now :-)
Good one!
Thank you for sharing both a way to think as well as the practical tips to execute an effective strategy.
This is makes sense and resonates very well with my experience working on Attribution in EMEA.
One key issue in your recommendation is to jump to modeling without stressing the importance of data quality. Since cookies – including GA 1rst-party cookies – are only able to track sessions and not cross-device activity, the conversion paths that will be used for more complex modeling will most likely be shorter than what is actually happening.
Let me take an example:
1/ Alex is clicking on a ad on its mobile phone, visits the website and watches a video (micro-outcome)
2/ Alex is going back on the website on its desktop, buys a product
So what will the data looks like with this simple example (with longer cycles, it will be much longer)
1/ Paid ad > Visit with micro-outcome
2/ Direct > Visit with macro-outcome
In that case, the company will not be able to account for any assisted conversions and will over credit "direct" for this purchase. Any advanced modeling will fail to credit the paid ad because of the tracking technology.
My point : looking at "assisted conversions" and experiments may be helpful but conversion paths based on cookies may be hiding a whole different story.
Benjamin: You are absolutely right to stress the value of data quality, I do take that as a given and it is worth reinforcing.
Re cross-device, you'll see that I cover that in Step 7 "Pan-Existence Modeling" in the post.
Thanks!
Avinash.
Hey Avinash,
That's a steep ladder! We are just trying to climb up to step 3 with assisted conversions ;-)
When I look at what the pros do with the awesome Google Merchandise demo account I see a ton of assists and conversions from Direct
https://analytics.google.com/analytics/web/#report/bf-assisted-conversions/a54516992w87479473p92320289/ – can you give any tips as to what you do with Direct traffic when it comes to assisted conversions/attribution?
I just think Direct is in many cases a catch-all for missing referrer information rather than earned loyal repeat visitors.
Michael: : )
Sometimes, and this is important. the size of the business or the opportunity in front might not require any further climbing than going to step 3. I would include our startup (I suppose now medium sized business) Market Motive in that bucket.
As to direct… The tool reports what it captures. I recommend two things.
1. All the things mentioned in this post: Excellent Analytics Tip #18: Make Love To Your Direct Traffic
Doing all the little things to clean up our ability to collect referring data cleanly is always a good idea, imperative in this case.
2. Implement an ability to understand People (vs. cookies). User-ID-Override FTW!
This will help you understand if they are actually all people you know, or truly new people.
At the end of this you might not have a 100% perfect answer, but it will a gosh darn better than where we currently might be!
Avinash.
Always something new to see… Great tips!
We have been working through attribution at our company and this really helps us figure out why we have not made more progress. We are skipping steps, and we have not put the surrounding structures in place. We have the tools.
Work to do now. It is nice to read your blog.
Very long and informative post Avinash.
I'm sure many people will find this helpful. It was definitely worth the read.
Avanish it seems to me that platforms are key as Alexa Siri voice activation IoT systems (possibly) develop one 2 one dialogue using in-house like for like aggregate platforms. AI linguistics have to be direct root domain platforms.
Those that have or can invest in authoritative platforms specific to the root domain which then serves (and persists) the common elements of an aggregate brand platform have a strategy to be seen/ found where everyone one else will pay the ferryman's ever increasing toll. SEO will have to work with data mining to be relevant?
Avinash, Thank you for responding please allow me to break down what was a broad brush / ambiguous comment.
AI voice activation systems [NLP] will become ever sophisticated with front of till Siri Alex Google Home Alibaba IoT developing the economies of scale to accept partial IoT in-house tech security transfer of risk. Public trust is engagement & marriage front of till has the hight ground demonstrated by Apple [phone unlocking refusal] Alexa [removing Wells Fargo] Google [medical work] financial muscle, banks & insurance companies threw brand trust out with the bath water. The ability to have direct one 2 one dialogue will mean generic search terms specific to the root domain have an advantage for a year or two ? It seems to me that in some categories such as insurance, auto sales, loans, energy like for like aggregate platforms will be developed in house IoT with AI Q & A's This would seemingly render SEO at best to basic AI algorithms, SEO mindsets would be best placed in data mining synergies ? Would you agree with this longer term view on the future of SEO.
Find it wonderfully ironic that it takes AI to bring back empathy/courtesy : )
Peter: There are a couple different strands in your note, let me see if I can unpack.
There is no question that Marketers, as well Product folks at their companies, will invest tons in Machine Learning initially, and then surely the application of Artificial Intelligence to reach and engage customers (Marketers) and solve their problems (Product). Some of this will take the shape of IoT devices, others still might take advantage of crazy things like a mesh network of all the IoT devices you have access to already. From Patagonia to AAA Insurance to my local utility, the possibilities for transformation of why each exists are limitless.
I am not an expert on security et. al. related to all this, so I'll skip those bits.
To SEO… Gartner predicted that 30% of the queries will be on screenless devices by 2018. That itself will be hugely disruptive (no ML or AI there :)). The process of allowing "computing entities" (the googlebot or a ML/AI powered newer version) to understand what your company is, what products you have, what problems you are trying to solve will continue to be critically important. Some people call this SEO – I think of it more expansively. We will focus on giving more insightful information (why don't we send along customer identity as we understand it along with the product data so that it is easier to figure out which human we are the right answer for?). At the same time I believe these new ML/AI drive entities will be smarter and understand more with less (negating the need for the simple, basic stuff)?
All in all, exciting changes to come.
Avinash.
Dear Avinash,
I recently discovered your blog and appreciate very much your effort and what you are doing for the community. Thank you!
I have also read your interview on the Italian issue of WIRED from November 8th in which you were asked about your opinion regarding the American presidential election.
My question: what is the reason that so many analysts have not only failed to predict this outcome but wrongfully predicted a completely different result? Where was the error in the analysis and what can we learn from this case for our profession as marketers? Do we need to change the methodology? Where was the blind spot? I look forward to receiving your opinion!
Best regards and greetings from Berlin!
Johannes: OMG, I wish that there was a simple answer.
First on data… There were many factors, but two that I believe were big causes. 1. Lack of enough users sampled on mobile phones. This has hurt polls in predictability recently in Columbia and the UK, and I believe that was also true in the US. 2. The Bradley Effect. Not enough people who were going to vote for Trump admitted to pollsters that they would vote for Trump.
Second on the moment… There was clearly an angst in the US that was underappreciated by the media/opinion makers/Barack Obama/everyone else who was talking.
It is important to share two contextual things:
If just around 57k votes (across 3 US states!) had gone for Mrs. Clinton, she would have been President. Hence, the result is a lot less emphatic than the press and media seem to make it out to be. 57k is not a lot. The impact is just magnified when you look at the archaic Electoral College system here in the US (it is false to say that the US is a democracy, it is not).
Mrs. Clinton won the popular vote by 2 mil (and growing a bit as of now). A second reason not to read too much into a *massive shift in the US belief-system*!
Despite those two things, big changes to come to the US under Mr. Trump.
Avinash.
Thank you for your answer, Avinash.
I have heard of the Bradley effect, but was not aware of the fact that not enough mobile phone users were sampled. If I understand right, does this mean that a big chunk of the electorate just has not been recorded in the polls?
It's a very long article but it's worth the read. I like how informative and detailed the whole post is.
I'm a newbie in the SEO world and I entered without knowing a thing about digital marketing and other related stuff. I'm finally and slowly getting the gist of the work. Will definitely share this with the management of my company.
Thank you for this and all the best!
A kind of long blog post but when you finish the reading you'll find a very interesting and informative post.
Awesome in the whole view.
Thank you for writing in such detail and providing a framework we can action.
We have B2B offering and we linked CRM with GA to see exact customers and their behavior. Even we created custom attribution model to compare channels.
All are fine but the problem is that GA allow only to user level data for 93 days. so that we are unable to understand what are the content they consume in initial stages and channels because sales cycle is around 9 months.
If we stop a channel based on Sales data (CRM) we might end up by stopping channels which brought users to the site. This is limiting us to get greater insights.
Do you have a workaround for this. appreciate your support on this.
Niroshan: I'm afraid I don't know exactly the limits from within the GA interface, but do please explore the free GA API for this as it has bigger limits and more fields of data you can download.
Another possibility to consider is to combine User-Id-Override in Universal Analytics with your CRM database and send an identifier back into GA (as GA accepts data as well) and track the long sales cycle – and original source etc. – in GA.
Ping a GACP as they might have more ideas for you, http://www.bit.ly/gaac
Avinash.
Thank you.
As someone just entering this field, this article is a lot to take in, but I look forward to the growth and challenges ahead. I will for sure be back for more reading!
Man you have explained this in an awesome way and the diagrams awesomely take the article forward.
Thanks for these great in-depth points.
I am new to this field but your post gives me much confidence to move ahead fearlessly.
Wonderful and thought provoking blog post.
Thanks for sharing.
Suresh
Hi Avinash
Excellent post . I have a question . Do you think that when framing the argument for attribution change within large de centralised organisations that you should start with writing this into data governance policies ? Basically make it a sin not to use your data the right way
Like you've said above I cannot believe the amount of companies that are using LC as the model of choice and not taking digging deeper to unearth gems and insight around coustomer journey !
Thanks
Austen
Austen: Yes. Very big yes. :)
You can ease the way you get the org to move forward. Show the Path Length report. That will instantly show how terrible last-click attribution is.
Avinash.
This is another of your thought provoking gold mines Avinash.
Looking forward to your next riff with Mitch Joel.
With many thanks
Allen
Hi Avinash. This is a great in-depth, well-researched article.
You raise a really interesting point here, about evolution versus revolution. It could be said that one of the reasons companies, agencies, and so on favor revolution over evolution is laziness. I’m not saying it’s necessarily easy to make a revolution to the way you do things, but it’s certainly easier than making an evolution.
If something’s not working, or not getting the results you desired, it’s much easier to scrap the whole thing and try something new (a revolution), as opposed to actually taking the time to identify what’s not working and improve it (evolution).
Revolution might be the easier choice, but evolution is usually the smarter one.
I’ve been a reader/evangelist of your blog for 10+ years. I truly appreciate the journey you’ve taken me on which has taken me around the world helping companies drive better digital performance.
You are the person who first introduced me to the goodness that is micro-conversions and the multi channel funnel (MCF) joy in GA.
For the last 3 years, I’ve been a cog in the Google AdWords money making machine in the UK. Sadly, very little use was made of analytics by 90% of my colleagues. I often used these MCF to help optimise campaigns. Shock, horror – I often used the first click attribution model to illustrate that last click attribution was not the end of the line in terms of understanding campaign performance.
I think you are doing a disservice to this model when you used this analogy:
“Take the First-Interaction model as an example. Choosing this model is like you giving all the credit to your first girl-friend for you marrying your wife. The definition of insanity.”
Perhaps I am insane :)
However, I think a more accurate analogy would be.
“Choosing this model is like giving your friend who introduced you to your wife credit for marrying your wife.”
Or perhaps a more modern example:
“Choosing this model is like giving Tinder credit for you marrying your wife (assuming that’s how you first met her of course)”
To my mind, it’s an indicator of how people first discover a brand/company, something which I believe is very important, after all, you will not marry unless that discovery process happened at some point previous.
Am I crazy?
Dominic: First, please accept my sincerest apologies as my reply is incredibly late. Your comment was stuck in spam and I only found it yesterday.
Having done more attribution modeling than I care to admit, across all continents, I stand by the metaphor when it comes to giving all the credit to first click.
I want to clarify that I am not averse to giving first click some credit. For example the standard model I encourage you to start with, time decay, will end up giving first click some credit (if first click is not too far away from reality).
It is important I stress that all of this is not relevant now that we have data-driven attribution. We can set feelings and opinions aside and use DDA which uses massively smart machine-learning to analyze your data, for all customers, across all their visits (including cross-channel) and uses actual behaviour to attribute credit. It that intelligent model assess it is all first, it'll send all the credit there!
Avinash.
Better late than never, Avinash! :)
I agree data driven attribution shows a lot of potential for helping us allocate spend optimally across the funnel from upper to lower – however it was quite a surprise to me when I learned of a limitation with the data driven attribution most commonly used in Google Adwords. It doesn't account for Display/Youtube at all!
So it seems to me that due to the challenges of the tracking complex/long customer journeys that I believe are often initiated by upper funnel campaigns are often overlooked by performance marketers
So for a marketing manager of a medium sized company like me – Adwords DDA is realistically the only option available (analytics/attribution 360 seems pricey) – and this option completely ignores display/video ads!
I am exploring the possibilities of DoubleClick tracking via Spotlight/Floodlight – but making the transition from Adwords to Doubleclick frankly makes my brain hurt :(
My main attribution insights are from the MCF report in Analytics but without Analytics 360 using data driven attribution to drive my campaigns seems a pipe dream at the moment… any suggestions?
Dominic: There might be some confusion here.
DDA accounts for all clicks and sources that send those clicks. That would include DoubleClick, Youtube, Facebook and even TV if you code your tracking parameters optimally.
DDA in GA is the only attribution platform where you'll have your owned earned and paid media all in one place. The other solutions you mention will be missing one or more of those pieces.
If you have a serious business, you have to spend some money to earn insights. I do realize that this is not an option for everyone.
Good luck!
Avinash.