Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models

balance There are few things more complicated in analytics (all analytics, big data and huge data!) than multi-channel attribution modeling.

We have fought valiant battles, paid expensive consultants, purchased a crazy amount of software, and achieved an implementation high that is quickly, followed by a " gosh darn it where is my return on investment from all this?" low.

A lot of that is because of all the stuff we don't know. There is lots of missing data. And as if that were not enough, there is lots of unknowable data. Neither of which has stopped Gurus and Masters and Agency High Priests from trumpeting here's the next thing directly from Lord Krishna that will solve all your problems.

So, let's apply Occam's Razor to this complicated challenge. Let's try to make some sense of it all.

By the time you are done with this post you'll have complete knowledge of what's ugly and bad when it comes to attribution modeling. You'll know how to use the good model, even if it is far from perfect. I'll close with a custom attribution model into which you can insert all your biases – sorry, I mean expertise – and get something better than good to make incremental progress from where you are today.

My macro goal is to make you dangerously informed. By the end of this post, if you pay attention, you'll know the often hidden nuances and you'll be dangerous to any analyst/consultant/vendor who walks into your cubicle/office with I've got the God's-gift-to-humanity, easy-to-implement solution with insights riding out to you on a Unicorn.

Here's the outline of our incredible multi-channel attribution modeling adventure:

Excited? Grab a Red Bull. Let's go!

Three Unique Attribution Challenges.

In a recent post, Multi-Channel Attribution: Definitions, Models and a Reality Check, I outlined three distinct attribution challenges.

MCA-O2S covers the challenge of attributing the offline impact (revenue/brand value/butts in seats/phone calls/etc) driven by online marketing and advertising.

MCA-AMS covers the challenge of attributing accurate impact of our marketing and advertising efforts across multiple devices (desktop, laptop, mobile, TV).

MCA-ADC covers the challenge of attributing credit to all digital marketing channels (Social, Display, YouTube, Referral, Email, Search, others) that contributed to a particular conversion (or multiple conversions).

digital marketing path to conversion1

In this post we are going to take a close look at MCA-ADC. Multi-channel attribution across digital channels. Looking at the picture above … we've spent money on Social, Direct, Search, and Referral efforts and received 767 conversions. But how do we distribute credit for the conversions across all those channels?

All three challenges are important. I strongly encourage you to read the post and deeply understand all three and what your marketing and measurement possibilities and limitations are.

Do You Have an Attribution Problem?

Yes.

Sorry, I meant to say it is highly likely that you do.

It is a pretty easy question to answer. I normally ask people to look at the Path Length report in the Multi-Channel Funnels standard report in Google Analytics (or equivalent tool if you are using SiteCatalyst or WebTrends or other web analytics tools).

If a significant percent of your conversions have a greater than one path length, you have an attribution problem. Combine that with the excellent multi-channel conversion visualize (in the Overview section) and you have yourself a view of your marketing that will freak you out.

path length multi channel conversion visualizer

It is also ok to weep a little at this point as you realize the extent to which every single decision you've made about allocating your marketing budget is awful. Weep a little for that inconsiderate "friend," last-click attribution.

[One of my favorite parts of this Venn -diagram is the implications on organization structure. Some CxOs see it immediately, other times I have to walk the horse to the water and force it to drink. The outcome in either scenario is a restructuring of the organization that is exquisitely geared towards taking advantage of portfolio optimization. Related implications of what you want to do in-house vs. out source to an Agency. Really fun stuff, really long- term strategic implications. From a Venn -diagram. Who would have thunk?]

The Best Next Steps/The First Best Steps.

The simplest way to start is to look at your Assisted Conversions report in Google Analytics. Look at the last column: Assisted/Last Click or Direct Conversions.

• If you see a value less than one, that channel has a higher tendency to drive last click conversions. Hurray, hurray!

• If you see a value greater than one, that channel has a propensity to be present earlier in the conversion cycle. These channels are getting zero credit in last click attribution platforms (read that as: all standard reports in all web analytics tools). O. U. C. H.

At this point you should educate your management team on this specificity. "Look we might not be valuing all the performance we get from our marketing channels. Here are the specific channels that we are undervaluing." (Where the ratio is greater than one.)

You can even use that column to adjust some of the budget allocation right now, without any attribution modeling, and measure the outcome. It is imperfect, but it is such a simple first step.

It is likely your CxO will want you to explain which channel comes first ("introduces our brand to the customer"), which channel comes second ("nurtures our potential customer"), which channel comes fourth, fifth … and last.

You can use the Top Conversion Paths report.

top conversion paths

It is very important to point out that this is a completely foolish exercise to undertake. For the same reasons that path analysis is a waste of time. There are too many paths, and you can't actually control the path that a potential customer can take. Even if, and this is not possible, I said to you that the path is Direct, Social, PPC, Organic, Referral for 5% of the site traffic … what would you do? It is not possible to force people down that path!

But show the actual report. Let them arrive at the obvious conclusion. Be a hero. :)

The next question will be, what are the best ways for us to allocate credit to all our marketing channels properly?

I'm glad you asked, Ms. Executive.

Multi-Channel Attribution Models.

There is a free tool inside Google Analytics called Model Comparison Tool. It is sweet. It allows you to attribute credit to all your digital marketing channels involved in conversions (macro and micro conversions). You can visualize the impact of applying three models at one time.

For example, what if we used a linear attribution model instead of last click?

last click vs linear attribution model

OMG! OMG! OMG! So cool!

All I have to do is look at the very last column and look at the green and red arrows and get guidance about how I should shift my budgets?Yes!

OMG! Really?Yes.

And you are telling me that the Cost Per Acquisition for my display campaigns is not $201 but rather a lowly $155?Yes.

Get. Out. Of. Here! That is so cool. Finally my amazing blinking hit the monkey display ads are getting all the credit they deserve!

Time to burst your bubble just a little.

The tool is actually that good. Apply the right model and you will not only distribute conversions across multiple touch points, but you can also look at the impact on the CPA (this really is OMG, I peed in my pants a little cool). You can even get great first-step guidance about how to rebalance your portfolio from that last column.

But the weakest link in the chain is the attribution model you use. The recommendations you get are only as good as the model you use.

With that in mind, let's look at the standard models available inside Google Analytics (and some of the high-end analytics or attribution analysis tools).

digital marketing path to conversion1

Just so we have a visual guide through this learning process, let's use the above image as a reference. Look up, memorize the steps to conversion. Ready?

last click attribution modeling 1

1. Last Interaction/Last Click Attribution model.

This is the standard attribution model in all web analytics tools. It is applied to all the standard reports you see.

[The only exception to this rule is Google Analytics which, and I deeply passionately hate this, applies the #2 model below in all its standard reports.]

You can see why this model is silly. If 767 people converted as a result of the above experience, saying that all the credit should go to the Direct channel is silly. [Bonus: Learn more about what direct traffic is: Make Love To Your Direct Traffic.]

Social, Organic and Referral were also involved. We should figure out some way to identify their contribution to the conversion process, because they were involved in some form.

Historically, all tools used last click attribution because the one thing they could confidently say is what drove the converting visit. And they did not have the technical horsepower to do Visitor-centric analysis. Both these problems are solved now.

The only use for last click attribution now is to get you fired. Avoid it.

last non direct click attribution model

2. Last Non-Direct Click Attribution Model.

Google Analytics is bipolar.

All standard reports in Google Analytics give 100% of conversion credit to the last "campaign" prior to the conversion. Campaign is defined as anything but Direct traffic. So, the campaign could be Social, Organic Search, Email, Display, Affiliate, Referring Site … anything really.

This deliberately understates the Direct visits that lead to a conversion. In our picture below this model would say all credit goes to Referral.

This is imprecise. Why give credit to a campaign if it took me another visit where I remembered your URL and typed it in and came to your site? Why should the visit where, say, I saw a great promo or you recommended something based on my prior visit not get some credit for the conversion?

Why undervalue Direct? Why undervalue a marketer's efforts to create brand recognition and brand value?

I believe this is a mistake. A historical legacy, perhaps. It should be courageously fixed.

Bonus: This model is also the irritating reason why none of your standard Google Analytics reports match your standard Multi-Channel Funnels reports, even if you look at conversions in the standard MCF Overview or Assisted Conversions reports.

last adwords click attribution model

3. Last AdWords Click Attribution Model.

My words for this model might get a little bit vitriolic, so I'm going to keep my mouth shut.

And to think you never thought that was possible. : )

This model is profoundly value-deficient. There. I can be nice.

first interaction click attribution model

4. First Interaction/First Click Attribution Model.

Reverse of last click. Rather than giving all the credit to the last click, give all the credit to the first click.

In our example above, switch 100% of the credit from Direct to Social.

This is a gigante mistake.

First click attribution is akin to giving my first girlfriend 100% of the credit for me marrying my wife.

Makes no sense, right?

If the first was so awesome, how come I needed #2, #3… to get to the most perfect person – I mean, campaign :) – for me?

With last click attribution there is at least some certainty that something about that campaign, something about that visit to the site, resulted in a conversion. With first click you just have faith. Or a HiPPOs (Highest Paid Person's Opinion) fervent "gut-feel."

Eschew irrationality.

linear attribution model

5. Linear Attribution Model.

This is less wrong.

That's it. Just less wrong. Use it if you are shooting for that.

When my son was smaller he would go to competitions (sports or IQ) and everyone would get a participation certificate.

Life, it turns out, is not utopian. When there is a competition, someone gets a gold medal, someone gets a silver, and someone gets a bronze. Everyone else goes home a loser, motivated to work harder the next time and win.

You should not treat your marketing optimization program with the same level of outcome optimization that is applied to five-year-olds. You can, and should, do better.

If someone threatens your life, use this model. Give everyone who contributed a participation certificate. But if you are not in a life-threatening situation, other models might help you actually understand which channels are contributing more value and which are not. And two of those models are just one click away.

time decay attribution model

6. Time Decay Attribution Model.

Ohh …. much better!

The core premise of the time decay model is this: The media touch point closest to conversion gets most of the credit, and the touch point prior to that will get less credit based on a smart and simple algorithm.

You only have to think about it for five seconds to realize it passes the ultimate test for everything: Common sense.

We could argue about how much credit the last few should get and how much the rest and how much the first. (Or we could not.) But overall it does seem to make sense that the further back a media touch point is (Organic Search and Social in our example) the less credit it should get. After all, if the touch points were magnificent, why did they not convert?

time decay attribution model adjust half life

One of the cool things about this model is that you can customize the half-life of decay and insert your own feelings into the attribution process. Notice I said feelings. :)

If you are going to start doing attribution modeling, the time decay model is a great, passes the common sense test, way to dip your toes. Go to the Model Comparison Tool, click on Select Model, choose Time Decay, and let thoughts be provoked!

Bonus: Adjust days prior to conversion on top of the tool based on your Time Lag report in the Multi-Channel Funnels folder.

position based attribution model

7. Position Based Attribution Model.

In some ways I really like the position based model because I have opinions – sorry, I meant to say expertise :) – and it is so easy to insert those opinions into this model and do some cool stuff.

That is what makes it a dangerous first model to use. If you don't know what you are doing, it is GIGO very quickly.

By default, the Position Based model attributes 40% of the credit to the first and the last interaction and the remaining 20% is distributed evenly to all the interactions in the middle.

1. See my perspective on first click attribution model above. 2. Understand why I believe that as designed the default position based model is sub-optimal. 3. Promise me you won't ever use the default one. 4. Feel really great you dodged a bullet.

Of the six attribution models available, there is one that you can use with little thought and still get value (Time Decay). One is not great, but won't completely kill you (Position). Three are so weak that you should not acknowledge them if they pass you in the street (and actively warn your friends to avoid them!).

Why are there so many models? The known world is smaller than the unknown world. There are always corner cases, there are always weird scenarios, there is always someone who wants to do something odd. All these reasons are good reasons for all these models to exist. But do go into using any model with open eyes.

There is one more thing you can do after you are done with the first step, playing with and experimenting with the results of the Time Decay model. You can create a customized attribution model.

custom attribution model

8. Customized/Personalized Attribution Model.

(I've said this twice already but let me say it again, don't go into this until you play with the Time Decay model and have spent a good few weeks learning the implications and trying to take some action. It is a very good learning experience.)

I love using the customized attribution model, and I'm grateful that the team at Google made it free for everyone rather than having it only for Google Analytics Premium.

With the custom modeling tool you can use the Linear, First, Last, Time Decay and Position Based models as your starting point, and then layer in other factors you consider to be important for your business to create your own attribution model.

I spend a lot of time with the business leaders, marketers, understanding historical performance, current media-mix and spend patterns before I create a customized model for them. Among the questions I ask the leaders are:

+ What type of user behavior do you value?

+ Is there an optimal conversion window you are solving for?

+ What does the repeat purchase behavior look like historically?

+ Are there any micro-conversions defined with engagement type goals, tied to the economic value?

+ Are offline conversions being sent back into GA using Universal Analytics?

So on, and so forth. These provide important context in making the decisions that will go into a custom attribution model.

From my portfolio of custom models, let me share one that has often served as a starting point for many customers.

Setting aside all humility for a nanosecond, I call it the Market Motive Mindblowing Model!

Click on Select Model in the Model Comparison Tool. At the bottom of the drop-down you'll see Create new custom model, click it.

Step 1: Select the baseline model.

I start with the Position Based. Then specify the amount of conversion credit based on the position. Here's what I use…

custom attribution model step one baseline

If you've read this post carefully to this point, this distribution of credit should not come as a surprise to you. From all my experimentation I've found that taking out the last channel (whichever one it is) causes a material impact on the conversion process, so it gets a "good amount of credit." The middle channels have an important role in driving people to the last interaction, they are recognized for that. The first interaction deserves some credit for the conversion, but not as much as the middle or last – for obvious reasons.

My distribution above is a good starting point. It is also really easy for you, as I often do myself, to experiment with different distributions, note the impact and optimize.

Step 2: Select the lookback window.

My process for picking the optimal time period to look for campaigns/interactions/media touch points to distribute credit over is to use the Time Lag report in the Multi-channel Funnels folder. It gives you the distribution of typical behavior.

My rule for picking the lookback window is to pick "close to the upper limit of the number of days to conversion, excluding the outliers, plus a bit more."

custom attribution model step two look back window

In this case it was a B2B client, long conversion cycle that lasted around 65 days, ignoring the outliers, so I picked 75. Just to be conservative.

Look at your own Time Lag report, come up with your own number. I'm a big believer in not going back to every single campaign, no matter how far back, and dragging it in to give it credit. If it was so awesome, it would have kicked off a conversion cycle for us that falls within the upper limits indicated in the Time Lag report.

The next two steps are critical. They are both really cool. But more than that, they help us wash away some of the sub-optimal decisions we might have made in the above two steps. Pay attention.

Step 3: Select the engagement based credit option.

We now go in and apply a rather clever rule to adjust credit for our campaign based on the behavior of the user that came to our site. This is particularly important for the touch points prior to last click.

custom attribution model step three user engagement

Time on Site is always a tricky computation. In all Web Analytics tools, unless you apply custom code, time on site is not computed for bounce visits or for the last page viewed in a visit.

Hence, I prefer to use Page Depth as a proxy for site engagement.

In this step we are telling GA to give more credit to campaigns that deliver users that have a higher engagement with the site. So if a user from campaign X see five pages during the visit on my automotive website and campaign Y sends a user that bounces, campaign X will get more credit.

Only seems fair. And now you can see how some of your credit distributions in step one will be auto-corrected based on the type of engagement campaigns deliver.

Step 4: Apply custom credit rules.

The last bit of mind-exploding fun. We are going to select some custom rules that apply uniquely to our company (remember the five business questions above?).

You can literally apply any custom rule you want. You can go in and say "for all bounced visits from rich media display campaigns give the campaign 2x the credit." You would not do that, but you can. You can do the reverse, "give every campaign with Bounced Visits zero times the credit of other interactions in the conversion path."

I take a simpler first step. I want to value my campaigns based on the interaction they deliver. If there is only an impression (people only see the ad), I value that a lot less than ads that get people to click on them.

To do that first I choose Interaction Type. Then I choose Click from the Exactly Matching drop down.

custom attribution model step four custom credit rules

Finally, I would like to have ads that get clicks to be extra rewarded and, in this case, get 1.4 times the credit of other campaigns in the conversion paths (in comparison to ads that just get impressions).

Why 1.4? After some experimentation, that was determined to be the optimal amount of value for this business (remember the custom model questions above?). There is no way out, you have to experiment.

That's our last step.

Other ideas for this last step include the ability to give generic or brand keywords more or less credit. Or giving Direct or Social more or less credit. Or giving all Social visits that are the last click prior to conversion only half the credit compared to other interactions in the path (Include Position in Path Exactly Matching Last and Include Source Exactly Matching Social, where Social is your campaign tracking parameter).

Totally your call. Just remember to drag your common sense along when you sit down to do this.

[sidebar] Once again in step four you see how clever use of custom filters can auto-correct some of your earlier assumptions related to distributions of credit in step one. If campaigns in the middle, or the first one, don't have the optimal interaction they will automatically be penalized. [/sidebar]

Here's a complete view of the Market Motive Mindblowing Attribution Model ….

market motive custom attribution model

That is all it takes, four simple steps, a pinch of understanding your business and a sprinkling of common sense.

It should be completely obvious to you that this model is based on a specific client's business environment, my experience, and business priorities. While I believe it will serve as a good starting point for your very own custom attribution model, it might not be optimal for you.

Hence, more than anything else, I would love for you to follow the thought process and the reasons for making choice x or choice y. Then apply that level of critical thinking as you go about creating a model for your digital business.

Multi-Channel Attribution Analysis.

Once you have your models sorted out, I recommend you get rid of the last click attribution model. It only ends up being a heavy useless anchor on your analysis. If you want to do comparative analysis, choose Time Decay for the first one (we know it is better than last click) and choose the Mindblowing Model (or your custom model).

Your view will look something like this.

multi channel attribution analysis

Focus on that last column, % change in Conversions.

Use the guidance provided (essentially a positive or negative shift away from the reference model, in this case Time Decay) to make recommendations for a different allocation of funds/effort for each marketing channel. Comparing the two models, you can see where your previous model/belief was wrong. Try adjusting your budgets accordingly for better success. As an example, in the above analysis Referrals are performing much better than we would otherwise have credited them for.

For the most optimal outcome for your company follow this 3-step process:

1. Create a hypothesis based on above analysis for how to better allocate budget across marketing channels.

2. Test that hypothesis using a percent of your budget and measure results.

3. Be less wrong over time.

Multi-channel attribution modeling and analysis is not a one-time effort, it is something you'll do all the time. Not every day, but at least do an operational review every two weeks and a strategic review (with recommendation for changes) every month.

In Closing, Five Quick Tips/Reality Checks.

I want to leave with some insights from the front lines of solving the MCA-ADC, MCA-AMS, MCA-O2S challenges. Hopefully these will help you get a jump-start in your own efforts.

#1. For multi-channel attribution modeling to work, all your marketing campaigns (Search, Social, Email, Display, Affiliate, others) must be 100% tagged with campaign tracking parameters . Tag your Bing campaigns. Tag your Email campaigns. Tag your Social campaigns. Tag the campaigns your mom is running on leaflets handed out to neighbors.

#2. One of my favorite exercises is to do the above analysis based on Cost Per Acquisition, rather than just conversions. You may be getting a lot of conversions, but the CPA can kill you. Notice above I only have two CPA values. For the rest I need to upload cost data into GA for my Social, Referral, Organic Search (yes, it costs money), and Email campaigns. You do too.

#3.You don't have to do attribution analysis for all your conversions in aggregate. On top of the attribution Model Comparison Tool, you'll see a drop down under the word Conversion. Click. Choose any conversion you consider to be important. You can do attribution modeling uniquely and optimize your marketing efforts just for an ecommerce transaction. Or you can do it for email subscription signups, or downloads, or videos played or anything else you consider to be important.

#4. Remember all of the above just covers Multi-Channel Analysis-All Digital Channels (MCA-ADC). There are two other, even more complex, attribution analysis scenarios: MCA-O2S and MCA-AMS. You can learn more about them here: Three Types of Multi-Channel Attribution Problems.

Don't be disheartened that all this complexity exists. Take things one step at a time. Standard Time Decay model first. Then your own Mindblowing Custom Model. Then Experimentation. Then MCA-O2S. Then MCA-AMS (it is so ironic this is harder than O2S!). With every step, you are making your company smarter. Less wrong every day.

#5. If you spend more than $10 million on advertising/marketing, it might be well worth it for you to completely skip all the attribution analysis challenges and jump to media-mix modeling by leveraging controlled experiments.

Optimize for your online media-mix at the start, then move to optimizing your online and offline media-mix. Media-mix modeling is harder and more time-consuming (hence the $10 million bar), but the payoff is huge and can be a competitive advantage.

We are done! Attribution modeling mastered! Hurray!!

: )

As always, it's your turn now.

Are you doing any attribution modeling at the moment? What frustrates you about it? What benefits have come from your credit re-allocation efforts? Run into any organizational/ego problems with senior leaders yet? Love First Click or Linear attribution, what am I missing in my thinking? Which model is your BFF? What are two fatally flawed choices in my Mindblowing Model? What would you do differently? Has it been easy to go from analysis, end of this post, to insights to action?

Please share your feedback, critique, brilliant new ideas and radical proposals via comments.

Thank you.

Comments

  1. 1
    Tyson says:

    Great post, Avinash! Thanks for giving your individual thoughts on each model.

    What do you think about downgrading interactions with brand keyword searches?

    This seems to be one our clients like, because many of them focus their marketing on driving new customers and conversions.

    It's nice to be able to pick a model like Linear or Time Decay and then customize it with a 50% downgrade for brand keyword search visits. Of course with (not provided) this is getting less effective.

    Thoughts??

    • 2

      Tyson: Does the company not want existing customers to buy more, or find the company when existing customers are looking for them?

      And how do they know for a fact that a brand search is by someone who is an existing customer? Have they validated conversions via brand searches and found that 100% of the orders were from existing customers?

      I suspect the answer to both is no.

      I would run an experiment. Take some campaigns, don't have any brand keywords in the portfolio (be it just a keyword portfolio, or a search, email, social, display etc portfolio) and see what happens to the conversions. Then compare that to one with brand keywords. Now you know the value. :)

      -Avinash.
      PS: In case I was not obvious, discounting brand keywords in the absence of the above experiments is flawed.

    • 3
      Daniel says:

      It is not only getting less effective, but may also make your data (more) inaccurate, especially if your traffic volumes are not that high

    • 4

      Great post, and now I am really glad I didn't try to write my own first :)

      Seriously this is a great description of how to think about attribution models and reporting but, and it is a big but, really what is the policy that this analysis is going to drive?

      Ostensibly it is about budget allocation, but since we followed a fixed policy when collecting this data, we really have no information around how our KPIs might increase/decrease as we reallocate based on any one of the distributions suggested by the various attribution models. I know you can say it is a starting point, but, and I am not trying to be pedantic, I am really not so sure. To me by advocating following this formalism, we might be implying that this analysis answers something that it really doesn't.

      The attribution model as is, tells us nothing about how it will respond to change – which is exactly what we are looking for.
      In order for this stuff to work, at each point in the process were we might try to affect behavior, we need to embed at least a null option, so that we are working with a decision process, rather than just a chain, no?

      • 5

        Matt: Break down the problem into two pieces. 1. Suck less. 2. Really rock.

        For #1, attribution modeling incrementally provides better answers, it helps you answer "if everyone who pitched in got some credit, how would the conversions look like?" (And being pedantic, I don't think of this as analysis, I just see it as modeling. Subtle but important difference. :)).

        For #2, you can and should have a practice inside your company to do controlled experimentation. Then you can have your null option, delight will follow! :)

        -Avinash.

        • 6

          Hey Avinash,

          Quick question – The selection of the number of days for the duration of the lookback makes intuitive sense. However, for the position based ratio selection – how would one evaluate whether a 10-50-40 is more relevant than say, a 40-40-20 model?

          In the absence of a control group, what is the best way to arrive at the 'right' weights?

          Regards,
          Neel

          • 7

            Neelakash: There is no blessed distribution, the only way to arrive at it is to use experimentation. Take credit distribution A, B, C (A can be control) and run a test to measure results. Go from there.

            As mentioned in the article, you don't have to wait for a perfect answer to start using attribution modeling. Anything you do will be better than last-click. The question then is how much better can you be. :)

            -Avinash.

  2. 8
    Charles says:

    Dude, That was hands down the most informative article I have read this year!

    Thanks for writing it!

    Keep up the great work.

  3. 9
    Josh Braaten says:

    Avinash – I feel like web analytics was complex, and it's just gone to another level altogether with new features such as attribution, cost data, and the entire Universal Analytics universe. There's no better time to focus on these tools and strategies — before the learning curve is simply too steep!

    I have a sad attribution story to tell you. An affiliate partner of mine has traditionally focused on a last-touch model for paying out their sales. I dedicated my entire content strategy to focus on the advocacy of their product (as did many of their other affiliates), creating a rich community surrounding the brand in the process.

    Recently they shifted to paying out on a first-touch attribution model because they wanted awareness, not advocacy. Now, part of me can't blame them because they really do have a great product and, to a certain degree, they don't need me to tell people how great it is.

    But on the other hand, this shift in how they compensate their affiliate partners has seen the advocacy for their product dry up nearly overnight from these hubs in the community. Their shift to partners who create awareness has fundamentally changed the dynamic of the brand, which is just one more thing to consider when looking at attribution.

    I will lean particularly heavily on this post for my future attribution activities. Thanks much for such specific thoughts and screenshots on this advanced topic.

    • 10

      Josh: What a great story. Sorry, I mean what a bad story for you that illustrates the challenge so clearly for everyone! :)

      When people choose first click, as your client, what they are saying is:

      "I don't care about the outcome for my business, all I care about is that people know about my product. If then they decide not to buy it, if all of that knowing was a complete waste of time, that is absolutely ok. And PS: Don't tell me if it was a waste of time. PPS: LA, LA, LA, LA, I can't hear you, LA, LA, LAAAA!"

      :)

      It sounds astoundingly crazy. But that is exactly what they are deciding to do.

      There is no way to know if First Click could have delivered the conversion all by itself. If it could have, why do you need all of the other advertising/marketing?

      Life is not about a OR, it is about an AND.

      Last click OR first click is silly.

      First Click AND "what else caused the conversion" is the right approach.

      Then we can, as we do in the post, discuss how to distribute credit for subsequent touch points, we can discuss how to overweight clicks (as I do) or overweight impressions (which I hate, sorry), we can discuss duration etc. All debatable to identify insights to optimize the portfolio of marketing that lead to a success.

      But you MUST optimize for the portfolio and not a silo. If you do the latter, you are cutting off your legs to try and run faster. Ok, maybe I'm being a drama queen there, but you get my point.

      -Avinash.
      PS: For true Analysts it is easy to figure out the impact of a corrosive decision like first click attribution. Just run a controlled experiment. Cell A: All budget put into first click. Cell B: Budget allocated according the advice above. May the best Cell win!

      • 11
        Dirk Brauner says:

        Avinash: First of all I want to tell you that I found this Article very informative.

        In regards to Josh´s comment I have to say that I understand why the Affiliate Partner might act this way.

        In the example of an Affiliate Partner, like the one above, we might be talking about a voucher affiliate. Voucher affiliates, who provide discount vouchers for various online shops, basically feed off the shops Brand keywords and are often visited once the customer has already reached the basket. The customer sees that there is an option to enter a voucher code and starts googling.

        This is why I would imagine the Affiliate Partner might want to decide to compensate Affiliates (or just some Affiliates) for reach (First touch). Otherwise we might be giving credit to a channel that doesn´t necessarily drive conversion but merely steals margin at the end of the funnel.

        From my perspective an attribution model based on channel mix and position would be justified. Have you tried something like that?

        Thanks for the great article!

        Dirk

        • 12

          Dirk: The way attribution modeling works in Google Analytics, or the other lovely tools, is that in the scenario you describe the Affiliate Partner will appear as the media touch point just prior to conversion, and not the first one.

          The first one will be the channel that drove the person to the site (search, social, email, display, etc.). Then during checkout the person saw the coupon/discount code box and went off to Google/Bing.

          This is the reason for my original guidance. You simply can't guess or "stack the system in favor of x or y." Just create the cleanest set of common-sense rules and then let the chips land where they might.

          You are ABSOLUTELY right that it is a margin stealing outcome because 1. we have to now give a discount to a customer and 2. we have to give a bounty to the affiliate for "delivering the conversion" and 3. for the customers we get in this manner (price sensitive, discount seeker) repeat purchases might be non-existent.
          Avinash.

  4. 13
    Dave Mundo says:

    Avinash, I love your perspective on digital attribution. I do have a couple of nits to pick, though.

    You indicated Time Decay is must better than Last Click…not necessarily of course. For a particular dataset/client, using Time Decay may actually optimize away from your most efficient spend. For instance, what if the "true" weight of the last click should be 85% instead of 100%? Using Time Decay would actually hurt you; Last Click would be the better option in this case.

    Also, I think your "Mindblowing Model" is just as guilty as relying on opinions/biases as Last Click, First Click, etc. Why choose the 10%/50%/40% distribution? For this particular dataset/client maybe the "truth" is 25%/40%/35%? Or maybe it's 5%/10%/85%? You see my point…deciding in advance what your distribution is allows you arrive at whatever conclusion you wish. In your example, if I decide up front I want referrals to perform well, I can use your distribution. If I want them to perform poorly, I could choose a different distribution.

    For this reason, I'm a strong believer in custom, data-driven attribution (statistical) models that let the data decide the optimial weights. You may argue that most digital advertisers don't need something this complex, though I'd argue 1) they do :) and 2) it's not really complex if you partner with the right experts. Otherwise, digital marketers will continue relying completely on their biases and will continue to spend dollars inefficiently.

    We recently created a nifty infographic on this type of modeling, please check it out!
    http://www.bkv.com/blog/scientific-attribution-modeling-grow-sales-20-infographic/

    Love your blog my friend!

    – Dave

    • 14

      Dave: Thank you for the nits you have picked. Debate pushes critical thinking, I so appreciate that.

      For the first part, perhaps I can restate the point. Time Decay causes less harm in all scenarios than Last Click. Simply because with the latter you ignore 100% of the prior ad/marketing touch points, with the former you at least give them some value. It might not be the right value, but some value.

      Once you give some value, we can ask the questions you correctly do in your comment and we can get better.

      On your second part… OMG totally! I do hope you did a Ctrl + F and saw how many times I used the word bias. :)

      Think of my custom model as *my* starting point from *my* 150 experiences of doing this. Those experiences are unique and most definitely bias my view. But I hope it serves as a starting point for you, if you don't have one.

      Ultimately the proof of the pudding is once you take the learnings, experiment with the new budget and realize how awful your assumptions were or how awesome they were.

      On the third part… I ever so passionately disagree. As I mentioned even Google Analytics has a data-driven attribution model. It is pretty nice. But it would be silly to suggest that it is not biased and does not suffer from missing variables, missing data points, and remember humans code these models (is there any statistical model that does not include, to a greater or less extent, the bias of the creator?).

      That does not make data driven models any less valuable, or in some cases better than the standard ones or biased custom ones. By default with people who have data driven models, it becomes my reference model (you'll see I use Time Decay above as a reference). But math and code do not make unbiased models. (Do not share this with anyone working on google.com!)

      Ultimately the only true way to wash out biases (mine, yours, smart algorithms') is to use controlled experiments and approaches like media mix modeling. It is not accessible to all companies (hence my $10 million bar), but when we can do that the ROI is magnificent.

      -Avinash.

    • 15
      David Jaeger says:

      Hi Avinash,

      I love your post. I've been deep diving into attribution modeling using the GA tool, and I feel this is absolutely huge for the industry. Most of my clients can't afford to pay the mega-bucks otherwise required, and I was actually custom coding some log based analysis to get some insight for some clients with convoluted purchase cycles.

      I am concerned with a similar viewpoint to Dave's above – how do we get rid of our own biases and systematically build a more accurate model. Ultimately, my preference for a model vs. your preference for a model is inefficient. We'd like to understand how our customer functions, and what monetary impact each channel / visit / ad impression has had on each sale.

      I love the fact that you've given a default custom model. I'm going to add in a deduction for any remarketing and brand campaigns, as those by default had a different channel that interacted with them first. (e.g. .9x of the regular). I would also channel group them separately.

      • 16

        David: Experimentation. That is the only way to validate our hypotheses (biases!) and learn where we are making the wrong assumptions.

        Please also checkout the comments (above and below). There are so many wonderful comments (I've also replied to many and added more context).

        Thanks!

        Avinash.

        • 17
          David Jaeger says:

          Hi Avinash,

          I've really enjoyed the post and comments. As far as testing, I've been thinking that the best way to test a hypothesis would be to segment (say) all of our US traffic in half (e.g. take all the states, look at current performance (e.g. conversion rates and AOV for an ecommerce site), and make sure I have 2 equal segments.

          I could then optimize all of my campaigns on one segment with one attribution model in Google analytics, and I could optimize the second segment with my original model. At the end of the testing period, I could then look at overall conversion rates. This could also give me insight into how well solving MCA-ADC correlates with MCA-OMS.

          While this is a pain, it's quite doable.

          The challenge I face with attribution modeling tool is getting insight into where I am off-base on the attribution model. Sure, we can make assumptions, but ultimately, I'd like to test those assumptions.

          I know you have an article about the process of conducting a real test somewhere on the blog, but I can't seem to find it.

    • 18
      Effie says:

      Dave,

      Thanks for your reply and for the neat info graphic. I do support your take on a data driven approach-coming from a statistical background myself. I have one question:

      1. The infographic seems like a variation of a Media Mix Model. Data issues aside, one of the limitations of a Media Mix Model (in my opinion, at least) is that it can not go deep enough into the specifics of advertising elements. MMM capture the essence of drivers into the response variables. However, when you consider a channel like Search with varying geotargets, campaigns, keywords, etc., the combinations of these factors can run into thousands. I am curious on the trade-offs between data granularity and model robustness when you are building the attribution model.

      Regardless, I think the data-driven approach is a step in the right direction. Thank you for sharing.

      • 19

        Effie: I'm sorry I'm not sure which image you were referring to. But in context of media mix modeling (and I'm a big fan), you are right that initially when you start off you are simply trying to get the mix in your overall big buckets right. For example, how do we distribute our budget between Paid Search, Display, and Paid Social. Or between TV, Radio, Search, YouTube.

        But that is not the end of the road. You can use media mix modeling to tease out the distribution between Brand Keywords, Category Keywords, YouTube Search. Etc. Etc. We have to find the balance between how much fiscal sense it makes.

        It is important to remember that we lean into using techniques like MMM because other techniques don't get us the answer. In as much there are always times when we know what to do, but not exactly why we must do it. All of course based on data.

        Avinash.

      • 20
        Dave Mundo says:

        Effie, thanks for the note. At a high level, yes, you could say digital attribution modeling is similar to marketing mix/media mix modeling in the sense that you have marketing phenomenon as indepedent variables and a response metric as dependent variables. The input data is a little different though: digital attribution typically includes 0/1 variables (did someone view an ad or not) and the dependent variable is also 0/1 (did someone make a purchase or not) — think logistical regression. Mix modeling generally is run on aggregate data (sales at a DMA level as a function of marketing activity in that same DMA), and the depedent variable is continuous. From a statistical modeling perspective, the difference in type (0/1 vs. continuous) results in using a different modeling approach (logistic vs. linear/non-linear regression).

        Also, I agree with your assessment of mix modeling…it can have its limitations when the goal is to drill down within a specific channel. Having said that, if you obtain the right data, you absolutely can assess the impact of these channel-specific attributes within mix models. For example, you can choose to measure the impact of TV in one variable, or you can create numerous variables for TV based on creative, daypart, etc. The same goes for digital…as long as you have the correct data. Honestly, I think mix modeling should be done in conjunction with a more detailed channel-specific analyis to get at these questions: execute mix modeling to decide how much spending to allocate to digital, then use attribution modeling to determine how the budget is allocated within digital (search vs. display vs. video etc.).

        Hope that helps!

        – Dave

  5. 21

    Avinash, please consider the "wow! great article, amazing insight, must read, super awesome" part of my comment covered!

    My dream – if at all possible, is all analysts would actually take the time to read and understand this critical aspect of business optimization (note that I intentionally wrote "business", not "web" or "online"). All that's missing is an exam at the end! :)

  6. 22
    Eric Tsai says:

    Great insights as always Avinash!

    A great primer to get started on attribution modeling. Especially with Google Analytic's new feature to track GDN impressions, it's up to us to upload cost data and start playing with the different models!

    If you want to know the tactical implementation of using GA Attribution modeling tool, you need to not just tag your campaigns but also leverage custom channel grouping. I wrote about how to do that here if you're interested :)

    http://www.designdamage.com/web-analytics-strategy-how-to-use-google-analytics-to-get-actionable-insights/

  7. 23
    Terry Hayden says:

    Avinash,

    This is a great article – as are they all! I love your humor and straight forward way of presenting the information. I look forward to each and every post – they are so enlightening.

    Thanks for all the hard work!

  8. 24
    Keith Campbell says:

    As always, great article – thanks for sharing so much with the community!

    Now, if only Adobe SiteCatalyst would get on board and offer similar capabilities in their base product. Until that time, if anyone knows of any work arounds for getting similar data in SiteCatalyst, let me know.

  9. 25

    Avinash –

    A gem of a post. Truly.

    I'm very glad you mentioned that this post is a deep dive into MCA-ADC and *not* MCA-AMS. *Question* –> to what extent does the MCA-AMS + cookie deletion world we live in (incognito, anyone?) make the insights we arrive at by applying a model, even a super-sweet MMMModel, just unreliable. My guess is that your answer will be -> we're "less wrong." Curious if that's what you think and if there is any way for us to quantify the quandary that MCA-AMS causes.

    Re: Problem with de-valuing Direct traffic in Last Non-Direct Click.. I hear what you're saying but don't wholeheartedly agree. I'm sure that you too have seen *plenty* of last touch direct traffic simply caused by session timeout. Heck, all the other sites I was on also just GA timed out while I read this post and commented! Personally, I prefer Last Non-Direct Click to Last Touch… (please don't throw rocks). Luckily, the session timeout issue is "somewhat" quantifiable (at least for ecommerce sites) by looking at how many macro conversions have a shopping cart or 1st page of checkout landing page. That data can then be applied to a custom model.

    Lastly, hurray for laying it on thick with regards to custom models. It is a good message to hammer home.

    Yehoshua

    • 26

      Yehoshua: I have to share a philosophy first (I think we both believe in it). It is always possible to find problems in your data, it is always possible to chase the last bunny down that very last, very deep fruitless rabbit hole. One of my biases is that I stop at the point when I reach diminishing margins of return. Because otherwise I could die data puking and never add any value to any business. This is a gigantic bias I have. I wanted to be transparent.

      Ok, on to your kind questions…

      First party cookie deletion does not have a material impact for most clients in the time-lag to conversion. Most of those fall within the 20-40 day window. The longer you stretch time, the more material the impact.

      With MCA-AMS, cookies are valueless. Either you have an opt-in permission driven model to track a human (awesome, go crazy with attribution), or you don't (attribution dies a early death).

      Regarding session time-out, luckily it is a easy thing to validate. Look at the percentage of sessions where the start page for the same User is exactly the same as the last page of a prior session. If your data says that happens 50% of the time, or 20% of the time or whatever, use last-non-direct. I've never seen a material number from this analysis and until then I refuse to take on faith what the GA team wants me to believe!

      I know I have you with me on that journey. :)

      -Avinash.

      • 27

        Hi Avinash,

        Thanks for your reply. I see that perhaps I didn't express my question as clearly as I hoped to. Let me give it another shot.

        I lumped cookie deletion / browsing in "private mode" together with MCA-AMS because ultimately no cookies = no attribution without the opt-in permission model as you mentioned. Not exactly the same thing, but close enough for this issue.

        You mentioned your bias regarding stopping at a point where you find diminishing margins of return. That is (somewhat) what I was trying to get out. At what point have you found there to be a diminished margin of return doing MCA-ADC analysis because of an MCA-AMS reality? Do you have thoughts with regards to quantifying that? Do you feel that trying to quantify that is a waste of time?

        For example: stop segmenting Organic Traffic by branded / non-branded keywords because (not provided) is at 80%. It makes no sense to try to analyze data you don't have.

        My guess is that you believe that being able to quantify issues with data quality leads to better decisions. For example, you mentioned uploading cost data for Organic traffic because, yes, it costs money. If one were to ignore the fact that iOS6 doesn't (didn't) pass referrers, they'd have a skewed sense of the value their getting from their SEO efforts. However, if I can estimate the amount of iPhone traffic that is probably Organic based landing page yadda yadda… you get the point. I'm speaking about using "broad strokes" here; I too agree that chasing small rabbits around all day would be tiring and fruitless.

        Alternatively, you may say that all exercises that try to quantify data quality issues are indeed just chasing rabbits.

        Back to your reply: MCA-AMS -> either you have opt-in (yay!) or you don't (dead!). Great. But it's not as if sites are either in the AMS camp or the ADC camp. There's a cross-over for sure. Does it matter?

  10. 28
    Barbara Frontera says:

    Thanks for this. As always, you've identified problems and provided solutions (or at least where to start). I will point clients to this post in the future when I mention MCA and they stare blankly into the distance.

  11. 29

    Avinash – It's great to see someone of your intellect and stature shedding some light in this area of analysis. I see a lot of confusion on this topic, from both newbies and "experts" alike, but not a lot of useful discourse.

    Now, the next logical step is helping people understand #2 from your 3-step process at the end:

    "2. Test that hypothesis using a percent of your budget and measure results."

    Expanding on that one line could fill several blog posts itself, I'm sure! (Actually, looks like you've already written about that and I forgot – http://www.kaushik.net/avinash/controlled-experiments-measuring-incrementality/).

    As always, many thanks for your selfless contribution to the education of us all!

  12. 30
    Lukas Stuber says:

    Very entertaining read, and valid points, no doubt – as long as the question is: "On the whole, how should I spend my bucks?" But other questions need other attribution models, I think.

    When doing a product launch, for example, the "First Interaction" model proved to be very helpful in optimising launch strategies, since it answers a very different – and in this case crucial – question: "How do I get this indispensable first contact?"

    Ultimately, attribution modeling is quite possibly not about finding The Truth, but about finding the best set of models to answer different questions, I guess.

    • 31

      Lukas: You are absolutely right, with attribution modeling we are indeed trying to answer the "I have $xx mil in my marketing budget, how should I best distribute my budget amongst the marketing channels to maximize profit for my company."

      Regarding first click, please see my discussion with Josh for some food for thought.

      I don't think you can find the truth with attribution modeling (you can get closer to the truth with media mix modeling), it is about figuring out how to be less wrong. :)

      -Avinash.

      • 32
        Lukas Stuber says:

        Avinash: Well, yes, as I said: As long as it's about profit, I completely agree with you. But as soon as other business goals come into play – purely growth-related ones, e.g. – other models have their merits, in my experience. Even First Click models.

  13. 33
    Dave Rekuc says:

    Avinash, thought provoking article as usual.

    One thing that I also feel compelled to mention is this kind of modeling is simply a measurement of the past, not a prediction of the future. So many marketers confuse the two and run themselves in circles because of it. A new dollar invested in a marketing channel will not necessarily perform the same as the existing dollars in that channel. In some instances, its more predictable (increasing bids in paid search) and in some cases its less so (viral campaigns, press releases, new keyword targets, etc). I understand you can only address so much in an article, but this, to me is incredibly important.

    The thing I struggle with the most is evaluating the models and advancing my modeling based on accuracy rather than subjectivity. If I identify an area to shift budget to, then do so and receive sub-par results. Was my attribution method the flaw or my execution of the new campaign? Or, perhaps, I receive stunning results, was it a false positive?

    The other cautionary piece of advice I would give is never expect an attribution model, no matter how complex, to replace the need for a human being to interpret the results; expect the model to augment that person's decision rather than replace it.

    Thanks for the good read :)

    • 34

      Dave: I could not agree with you more emphatically. That is why I had the couple lines of guidance just before I started my closing thoughts.

      Identifying all the variables in play is very important. If the model goes from 90% Search, 10% email to 40% Search, 20% Email, 40% Display, it is very important to be aware of all the new variables in play so that you can make a smarter decision.

      This is partly the reason that we all start with attribution modeling, but quickly end up with controlled experiments because it allows us to identify some causality.

      Your comment goes to the heart of my 10/90 rule, and hence a 100 hugs from me to you! :)

      Avinash.

  14. 35
    Michael R Hoffman says:

    Nice – And while complex multi-channel by customer type and segment (informing the channels and better – the interaction point) about the customer/stakeholder changes the assumptions and formula dramatically and then there is the multi-screen…

    I built a framework to predict each customer's movement/lack of mevement through an exhaustive contact matrix (CxC matrix: customerworthy.com/cxc-matrix ) following customer path (horizontal) by company/affiliates channels (vertical) which is unbelievably easy to populate – just count the contacts per channel to start – and monetize customers passing/not passing from stage to stage.

    This method connects all the pieces and links all the stakeholders in and around a company (agency, affiliates, distributors, call centers, even manufacturers after market)

    If you want to build the killer app – use the CxC Matrix to populate a complete simulation (time shift for testing and expose variables per interaction, environment, competitors, seasons – again, easier than it sounds – and visualization socializes experimentation and input for innovation )

    Please feel free to add this dimension to your analysis – I'll send you free pdf copy of book Customer Worthy (or $50 @ Amazon) and of course answer any questions about how you might use this – consider CxC matrix open source for customer experience management.

    Great post and catalyst for discussion.

    Michael R Hoffman, mrh @ customerworthy dot com

  15. 36
    Lisa says:

    Great post, and extremely helpful.

  16. 37
    Gretchen says:

    Hi Avinash, great post and something we've been (finally!) talking about a lot at my company.

    I see on the "custom credit rules," you can include interaction types = impression. My question is, how does GA know that an impression occurred if the user never visited our site (hence no page-view)? We advertise heavily via display media, both on the Google Display Network as well as other ad networks.

    If it is possible to track impressions via GA, it would solve a lot of problems for me. I've been pushing hard within my team for the last eight months to hire our own ad server, so we could have a 360° view of our marketing campaigns (impressions + clicks) plus the on-site activity.

    I'd love to hear your opinion on this matter.

  17. 39

    Besides the complexity for the web analysts to understand en set-up the best model, I think the biggest challenge is to translate this into understandable conclusions/reports for other not-so-analytics-savvy stakeholders in your organisation. If you explain this to other people, you will get a lot of questions and you won't be able to answer them with a simple yes or no (what a lot of people expect of web analytics).

    Showing your new and better model, also means that all the things web analytics told us before about channel attribution (on last click), were 'wrong'… So also be prepared for this.

    In general this tool is a giant step forward in channel attribution, but I also think it requires the right approach of the web analyst towards other stakeholders.

  18. 40

    Avinash, nice post… but what about display media? What about view-through, the impact of "view but not click" behaviors?

    This post and model appears to be focused on the click, which for some brands is great, but for many others, is only a small portion of their experiences: they have social interaction with customers, they have sponsorships of digital experiences, they have display both bidded and guaranteed: where do all these view-based experiences fit into your approach and GA?

    I guess you call this "MCA-ADC" in other posts, but any thoughts on how to fit it in here?

    • 41

      Michael: There might be a misunderstanding. You'll see Display Advertising in a bunch of my screenshots in the post, for example the very last one. You will also see that when I apply custom rules to my mode, I choose click but the screenshot shows that you can use impressions if you want.

      I'm biased when it comes to View-Thrus, I do not believe in the party line coming out of most advertising platforms or from most Gurus. By default I give zero credit to View-Thrus because I'm being asked to believe on faith that View-Thrus by their mere existence are awesome.

      My approach with valuing impressions (with no clicks) is to use controlled experiments to validate that I'm able to attribute a specific lift to those non-clicked on impressions (View-Thrus). Sometimes that comes out looking great, other times the impressions were a complete waste of time.

      Data over faith. :)

      -Avinash.

  19. 42
    Cameron says:

    What a ridiculously helpful post! Will have to re-read everything when I get home.

    Can't wait to try this out. Have been fiddling with the tool myself, but had no idea where to start.

    Thanks a lot!

  20. 43

    The "missing information" piece seems quite significant. Cookie deletion, multi-device browsing/buying, etc. THEN you've got the offline stuff … direct mail, call center (inbound/outbound), and even mass media (TV, etc.). By siloing customer experience into digital only, I think many analysts WAY over estimate the impact of their digital channels.

    Digital attribution is great, and there's a lot to be learned from this excellent post, but there's a big missing piece. Discussions of attribution must be omni-channel to truly reflect the experience of customers, and accurately reflect marketing ROI.

    • 44

      Patrick: Just as digital analysts can't say that it is all digital, it is equally sub-optimal to say that digital is "way over estimated." Unless you measure it, you just don't know.

      I share your stress on true multi-channel attribution, hence the post opens with an explanation of MCA O2S, AMS and ADC. There is also a link included, pasted below as well, where folks can learn more about the value of, and the challenge of, optimizing for all channels. (And while we are on this point, you can't do it with attribution modeling.)

      -Avinash.
      http://www.kaushik.net/avinash/multi-channel-attribution-definitions-models/

  21. 45

    Hey Avinash,
    A great and very informative post indeed. It reminds me of my Maths Class on Volatility Modeling – EWMA and GARCH. :-) A lot of gobbledygook on hard core maths and modeling. ;-)

    We have been using attribution to the final stage of conversion and I do agree with you that it is not the best method. It is almost like giving sales incentive to somebody who closes the sales, without any incentive to the earlier marketing efforts! But quite often I find this to be true!

    Had a few questions:
    – Based on the importance of each media (lets say if I am using the exponential decay), how would I decide my budgets? Do you suggest that if Social Media gets an importance of 2 and PPC gets an importance of 1, then I appropriately adjust my budgets?

    – What is the best way to estimate decay factor? Just out of curiosity, are you guys also using MLE or other hard core maths to estimate? Or is google internally using it? I remember that in EWMA, again there were many models and JP Morgan had one that sold very well!

    – We are combining a lot of factors and metrics here. Would this not reduce the overall visibility of decision making (I remember your post on moving from cumulative metrics to focus on individual metrics!)

    Overall an eye opener that raises a lot of questions! :) Keep up the good work.

    • 46

      Paramdeep: Answers to your questions….

      When you use the attribution modeling tool, as we did in the latter part of this post, the last column shows how much credit would swing (positively or negatively) based on your new model. Use those swings to come up with new media mix and experiment. You can just go out and try the new mix, but testing is better.

      I don't know that there is one perfect answer for this. If you click on Edit next to Time Decay it will show you how GA is doing it, you can experiment from there.

      There might be some misunderstanding. You are not mixing any metrics here, not even one. You are giving credit to all media touch points that lead to a conversion. You metric is still Conversion, it is still CPA.

      -Avinash.

  22. 47
    Jeremy Gold says:

    All of these attribution models have the same fundamental flaw: they don't account for confounding factors. It's impossible to go back in time after the conversion event has occurred so as to see what would have happened had the user not been exposed to any adverting at all. Indeed, many of those users would have converted anyway. Thus, all of these models simply move the deck chairs around on the Titanic and don't get any closer to the real question: how do I quantify the level to which my media strategies are actually influencing users? The answer of course is to conduct control and test experiments and ignore most of the noise around attribution.

    I see that Avinash has already written a separate and excellent post on Incrementality: http://www.kaushik.net/avinash/controlled-experiments-measuring-incrementality/. Thus, I'm trying to understand the point of the above post since none of these attribution models truly solve anything.

    • 48

      Jeremy,

      I think the point of the article is that it's a way of systematically generating hypotheses regarding peoples' responses to your marketing channels, which can then be tested.

      I think your criticism applies to all web analytics data – none of it shows causality until you do controlled experiments. It's simply a way of developing a mental model of how website visitors act, using that model to generate hypotheses, which can then be tested by taking action and measuring the impact. Then you refine your mental model and start again. In this case, substitute attribution model for mental model.

    • 49
      Ken P says:

      The identified models and custom model are rule based attribution models.

      What you're looking for is causal inference. Causal systems will include visitors who did not convert to truly quantify the touch point.

    • 50

      Jeremy: My apologies for the late reply, crazy schedule.

      I appreciate your feedback, thank you. I have to admit that I do believe that attribution modeling is incrementally better than what we have today (last-click). In as much I do disagree with your observation of the more obvious futility of deck chairs and titanic.

      Even if you use a completely silly model like Linear or First Click, you are at least aware of how suboptimal Last Click is.

      Maybe this is not right, but I've come to realize the value of solving for the non-utopian world I think people should live in. And not just on this topic, but on others as well you'll see me coming down from my high horse to solve a intermediate "just suck less" step so that I can bring people (some day!) to my utopia. It is so hard. :)

      Avinash.

  23. 51
    Daneil Huss says:

    Given the suggestion by some data (seewhy.com/blog/2013/04/03/understanding-online-buyer-behavior-part1/) that 99% of users do not make a purchase on their first visit. Is it fair to weight a higher percentage of attribution to the first click?

    In other words, if users have a natural propensity to not convert on a first visit, is it fair to say that the first click has more value than maybe we give it credit for?

    • 52

      Daniel: Please see my reply to Josh on thoughts about valuing first-click.

      Beyond that…. remember what you are saying with overvaluing first-click…. your first girl friend deserves a lot, or all, of the credit for you marrying your wife. I know that is a folksy metaphor, but think about it. Even if your first girl friend was amazing, even if she is the one who introduced you to your current wife and even if she massively pushed you two to get married, would you give your first girlfriend all, or most, of the credit?

      -Avinash.

  24. 53
    Ken P says:

    Avinash can attribution be used to create lift in new customer acquisitions, aka the visitors who didn't convert and are not included in GA's attribution data. How can we use attribution to identify gaps in our media or is this primarily an ROI tool?

    When you refer to click path are you referring to internal pages before conversion of channel click paths?

    Thanks.

    Thanks.

    • 54

      Ken: I'm not sure I understand your question. Only people who deliver a macro or micro conversion are included in the attribution analysis tool. The goal is to learn from them, get more people to convert via, now, better more informed acquisition strategies.

      The picture you see included in the post is for the media touch point that drove each visit to the site.

      Avinash.

  25. 55
    Anthony Centeno says:

    Another great post from The Master! Really enjoyed the breakdowns and storytelling. Thank you again for feeding the hunger for more insight!

    I agree that experimentation is the only way to test what the optimal budget allocation should be for a business and you've clearly demonstrated that attribution modeling can help design those experiments.

    Thanks again,
    Anthony

  26. 56
    Andrew Strickland says:

    Hi Avinash, like the post, totally agree with the 3 step process (analysis must lead to tests otherwise it is a waste of time) and evaluating CPA (or ROAS).

    Regardless of whether last (last) or last non direct is better as the default it definitely did cause confusion when MCF first arrived as the audience got confused when the 'last touch' number was different (at least in attribution modelling you can also do last non direct if you want to that is!)…

    I am interested if anyone has leveraged the GA Premium and DFA integration to include DFA display campaign impressions within the attribution dataset, if so how much of an effort was this to configure?

    Thanks
    Andrew

    • 57
      Tom Kelly says:

      I also would like to know how this configuration worked out and what insights were gained from it. My initial thoughts would be to eliminate the challenges faced in MCA-O2S, MCA-AMS (offline and cross-device) with Universal Analytics by marrying all touch points to a single user ID.

      A potential issue would be the discrepancy between GA and DFA conversions due to the different pixel but if the conversions could be streamlined somehow I can see this being a great asset.

      Thanks,

      Tom

  27. 58
    George says:

    Avinash,

    I've only just discovered your blog and have bookmarked your site already.

    This was a great informative read.

    Many Thanks
    George

  28. 59
    Peter says:

    This went over my head early, but I've found some people here we will use- that are staying on top of what obviously is the brave new world.

    But to you marketologists, please remember this from a human to human perspective: mel shampain moves product

  29. 60
    Anna Macko says:

    Great post, however did Analytics change from this week to last week? I cannot find LINEAR for the life of me today.

    The design changed and so how do we find it now?

    • 61

      Anna: Analytics does have a nasty habit of changing from time to time, sad outcome of the speed at which GA is evolving.

      But it has not changed since last week. :) Go to Attribution Modeling Tool which is in the Conversions > Attribution section. Then click on Select Model. The fifth one is Linear.

      -Avinash.

  30. 62
    Joe Meier says:

    Avinash,

    I LOVE the GA Model Comparison Tool. It seriously reduced the complexity of the enormous spreadsheet I had been using for the task ever since the release of Multi-Channel Funnels. The biggest headache of my modeling has always been Direct Visits / Brand Search, which correctly or incorrectly, I treat similarly. My question is, how much do you trust that your Direct visits are really Direct visits (typed in or bookmarks)? In your "Make Love To Your Direct Traffic" post, you refer to "corner cases" where odd browser configurations, etc. make for not-really-Direct Direct traffic.

    Even after crediting Direct traffic with a 0-multiplier in the attribution model (highly debatable), I'm usually still left with 15%-20% of my conversions attributed to Direct (because it is the only source in the chain). Assuming I'm as diligent about tagging every last link to our website that I can get my hands on as you are (I make a valiant effort), would you interpret this as really that many people converting directly? As much as that thought makes a marketer happy, I've always been skeptical. How much of that is mis-attributed traffic? For example, I read that clicks to our http site from https sites show up as Direct. Facebook is often https … that alone could be large. Assuming that example is correct, how much other stuff like this is going on creating Direct traffic that really belongs to other sources? Are these really the corner cases or is Direct traffic mostly unidentified traffic with some real Direct traffic mixed in? The whole recent debacle with iOS 6 Google Searches showing up as Direct just adds to my concern, though at least the impact of that can still be guestimated and it only occurred in the last year – but the impact was huge. Thoughts?

    Thanks.
    -Joe

    • 63

      Joe: With or without direct traffic issues, it is imperative that you use campaign tracking properly on all your campaigns to ensure that your MCF reports present a realistic view of your marketing touch points.

      One good benefit of complete campaign tracking, as my post on direct visits outlines, your direct visits will also get cleaned up and only be direct visits.

      On your second point… there are always things beyond your control and you have to make do with the hand you are dealt. If you want to figure out how much to attribute to direct, you can easily create a clean segment like "all direct visits with not via mobile platforms" and you can eliminate one of your hypotheses related to the cause of direct visits.

      -Avinash.

    • 64
      Rachel Huang says:

      Hi Joe,

      I'm pretty new to online analytics (my previous experience is offline). I'm curious about the multichannel funnel model u are using. Could u please elaborate a bit that how u cut ally use it pls.

      Much appreciated!

      Cheers Rachel

      • 65
        Joe Meier says:

        Rachel,

        I've created a custom model using Google Analytics to separate all of my various traffic sources.

        There are lots of steps since for some sources I want to break it into chunks like exact match vs. broad match, some I'm content to view it by campaign, and some I'm content to roll the whole source up into one group. I also try to grab as many variations of our brand name as I can and call that Brand Search.

        After that, it is just a matter of playing with different attribution models as Avinash suggests.

        -Joe

  31. 66
    Erik says:

    Thanks for the post.

    I agree that testing should always be done first. Your post is very detailed and informative.

    Thanks again.

  32. 67
    Sean says:

    Great read! This is a very comprehensive guide to getting the most out of attribution modeling. Definitely going to play around with modeling today and share this with my team.

    Thanks!

  33. 68

    Running through your steps on Google Analytics and conversion rates!

    I have struggled a bit with attrition cycles. Will spend some time with the Model Comparison Tool to better identify trends. That I think will better help me understand what is working. Looks very cool.

    Looking at last click credit when each page through the process should get credit for conversion is a bit troubling. The Position based Model makes much more sense.

    Going to take a bit of refining to achieve better detailed information.

    This is a great post!

  34. 69
    Ramakrishnan says:

    Avinash

    What is half life of time decay? Where in GA, can you go and edit the half life of time decay? Pls. brief

    • 70

      Ramakrishnan: Here is a page that outlines all the standard attribution models in GA:

      Conversions: Attribution Modeling: https://support.google.com/analytics/answer/1662518?hl=en

      To edit the standard half life of 7 days, you go to the Attribution Modeling tool. Click on Select Model. Look for Time Decay. Click on the link next to it called Copy.

      Now you will see a box where you can change the standard model. Give it a name you'll remember and then apply it to your modeling tool report.

      Avinash.

  35. 71
    Jim says:

    Wow, I am reading this article for the third time and still can't wrap my head around it.

    I think I will have an interesting day tomorrow going through GA while reading your article once again.

  36. 72
    Bharath says:

    Nice Post!

  37. 73
    erik says:

    Wow, this is a very informative post on a topic that can be confusing and overwhelming. Thanks!

  38. 74
    Brent says:

    This is mindblowing! Way back I did keyword attribution modelling to optimize bidding on keywords based on impression and first click. I found this to be very usefull in casting a wider net for people discovering my market.

    Is there a way to dovetail multi-channel attribution to keyword attribution or should they just work as "layers".

    This is great stuff – Need the ebook!

  39. 75
    Jalpan says:

    Lot of buzz abt data driven attribution or algorithm based attribution..

    Not sure how does it differ from simple concept of incremental response rate? If my RR from path A>B is 3% and RR from A>B>C is 5%, then incremental credit from 2% will go to C provided it's not one off case.

    Does data driven attribution work only on this concept or any other complexity is also involved?

    • 76

      Jalpan: I encourage you to read the post again, especially the first part. It might help internalize the uniqueness of the problem we are trying to deal with when it comes to attribution.

      At the end of the post I do make a very strong case for methodologies like media mix modeling where we try to compute incrementality. But it is important to understand that we should not solve for path analysis. How many paths can possibly exist?

      For a client I just opened the data for, 98k conversions come from 33k paths. Even if mathematically you can look at all the paths and for each unique combination you can compute incrementality, how do you go about controlling the path a person can take? How do you execute your marketing plans?

      There are a number of other challenges, but that should give you some idea of why this might be unwise.

      Attribution modeling, as I mention at the start, is imperfect. But it is better than last click. (And controlled experimentation si better than, and more actionable, than attribution.)

      Thanks!

      Avinash.

  40. 77
    Neeharika says:

    Avinash- This was a great read.

    If there is a website that is just trying to create brand awareness with a specific call to action as the only conversion tool. Would last call attribution be the best approach or is the time decay approach where you can give weightage to all previous page views before conversion would be a good approach.

    I just wanted some direction from you.

    Thanks!

    • 78

      Neeharika: Let the data help guide your decision.

      Look at the Path Length report in the MCF folder in GA. If you see a good number in the rows that are longer than 1, you should use attribution modeling beyond last click.

      It is completely irrelevant that you have a site that is driving brand awareness or driving an ecommerce outcome. You want to know what marketing, all the marketing, that influenced the outcome.

      After that, please see this post as to which attribution model might be optimal for your analysis.

      Avinash.

      • 79
        Shankar says:

        Avinash :

        In financial services and specifically banking, the customer journey transends from digital to call center or even a branch for complex purchases ( Mortgage, Auto loan etc).. Have you come across any financial insitution which tracks the metrics of referral sale ( customer leverages digital as a channel for research and start the app and saves it but completes it in the assisted channels) and influence sale ( researches online but starts and completes the application in the assisted channels).. Any pointers in this regard would be helpful…

        Also given the proliferation of devices, what are the various means that you can track and assign a unique common identifier across digital devices ( tablet,smart phone, desktop etc) to dentify the prospect… This is also a problem in the FS space… My guess is that retail would have solved the same.. Any pointers will be helpful….

        As always all your posts are insightful..

        Thanks !

        Shankar

  41. 81
    Veena Ranjeeth says:

    Hi Avinash,

    Nice article, and great insights.

    What are your thoughts on looking at consumer and business segments differently for Marketing attribution modeling, Would it really help to look at attribution modelling for b2b.

    Here are my thoughts, Business segment of consumers are more influenced by personal relationship, sales/after sales experience, (and other related parameters) more than their experience online? Also their time to conversion, online path taken, impact of different digital media on business segment are so very different from consumer segment.

    Secondly, most of b2b customers come back to you for purchase at regular intervals (based on need) if their past experience was good and if their requirements are met. Hence, their used of paid search for instance wouldn't necessarily mean a success story of your investment in that media (for b2b) and moreover paid search could've been used as a navigational medium to land on your page.

    How do you demystify this to be able to do the right analytics & invest optimally to target b2b.

    Regards,
    Veena

    • 82

      Veena: Let's unpack the couple threads.

      Let's leave the repeat purchases aside, simply because if you have a decent CRM system you should be able to track that efficiently and also tie that behavior either to the first purchase or subsequent marketing touch points.

      The challenge is the first one because at that point you don't know who the person/company is and what they are looking at, doing.

      On this front, just thinking of digital attribution, there is little difference between B2B and B2C. Especially because remember that what your analytics tool, Google/Adobe/IBM, is tracking is cookie/person based (rather than company based). Consider the same suggestions in this post for more on what to do.

      Avinash.

  42. 83
    Juan M. Sanchez says:

    As Google Marketing Evangelist say that "Last AdWords Click Attribution Model" is profoundly value-deficient, it's a good signal to follow your advises.

  43. 84
    Thomas says:

    Hi Avinash,

    Nice article, great insights and all so simple explained. I love your style to explain complex topics and I´m a great fan.

    But I have problems in understanding your rule for picking the lookback window: "close to the upper limit of the number of days to conversion, excluding the outliers, plus a bit more."

    Because even in the Time Lag Report I have the opportunity to switch beetwen a 0-90 days view. So here my problem starts. When I move the time controller the shares will change. So how can I find the right setting for this report???

    In clear words…I have a broad product range with ragrd to the selling price. Low priced products as well as high priced products. Nearly 60% of Sales took place at the same day when I use the 30 days lookback window. Of course the share will descend when I raise the lookback window. All other days 1- ~ get around 1 %.

    So do you mean with upper limit my 60% so should I take about 7 days? The differences in the Assisted Conversions report are huge when I take 7 or 30 days.

    What would be your recommendation or what about the others on this site? How you are dealing with this problem.

    Thanks for your help.

    Regards,
    Thomas
    A great fan! :-)

    • 85

      Thomas: You can only get good enough on this stuff, you can't get to prefect. Part of this is the data we have. Part of it is that some desirable segmentation options might not be there (assuming they have a good impact). So on and so forth.

      Please be aware of this. It is also important to point out that the things I do at the bottom of the Mind Blowing Model make some of the above complexity or perfection less important in terms of impact.

      All that said…. I start with the average (around 75% of the purchases, to leave the outliers out there) Time Lag and use that for my initial models. After that, depending on the importance of the business, I use various custom conversion segments (on top of that report) to hone in on a better number.

      -Avinash.

  44. 86
    Magnus Friberg says:

    Hi Avinash,

    Just a question about Time Decay model. What's the math behind it?

    Thanks!

    • 87

      Magnus: Here is a simple way to check (all the models). Go the the attribution modeling tool. Click on Select Model (on top of the table). Next to the model you'll see a little clipboard thingy icon, press that. It allows you to customize the model, but it also shows you what the model is doing.

      If you do this for Time Decay, here is what you'll see: Half-life of 7 days. And "An assist click occurring this many days prior to conversion will receive 1/2 the credit of a click on the day of conversion."

      I hope this helps. And now of course you can customize this to whatever you what! :)

      Avinash.

      • 88
        Magnus Friberg says:

        Hi Avinash,

        Thanks for the tip! I still struggle to see how the value distribution would look like… So let's say we have only 1 conversion á $500 and using a Time Decay model with a 5 day half-life.

        1. How the value distributed if there are only 2 interactions during a 8 day period, the first occurring on the 8th day before and the 2nd interaction (obviously on the day of conversion)?
        2. "1/2 the credit of a click on the day of the conversion" – and how do I know what the credit is for that click on the day of the conversion?

        Many thanks for taking your time :)

        Kind regards,
        Magnus

  45. 89
    Brad Sockloff says:

    This is a great article and I am so glad to have found it.

    The only place my opinion differs from you is with regard to the importance of first touch. If you are promoting a new product/service, isn't getting the word out there and getting traffic to your website of paramount importance?

    I am not advocating for using first-in attribution, but rather for the value of position-based instead of time-decay.

    It "feels" right for the touchpoints that drive initial awareness and close the conversion to get the majority of the credit. (And you can control the look-back window to make sure credit is not given to a first touchpoint that is really miles away from the conversion)

    Thoughts?

    • 90

      Brad: For additional layer of my perspective on first-click please see my reply to Josh. I hope you'll look at first click with a different light. : )

      With regards to the amount of value… you are right, we can control it when we create our custom models. And it is not overrated to worry about it to some extent. For example in our Market Motive Mindblowing Model, above, and http://goo.gl/4PjiQh, you would play with the percent value in the top part, but it is the custom rules where the important bits are.

      If I put in 10% and you put in 20% I don't think it matters all that much because of what we are deciding to do in custom credit rules.

      -Avinash.

  46. 91
    alan king says:

    I don't have valuable anything to add but I just want to express how informative, jargon free, well written and absolutely brilliant this post is.

    Avinash I take my hat off to you.

  47. 92
    Gemma Holloway says:

    Epic post Avinash!

    Multi-Channel attribution is been something I have been trying to get my head around for a long time and I can finally feel it starting to sink it.

    Can't wait to get started on my own analysis now :)

    Thanks!

  48. 93
    shiv says:

    I have little bit confusion over Time lag And Path Length:

    For Example: Visitor A comes to our sit Today from Referal XYZ , and then he comes second day with same referral, Then Time lag would be 1 or 2 ?

    According to me It must be 2

    And what About Path Length, it would be 1 Or 2 ?

    I am little confused over it.

  49. 95

    Hi Avinash,

    Thanks so much for this post – it finally helped me get my head round Attribution Modelling!

    I've set up my personalised Attribution model based on your recommendations above and am so glad to finally be moving away from Last Click analysis.

    Brogan

  50. 96
    Bjorn says:

    Great article. Some thoughts, would love to hear your feedback:

    Is there any way to:

    1) Get the custom attribution into other parts of GA? OR
    2) Export conversion data with the "new model" as actual transactions including purchase ID, so I can sync up with my database data (look at LTV by channel, etc)?

    • 97

      Bjorn: At the moment the customized attribution modeling you are able to do only applies to the MCF section of GA.

      But you are able to create custom segmentation inside MCF, which enriches your ability to do more.

      With regards to export, lots of data is available via the free GA API. I would encourage you to work with a GACP to help you create what you need. You'll find a list here: http://www.bit.ly/gaac

      -Avinash.

      • 98
        Bjorn says:

        Hi Avinash,

        Thank you for the great feedback. I did spend some time on the custom segmentation, which is great. I basically grouped all non-paid interactions together to analyze the paid ones – very powerful.

        In our business, we do care quite a bit about the first interaction because a) it takes a lot of time from the first time a customer hears about us until they try us, and b) it turns out that almost 100% of the following interactions outside of retargeting are unpaid.

        However, being cautious about first click attribution based on your comments in the article, I was wondering if you know whether or not Google counts impressions on Google-run channels as a first interaction?

        If the interactions were clicks/visits only, I'd feel much better about attributing towards them. Google seems to be (perhaps deliberately) vague on the topic, as I suspect they would get a lot more credit than other channels if it was impression based, as most other channels won't have impression data in GA.

        • 99

          Bjorn: I appreciate the first interaction, but only as you might have noticed in the article (and the comments, see my reply to Josh) I do not assign an irrational value to it. In my custom model I do appreciate certain behaviors (clicks more than impressions) and if the first click delivers that, I overvalue it in my model.

          The distribution you see based on the click behavior. But as in my custom attribution model, you can value interactions for channels where Google has data.

          -Avinash.

  51. 100
    liberty says:

    Dear Avinash,

    I'm an university student of China,and luckily I read your book Web Analytics 2.0.I'm in great interest on the Attribution Model,but I have few questions for your help,hope you solve my problem asap:

    1.I don't have a Google Analytics account n even have no money to pay for the attribution model function. How can I do this model via Excel?(This is an urgent problem for your help!!!)

    2.As the path of user click is complicated, should i list all the possible path?In a word, i still not understand how to calculate the credit of channels in this model.

    3.In China, it may be difficulty to track all the metrics, especially in multi-channel case. About the offline advertisement, what metrics do you suggest to use?

    I'm new in web ananlytics, so I have a lot of stupid questions to ask,hope you understand :P
    Looking forward to your reply soon!

    • 101

      Liberty: Answers…

      1. You don't need a paid version to use Attribution Modeling, you can use the free one.

      I don't know of an alternative in Excel, it would be quite complicated to build all the functionality in there.

      2. The tool will automatically compute the path for you, you don't have to worry.

      3. For attribution type analysis the metric to focus on is Assisted Conversions and Last-Click Conversions. This gives you a complete view of success by each channel.

      As to how to pick very few metrics please see my first two answers in this post:

      ~ Dear Avinash: Your Digital Marketing + Analytics Challenges Answered

      Good luck!

      Avinash.

  52. 102
    David Lee says:

    Thanks for this post, very informative. I'm wondering that once you select any attribution model you're immediately "biasing" the data, no? Even when comparing one model versus another, I'm not sure I understand that CPA is actually higher or lower. Isn't it only that particular number because of that particular model?

    Intuitively, it seems like the best way to get the most appropriate attribution would be not based on a model but based on the data. I know its not possible to connect all the dots since this is such a complex situation with so many variables but there's got to be some way to give credit to the right channels based on what the data says not based on the model were using.

    Does this make any sense?

    • 103

      David: I'm not sure I completely understand what you are expressing.

      In part two of your comment I feel like you are saying do part one (but there you express that that is not the right thing).

      Let's step away, think of it this way…

      Last-click attribution is silly (as is first-click, for the same reasons and more).

      So you should do something better.

      If you don't want to work very hard, go with time-decay type models. They are the next best step away from last-click. Your decisions will be better. (Not perfect, better.)

      If you want to work hard, experiment with various custom attribution models. If you balance for factors that are in the top part of my custom attribution model in this post AND the bottom part, you will get much better answers.

      If you are willing to work very hard, because it is worth it, you are that big of an advertiser/player on the web, you should in a persistent media-mix modeling practice inside your company. You eliminate the whole modeling thing, you will get very close to the perfect answers. You have to be willing to do this persistently, invest in people, process and technology to pull it off.

      Pick your poison. :)

      Avinash.

  53. 104
    liberty says:

    Dear Avinash,

    I'm so happy to receive your reply this afternoon. I've asked for a website from my frd and sign up a Google Analytics account to have a try this week! But I still have a question to ask you:

    I still not understand how the Google computes the path. As you said, the path of the user online is so complicated, and you're not sure their browsing order. So how can Google compute the path? Is it based on the linear correlation between the channels or the conversions rates, or any other factors?
    This question has confused me for about a week, and what answer I search on the Internet is just THE MACHINE/GOOGLE CAN AUTOMATICALLY COMPUTE FOR YOU. But I wanna know the calculation method and how!

    Looking forward to your reply soon.
    Have a good day:)

    • 105

      Liberty: You can get help from an authorized consultant, they are quite affordable and can help you get on the right tack pretty fast. Here's a list: http://www.bit.ly/gaac

      Here's a simple version of what all web analytics tools, including Google Analytics do… Each time you visit the site they capture the referrer and match it to your cookie id. Then when they create the reports, they are able to know the "path" you took during your visits.

      -Avinash.

  54. 106
    Samuel says:

    Hi Avinash,

    Thanks for your awesome post. Despite not being a data analyst, I was able to use your post to make a customized attribution model (based on time decay) that does wonders and makes my strategic dashboard much more solid :)

    I have one big question though.

    Basically, I I'd love to use the same kind of model to attribute a value to each and every of my micro-objectives. Right now I kind of try to guess their value and see if that makes any sense. That works more or less… But I'd love to have something much more robust (there are things I like better than saying to my CEO "well, # of whitepaper download is worth 50€ because I think it's worth 50€" :p)

    The problems is I can't find how to do that on GA (did I already say I'm no Data Analyst?). I do track most of my objectives there, but I can't find how to track assists/attribution for them.

    In other words, I can measure "channels" attribution / assists for "micro-objectives", but not "micro-objectives" attribution / assists for "macro-objectives".

    Thank you for any answer you can provide.

    ~ Samuel

    • 107

      Samuel: I'm sorry, but I'm not sure I understand your comment. Let me try to share some thoughts, perhaps some will help.

      If you want to identify the value of your micro-outcomes, I have a long and detailed post on that. Here it is: Excellent Analytics Tips #19: Identify Website Goal [Economic] Values.

      The post will help you figure out the best way to get the value of your whitepaper.

      If you want to apply your custom (or the included standard) attribution model to only one, or a cluster, of micro-outcomes (or just the macro one), you can use the Conversion Segments feature in GA. If you go to the Model Comparison Tool in the attribution folder, at the top you will see something called Conversion Segments, just click on it and now you can create segments. Perhaps this will help.

      All the best!

      Avinash.

  55. 108
    Juan Fco Romero says:

    Wow!. This is a very complete artícle and it's very clear with the images. I want to congratulate to you.

    Thanks.

  56. 109
    Anirban says:

    Hi Avinash,

    Thanks for a great post! I have a background in predictive modeling and just starting to learn the ropes of marketing analytics.

    I just have a quick question – do you have any thoughts/documented case studies about blending path length with the time decay model?
    Reason I ask – instead of setting an absolute half-life #, shouldn't we be trying to gauge the true impact of each touch point?

    And thanks again for the clear and detailed (and very readable, of course!) presentation of all the aspects.

    All the best!
    Anirban

    • 110

      Anirban: At the moment in Google Analytics you can only set the half life number for the time decay model. In my personal use of some of the advanced attribution modeling techniques, we have taken data out of Google Analytics and applied various strategies in a customized data environment. I'm afraid none of that work is in the public domain. My apologies.

      If you wanted to gauge the true impact of each touch-point, please see some of the strategies I've applied in the bottom part of my Market Motive Mindblowing Attribution Model. I am creating multipliers for the touch-points that deliver better engagement, higher value (even if no conversion). You can consider using something like that in GA at the moment. That entire section is totally customizable and quite feature-rich.

      Avinash.

      • 111
        Anirban says:

        Thanks for the response.

        And I was afraid you would say that the details of such model mixing is not in the public domain! :-) Nonetheless, thanks for indirectly supporting my thoughts!

        I will take your advice on exploring GA's customization capability.

        If I see something interesting and worth sharing, will reach out to you again.

        Thanks!
        Anirban

Trackbacks

  1. […]
    So while attribution vendors are getting very good at tracking exposure to banner ads and search – and doling out credit – they are actually following cookies, not people. In a world where more than 60% of us own a smartphone, almost half own a tablet, and 82% of global consumers in a recent Microsoft-led survey said they like “multi-screening,” the damage in terms of inefficient marketing is major.
    […]

  2. […]
    Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models, http://www.kaushik.net
    […]

  3. […]
    Google “Digital Marketing Evangelist” Avinash Kaushik lays out on his personal blog what he thinks makes effective multichannel attribution modeling on his site. He begins, “My macro goal is to make you dangerously informed. By the end of this post, if you pay attention, you'll know the often hidden nuances and you'll be dangerous to any analyst/consultant/vendor who walks into your cubicle/office …” Read dangerously, my friends.
    […]

  4. […]
    Danielle Uskovic‘s insight:
    I’m a big fan of Avinash Kaushik and avidly read his blog. Attribution modeling is a hot topic for brands and this article looks at all types of measuring this beast.
    See on http://www.kaushik.net
    […]

  5. […]
    A januári cikkünket Luigi Reggiani Superweekes előadása alapján írtuk. A különböző Google Analyticsben elérhető modellek hasznáról és használhatóságáról nemrég Avinash Kaushik írt egy hosszú és hihetetlenül hasznos bejegyzést. Az alábbiakban az ő véleményükre alapozva mutatjuk be az egyes modelleket.
    […]

  6. […]
    n the world of PPC, it’s always great when you see leads/sales coming in through your campaigns: search, display, remarketing, social, etc.. But when it comes time to make optimizations and allocate budget dollars, you need to know which channel influences your customers the most. That’s where Avinash comes in. His article Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models is not solely a PPC specific article, but all of this points can be really helpful when setting up attribution modeling. He even walks you through how to do it! If you’re tired of bid and ad copy changes and want to step outside of the box for optimization inspiration, then grab a coffee and hunker down for this long yet informative post.
    […]

  7. […]
    Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models by Avinash Kaushik another super must read article from Avinash on what's ugly and bad when it comes to attribution modeling.
    […]

  8. […]
    In seinem letzten Blog-Eintrag hat der bekannte Webanalyst Avinash Kaushik einen Artikel über gute und weniger gute Attributionsmodelle geschrieben (http://www.kaushik.net/avinash/multi-channel-attribution-modeling-good-bad-ugly-models/). Da jedoch in jedem Unternehmen der Online-Marketing Mix unterschiedlich ist, muss in jedem Unternehmen zu erst einmal die Customer Journey jedes Besuchers genau analysiert werden, um herauszufinden, wie der Online-Marketing mix funktioniert. In folgenden Blog Einträgen werde ich unter anderem darüber schreiben, welche Schritte durchlaufen werden müssen, um ein passendes Attributionsmodell zu finden und die Erfolgmessung bzw. Analyse zu verbessern.
    […]

  9. […]
    Na apresentação sobre métricas online do Leonardo Naressi, ele lembrou sobre o papel de olhar as métricas com a inteligência de um analista. Também houve uma demonstração das potencialidades da ferramenta da Ignition One e uma explicação sobre o que é marketing de atribuição pelo Edmardo Galli, além do encerramento com uma boa reflexão sobre inovação com o Fernando Kimura.
    […]

  10. […]
    Multi-channel attribution modeling: the good, bad and ugly models, from Occam's Razor by Avinash Kaushik.
    […]

  11. […]
    What I felt today was different, though. It came when I was looking for some info on a piece of software I use (Google Analytics). I started at the website of a well-known expert in the field, Avinash Kaushik at his The Good, The Bad, and The Ugly piece. That article led me to his Definitions, Models, and a Reality Check piece. That article led me to his Tracking the Online Impact of your Offline Campaigns piece, and that article led me to David Hughes’ website because he coined the term “non-line” that Avinash uses (it means marketing efforts that exist both on- and offline, like the color of a logo).
    […]

  12. […]
    Avinash Kaushik shares insight on multi-channel attribution modeling: the good, the bad and the ugly.
    […]

  13. […]
    It’s a common marketing question that has several possible solutions. There’s last click attribution, where the last step gets all the credit (the search ad). There’s first click attribution, where the first step would get all the credit (the TV placement). There’s also linear attribution where you give equal credit to each touch point in the process. Avinash Kaushik put together a wonderful post detailing some of the various types of attribution models.
    […]

  14. […]
    Solutions such as Google’s pack in as much data as possible about a particular users’ exposure to your ads — what, when and where — and look at user-level data across vast swarms of humans. It then constructs a best-fit description of the info it has, saying things like, “Display ads contributed 21.2% ($45K) to the campaign’s success.” But as Google’s resident evangelist Avinash Kaushik recently warned, “There are few things more complicated in analytics … than multi-channel attribution modeling.” So in the words of a musical team from the pre-Google era, don’t believe the hype. At least, not all of it.
    […]

  15. […]
    Therefore, basing a marketing budget solely upon this method would undervalue the contribution of social media to the conversion process. Google Digital Marketing Evangelist Avinash Kaushik wrote an excellent blog post on attribution modeling, addressing these issues. He opined the Time Decay Attribution Model does a fairly good job above and beyond the last click, and I would agree.
    […]

  16. […]
    If you want to read up on multi-channel attribution models, I recommend Avinash Kaushik's work Multi-Channel Attribution: Definitions, Models and a Reality Check and Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models.
    […]

  17. […]
    Secondly, Cross-device tracking & attribution modelling are here to stay. The holistic value of search across multiple devices and channels and; the complex paths to purchase are likely to become an important aspect of PPC budget allocations in 2014.
    […]

  18. […]
    Multi-channel, also as the name suggests, spreads out the conversion value to better represent the customer’s relationship. My personal favorite method, and as Avinash Kaushik points out as the easiest to go astray, is the Position Based Attribution Model. This particular model is flexible, breaking down the total conversion value into two lump sums (equal for first and last) and then evenly distributes the rest through the other conversion steps.
    […]

  19. […]
    Favorite: “Don’t Silo Mobile Marketing” – This one could not be emphasized more. One of the key elements I’ve seen in mobile marketing is its presence in the conversion path. In plain English this means people may find you with a mobile search but convert on a web browser. Some people research a purchase several times before they convert. You want your product or service to be where their eyes are. Even when mobile doesn’t convert as last click it is very often present somewhere in the funnel. If you want more information on how this process works you should research multi-channel conversion attribution models. This is a great article by my favorite author on that subject: “Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models” by Avinash Kaushik.
    […]

  20. […]
    “First click attribution is akin to giving my first girlfriend 100% of the credit for me marrying my wife.” — Avinash Kaushik
    […]

  21. […]
    Now, if you read the first paragraph, then continued onto the second paragraph, and you've respected the basic tenets of literacy by moving onto this, the third paragraph, you might already know where I'm going with this preamble. We have entered the age of attribution modeling, where we try to right the wrongs inherent in these tools by affording a certain slice of the attribution pie to all the channels that participated in the conversion (i.e. making the visitor fulfill a goal on the website). Check Avinash Kaushik's nice introduction to the topic here.
    […]

  22. […]
    En resumen, es algo muy útil para saber si comunicamos bien y, sobre todo, para conocer cómo acceden los usuarios a nuestros servicios. Eso sí, siguiendo los consejos de Avinash Kaushik, antes de hacerlo hay que:
    […]

  23. […]
    That’s a really big question. The best answer I have is to direct folks to this excellent post by Avinash Kaushik. While it is a fairly long read, it is very well written and explains a very complex topic in very clear language.
    […]

  24. […]
    *** You still have to be using the right model. This is the point in the course where my brain exploded all over my laptop, so I will hand you over to the wonderful Avinash Kaushik for enlightenment.
    […]

  25. […]
    On GA attribution modelling.
    […]

  26. […]
    When you’re looking at tightly cut back budgets, affiliate marketing is often the first channel to get the chop. Before you consider that, though, let’s get our attribution on. I recommend getting a strong cup of coffee and setting aside a solid day to work through each example given by Avinash in this superb post on the basics behind attribution models. Try out a few examples and see how they would influence your decision, specifically, on cutting back a channel that’s currently driving revenue.
    […]

  27. […]
    You'll want to track all the usual online metrics, including conversions from awareness channels. While direct conversions might be relatively small, be sure to look at multichannel attribution. It's likely you'll see many conversions coming from organic or direct traffic that were influenced by your awareness efforts.
    […]

  28. […]
    Of course there are many marketing attribution models you can follow. But that’s a post for another day. For now, just continue business as usual, making sure to follow a consistent tagging structure. Spelling and capitalization count! Be sure to establish and follow patterns to avoid a big mess when it’s time to analyze your efforts. After all, “LinkedIn” vs. “linkedin” will show up as two separate records if you get lazy.
    […]

  29. […]
    Create a winning custom attribution model (Try this one from @avinash)
    […]

  30. […]
    Avinash Kaushik’s post on the good, bad, and ugly models for multi-attribution modeling
    […]

  31. […]
    Some marketers opt to give the first “touch point” rather than the last “touch point” 100% of the credit. So in the case above Facebook would get the “goal” for being the first source to let the buying customer know about your online course. However, this is problematic as well because it ignores the steps in between. Another blog put it perfectly when the author wrote “First click attribution is akin to giving my first girlfriend 100% of the credit for me marrying my wife.”
    […]

  32. […]
    If you’ve read up on attribution modelling in the past, you probably already know what’s wrong with the default models. If you haven’t, I recommend you read this post by Avinash, which outlines the basics of how they all work. In short, they’re all based on arbitrary, oversimplified assumptions about how people use the internet.
    […]

  33. […]
    In this scenario, a few things would happen to the KPIs of most e-commerce companies: The main website’s conversion rate would look better during the day of the campaign, since some customers like Alex logged on to it directly to purchase. The conversion would be attributed to the “direct” (also known as organic) channel, since that’s the last channel the customer came in from.
    […]

  34. […]
    On the other hand, by using a multi-touch attribution model, the eventual sale gives credit to all the marketing channels that influence your customer during their awareness, education and consideration stages of the buying cycle. Plus, since Social Media is only one part of the marketing mix, knowing how other touch points along the sales cycle influence your prospects will provide insight into understanding how you can accelerate and optimise the entire process. There are many tools out there that measure multi-touch attribution, such as Marketo and even Google Analytics. Here’s a useful article to learn more about the topic.
    […]

  35. […]
    If you’ve read up on attribution modelling in the past, you probably already know what’s wrong with the default models. If you haven’t, I recommend you read this post by Avinash, which outlines the basics of how they all work. In short, they’re all based on arbitrary, oversimplified assumptions about how people use the internet.
    […]

  36. […]
    Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models – Goes into detail about the various types of attribution models and how to best apply them to your marketing efforts.
    […]

  37. […]
    功能歸屬模式會將大部分的功勞歸給距離達成銷售或轉換時間最近的接觸點。在這個銷售範例中,「直接」和「電子郵件」管道會分到最大的功勞,因為客戶在轉換前幾個小時內與這兩個接觸點有互動;「Social Media」管道分到的功勞會比「直接」或「電子郵件」管道來得少。而由於「AdWords 廣告」互動發生在一週前,因此這個管道的功勞就少很多。
    上述5種功勞歸屬模式,各有長短,須視乎網站性質、情況而用,而筆者最常用亦覺得最有用的是「時間衰減」(Time Decay)。
    圖片來源:http://www.kaushik.net/avinash/multi-channel-attribution-modeling-good-bad-ugly-models/
    […]

  38. […]
    There are quite a few attribution models but they are more complicated and I don't know anyone who uses them in the gaming industry. I think it would make sense to use them for the travel industry when you want to know whether a hotel is good or not before booking so you read reviews and information online from multiple sources before making your final decision. In which case the Customized/Personalized Attribution Model looks great but I personally only know this from following Avinash Kaushik (Google's Evangelist), I don't speak from experience
    […]

  39. […]
    Fact #3: Consider that 50% of all shoppers are in market 90days or longer, and visit >20 automotive sites prior to purchase. Your not looking for the lead source, your looking for the MOST POPULAR paths to your store. In other words, if you obsess about lead source, what of the 19 other sites the shopper was on prior to sending you a lead? Where did they go after they sent you a lead? Look into attribution modeling*.
    *https://support.google.com/analytics…/1662518?hl=en
    *Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models – Occam's Razor by Avinash Kaushik
    […]

  40. […]
    Some readers were surprised this post didn’t discuss multi-touch attribution within digital channels. For those interested, this post provides a good overview of good and bad multi-touch (digital) attribution. I recommend starting out with the time decay model, where the media touch point closest to conversion gets most of the credit, and the touch point prior to that will get less credit based on a smart and simple algorithm.
    […]

  41. […]
    The Power of Attribution
    While Gaffney’s advice is sound, each enterprise will have different requirements as to what they put into an attribution model. As Avinash Kaushik writes:
    […]

  42. […] Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models […]

  43. […]
    The __reff cookie takes care of this part by allowing you to actually save all the referrers that brought the user to your website before he converted. You will actually be able to apply some advanced attribution modelling because of this, though that would require some advanced database skills on your side.
    […]

  44. […]
    So you can see the logistical challenge. It gets worse. Of course, the model wouldn't be accurate unless it tracked people across all their devices, also, so that it knew that Martin on this version of Explorer is the same Martin on that iPhone and on that tablet and booting up that app on his X-Box One. And I haven't even mentioned the nontrivial (nerd-speak for "really, really hard") problem of achieving statistical significance, and that experts don't agree on a formula or algorithm to use. (Google introduced one last year based on game theory. For a great overview of attribution models in general, see Avinash Kaushik's blog post.)
    […]

  45. […]
    It’s possible that some of the shoppers rummaging through the marked-down sweaters in your bargain basement saw your banner ad this morning. Possible, but not likely; it’s more likely they’re there because they know exactly when you mark down sweaters every season. Complaints about revenue attribution usually center on the “last click” versus “full-funnel” debate, a tiresome argument you can avoid by insisting on measurement through market testing.
    […]

  46. […]
    These and other data-driven attribution models are many and varied. And they can be fairly challenging to grasp. A great view of all the options is in this post by Avinash Kaushik – author of two best-selling books on Web Analytics.
    […]

  47. […]
    Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models : when marketing get the detail of the attribution models and apply them to get the highest traffic possible.
    […]

  48. […]
    Useful links: Customer Journey to Online Purchase – Google Default channel definitions – Google Analytics Avinash on attribution models
    […]

  49. […]
    Time to Conversion (or sales cycle length) – Somewhat tricky and varies widely across industry. This aids sales teams to see what’s in their pipeline. Attribution Modeling is a great way to grasp the sales cycle. But be careful and don’t over-emphasize reducing time to conversion.
    […]

  50. […]
    Avinash Kaushik has written extensively on the subject and gives a good overview of multi-channel attribution modeling to get you started thinking about what is best for you. Whatever path you choose, make sure that you learn what efforts provide the most value. That way, when it’s time to review your strategy, you can do so in the most educated way possible.
    […]

  51. […]
    Taking this analysis of ROI a step further, using the multi-channel funnels assisted conversions report, we can begin to understand which channels contribute to conversions overall and how: Now we can explore the next level of analysis by attribution modeling, so we can grow resources for channels based on their overall contribution. For more on modeling, check out this post on multi-channel attribution by Web analytics guru Avinash Kaushik.
    […]

  52. […]
    HOW-TO, BEST PRACTICES, ATTRIBUTION MODEL REVIEW
    It’s not a recent article but always good to read: Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models
    […]

  53. […]
    Google Analytics permite crear modelos de atribución alternativos. No hay un paradigma correcto y otro erróneo, puede variar en función del tipo de negocio y del segmento de mercado. Sugiero un artículo interesante de Avinash Kaushik, referente en el campo de la analítica web, donde explica las ventajas y desventajas de diferentes tipos de atribución.
    […]

  54. […]
    Sin embargo, aunque este razonamiento nos parezca obvio, superarlo y poner en práctica otro modelo en una estrategia de marketing es todavía complicado porque las herramientas de analítica más extendidas –incluido Google Analytics- siguen apostando por defecto por el modelo de atribución al last click. Además, ¿de qué alternativas disponemos? Tomando como guía el completísimo análisis de Avinash Kaushik, podemos evaluar las siguientes:
    […]

  55. […]
    For marketers, this information is critical to understanding where our customers are, where they are interacting, and, therefore, where our time needs to be spent. If we only look at last-click conversions, we do not truly understand the customer journey. Here are a few resources worth checking out: Multi-Channel Attribution Modeling: The Good, Bad and Ugly Models
    […]

  56. […]
    You can read more about the different types of Google Analytics models and how to use them here. Also, Avinash Kaushik, Digital Marketing Evangelist at Google wrote a blog post describing how to use the different reports. The biggest difference between using reports on Google Analytics vs. HubSpot is that you cannot connect the report back to specific contacts unless you use HubSpot. That means you cannot apply attribution trends to specific personas, contact groups, lifecycle stages, or other categories that are relevant to your business. So when deciding on what report to use, take that into consideration.
    […]

  57. […]
    You can read more about the different types of Google Analytics models and how to use them here. Also, Avinash Kaushik, Digital Marketing Evangelist at Google wrote a blog post describing how to use the different reports. The biggest difference between using reports on Google Analytics vs. HubSpot is that you cannot connect the report back to specific contacts unless you use HubSpot. That means you cannot apply attribution trends to specific personas, contact groups, lifecycle stages, or other categories that are relevant to your business. So when deciding on what report to use, take that into consideration.
    […]

  58. […]
    The overwhelming number of attribution models available further complicates an already thorny issue. As a hypothetical example, consider your marketing team has decided it’s time to switch from last-click attribution to another measurement methodology that better accounts for cross-channel conversion paths. There are many multi-touch attribution models that would represent an improvement over last-click, so which is best?
    […]

  59. […]
    In this scenario, a few things would happen to the KPIs of most e-commerce companies: The main website’s conversion rate would look better during the day of the campaign, since some customers like Alex logged on to it directly to purchase. The conversion would be attributed to the “direct” (also known as organic) channel, since that’s the last channel the customer came in from.
    […]

Add your Perspective

*