Mastering Marketing Attribution: A Comprehensive Guide

Facebook, Google, Bing, Outbrain, TikTok, LinkedIn and all other platforms send traffic to your digital assets day and night. We spend a lot of money on our campaigns, which we might run by ourselves, through freelancers, agencies or other units of the organization.

As the owners or analysts of an app or a website, we consider its performance holistically. Google Analytics and other analytical tools analyze the overall activity on our site, taking into account all of the campaigns we run. However, each platform that provides traffic to our site primarily focuses on its own interests and is not designed to coordinate with or work seamlessly with other platforms. Facebook Ads does not “talk” with TikTok Ads. They do not share data between themselves. No one does.

Typically, each platform contains three modules:

1. The campaign management system is used to create campaigns, write ads and test new creative ideas. This is where campaign managers manage their campaigns.
2. The analytics system, which includes tools for tracking events and conversions, that usually operates independently of whether a campaign is currently being run.
3. The attribution system connects the other two modules, allowing tracked events to be associated (or attributed) with specific campaigns. This is the reason for its name.

How Attribution Works

Please welcome Betty.

1. Betty sees an ad on Facebook. Facebook recognizes her because she was still on the platform when she saw the ad. Facebook then logs the following information: “Betty, ID Number 123456, just saw ad number 55555. This ad is part of a Black Friday Instagram campaign with ad account no. 99999.“
2. Three seconds later, Betty clicks on it. Facebook then logs the following information: “Betty, ID Number 123456, clicked on ad number 55555. This ad is part of a Black Friday Instagram campaign with ad account no. 99999.“
3. Each time a user clicks on any content type that redirects outside of their platform, Facebook generates a click ID, also known as an fbclid (stands for “Facebook Click ID”), whether it is an ad or organic content. The fbclid is a long and unique number that allows Facebook to track the source of the click. When Betty lands on your site, the fbclid is included in the landing page address: domain.com/page/?fbclid=1234567.
4. You likely have a Facebook tracking pixel installed on your site, which can do a variety of things, such as creating and updating cookies and storing information on your browser. Among other things, the pixel immediately saves the fbclid.
5. When the page is loaded, the script runs and informs Facebook that a pageview has occurred, adding the fbclid to the information package.
6. Facebook knows now that Betty viewed the homepage of domain.com and that her fbclid was 1234567.
7. Later, Betty submits a lead form. This action triggers the process described above again, sending a Lead event to Facebook with Betty’s fbclid. Facebook’s analytics module records the Pageview and Lead events, and that’s what you see on Facebook Events Manager screen. In the image below you can see that Facebook received 459,000 Pageviews and 11,000 leads.
   
   
   
8. Important! Facebook Events Manager counts all events, including those that did not result from Facebook traffic. This means the count should match the ones in your Google Analytics, or any other analytics platform you use.
9. At the same time, on their servers, Facebook decodes the fbclid, reverse-engineering it if you like, learning that it was associated with Betty’s click on a specific ad for a specific ad account.
10. Now the fbclid acts as a KEY that connects between the campaign, user behavior and Facebook user ID. With this key, Facebook attaches the event to the appropriate ad and campaign in their Campaign Manager, allowing the campaign managers to see whether the ad caused a conversion.
11. This attachment between an event and a campaign is called “attribution” or local attribution because it only applies to Facebook’s own campaigns.
12. Generally, the same process (1-11) happens with all ad networks.

The Egocentric Attribution

Let’s stick with Betty for a while.

On Monday, Betty clicked on one of your ads, reached your site, browsed a little bit, viewed a certain product page, but eventually, didn’t buy a thing, hence, we have no conversion. At this point, Facebook already registered the click and knows what Betty did.

On Tuesday, a day after, Betty opens her computer and miraculously, she sees your Google remarketing banner, that offers her the same product with free shipping. Betty clicks, redirected to your site and since it’s a Google Ads ad, the same process I’ve described earlier – happens here. This time, instead of fbclid, we have gclid (stands for Google Click ID.)

Betty reads about the product (again) and clicks the “Add to Cart” button. At this point it triggers Add-to-cart event to ALL platforms – Facebook, Google, Snap etc.

Unfortunately, Betty doesn’t wanna buy anything, so she leaves.

Several days later, Betty remembers that she was interested about your brand, goes back to Google Search, looks for your brand (“Banana Sport”), click the first organic result (not an ad), and finally makes a purchase.

Now, what happens on the Thank-you page? You are right. Your website fires a Purchase event to all networks. As I said, all pixels are fired regardless the traffic source.

But now we have a problem

Google Ads, may claim the conversion as their own, as the ad that was clicked was from their platform and the sale event alert was sent to them. Similarly, Facebook may also claim attribution as they were the ones who recognized Betty’s click in the first place, even before the purchase was made.
This results in both platforms claiming attribution for the same purchase, creating discrepancies in attribution data. Additionally, when different marketing agencies operate different ad platforms, each agency may exaggerate the effectiveness of their own platform, making it difficult for marketing managers to allocate advertisement budgets. The complexity increases with the number of platforms involved, for example, advertising on ten platforms instead of two.

The Attribution Window

To ensure accurate attribution, every platform provides the option to set a timeframe for attributing a sale. For instance, if a certain number of days have passed from the moment an ad is clicked to the moment the sale is completed, that platform cannot attribute the purchase to itself. The platform will still count the purchase, but it will not be attributed to any campaign. Defining an appropriate timeframe can be challenging, but we will examine the process in Google as an example, with the assumption that it is similar in other platforms. Open your Google Ads account, go to Settings->Conversions and choose a specific conversion. Now let’s go through the settings:

1. Count – For example, for “transaction” conversion type, “Every” should be selected. This ensures that Google counts every transaction. If “One” is selected, it will only count one conversion, such as filling out a lead form, for each user.
   
2. Next, set the Click-through Conversion Window. Google prompts for the duration of time from the click to the conversion, which can be up to 90 days. In this scenario, if the ad was clicked on in early January and the conversion was completed in late March, Google will attribute the conversion to itself.
3. For YouTube campaigns, the Engaged View Conversion Window can also be set. This is the duration of time from the moment a video ad is viewed to the moment of conversion, for Google to attribute the conversion to itself. For example, in the case of a user watching an In-stream ad and converting without clicking any ad.
4. The View-Through Conversion Window can also be set, determining the duration of time for Google to attribute a conversion to itself in the event that a user only viewed the ad without clicking it, but converted later on.
5. Finally, the Attribution Model is selected. This is crucial for those running multiple campaigns simultaneously, targeting potential clients in different ways, such as remarketing, shopping ads, search results, and video ads. In such cases, the potential customer may have multiple touchpoints (viewing the YouTube ad, clicking the shopping ad, and clicking a brand ad a few days later). Therefore, it is important to determine which campaign the conversion should be attributed to. For example, if the Last Click model is selected, the conversion will be attributed to the last campaign the user clicked on. If Linear is selected, the attribution will be evenly shared by all ads the user was exposed to. If you choose “Data-Driven” you’re saying “Listen Google, I don’t have a clue what to choose, so I’ll let you choose”.
6. No matter which model you choose, it applies only inside Googel Ads, and sets the credit mechanism between Google Ads campaigns.

So, Where’s The Catch?

Assuming that all ad platforms have similar settings, the longer the attribution window is set for each platform (e.g. 60 days instead of 3, and considering view-through conversions rather than click-through conversions), the more conversions the platform will attribute to itself, potentially including conversions that were not generated by the platform.

Let’s see another example:

(X-axis is dd/mm)

(The length of the bar represents the attribution window according to your settings.)
A user clicked on a Snap ad on 1/11, leading to your site, then clicked on Facebook ad on 3/11, and on your TikTok on on 4/11. T

The actual conversion took place on 14/11. Since the conversion took place outside of Facebook’s attribution window, Facebook will not claim the conversion, while both Snap and TikTok will each claim the conversion as their own, since the conversion took place within their attribution settings. Additionally, platforms count the conversion on the day of the click, rather than the day of the actual conversion, which can result in post-factum attribution. That’s the big debate.

It is important to decide which platform should be credited with the conversion, whether it be the Google campaign that initially attracted the customer to the site, the remarketing campaign ad that the user clicked on multiple times, or none of those, but rather, your SEO.
In general, it seems that the best way is to set your attribution window as narrow as possible to minimize overlapping attribution, and make the platforms work harder for each conversion. But it’s not the best way. Far from it.

Another aspect of messing with attribution window is relevant to B2B sites where it may take several months for users to convert from the moment they are exposed to the site or a webinar. Bummer, eh? Now let’s meet a new and a very important player that joins the attribution game.

Attribution In Google Analytics

In this in-between period, I’ll talk about attribution on both Google Analytics 4 and Universal Analytics. . These are two distinct systems and the attribution process for each one of them is a bit different.

In contrast to the attribution windows I’ve described before, Google Analytics does not have a specific attribution window. Instead, it displays the user’s visits according to their sources (if they are properly tagged with UTM parameters, for example). For example, if a customer makes 5 visits before conversion, these visits are stored as touchpoints:

Universal Analytics adheres to the Last Click Attribution model. In this scenario, the conversion will be attributed to Taboola (without notifying the channel). All of Google’s conversion reports follow this model, for instance, Conversions by Country only displays the last country in which the site was visited and in Conversions by Time, the conversion will only be attributed to the most recent visit to the site. To provide a more comprehensive understanding, the Multi-Channel Attribution reports were designed to give users, particularly analysts, a more complete picture.

Multi Channel Attribution Reports

These reports can be found under Conversion Reports. Under it, you can find Top Conversions Paths, which only shows the touch points of users who converted:

First, let’s configure the settings. Under “Conversions,” the type of conversion to focus on, such as sign-up, download, or purchase, should be selected. It is not recommended to use the “all conversions” default.
Under Path Length, select “one or more”, which will display conversions that had only one touch point. The Lookback Window allows you to set the period of time during which you want to track each conversion type, with the default being 30 days.
As shown in the display, two conversions took place, each resulting from 2 visits, one from Google/cpc and the other from Facebook/cpc. In other reports, the conversion would be attributed to the latter, however, both sources are mentioned here.
Under Conversion Segment, you can select the channel you want to focus on, for example, Paid Google.

Here, you can select “Any Interaction,” which will display Paid Facebook activity throughout the conversion process, or “First Interaction,” which will only display conversions where it serves as the initial touchpoint. Advanced filtering options are also available.
Please note that this report only provides an understanding of how your channels are interconnected but does not assist in determining how to allocate attributions. For example, you may prefer to attribute conversions to the initial touchpoint, the final touchpoint, or divide the attribution equally. To assist with this, Google has designed their Model Comparison Report.

By selecting “First Interaction” the conversion will be attributed to the channel that initially attracted the user. By selecting “Position-Based” the attribution will be divided between that channel and the one that attracted the user last, with no attribution given to the channels in between.
You can choose either a Single Touch Attribution Model or a Multi-Touch Model report. Under Single Touch Attribution, you will find all models that credit a single channel only. Although they are still effective analytical tools, it typically takes more than one visit for most users to convert. Under Multi-Touch Attribution, you can select a model that divides the attribution among multiple channels.

But which model should we select? Let’s see how Google Analytics 4 deals with that, and then I’ll try to answer this question.

Attribution In Google Analytics 4

In Google Analytics 4, the default attribution model is Cross-Channel Data-Driven, which distributes the attribution among all channels that participated in the conversion, excluding the direct channel. This means that GA4 operates under Multi-Touch Attribution. The distribution of attribution is complex and follows undisclosed principles, similar to other platforms. Additionally, if your GA4 has various settings for conversions, the model will analyze each of them differently.
GA4 also offers more advanced models than the Universal version, in addition to Single Touch Attribution.
Alternatively, if you prefer not to use Google Analytics, you can set the same models on your servers, assuming all your platform and analytics system data is downloaded to your own data software.

The Challenges With Attribution

(1) Attribution In the Cross-Device Era

The average user uses several devices, such as mobile phone, PC, and the tablet computer. Since most social networks require login, they can easily bind all user devices under one user. Though Google has no social network, they have Android, Gmail and other means to locate the users’ devices and bind them under one user (although it’s a bit trickier). Unfortunately, Google Analytics does not receive this information from Google/Alphabet, and has major issues with consolidated all user devices under one user. In Google Analytics 4, Google account “Signals” feature that should have help with it, but it’s really not working as it should. Therefore, these systems may fail to recognize that different users browsing through several devices are basically the same one. They do try, but it’s far from working.

(2) In-App Browsers

If a user clicks on a link inside TikTok or Facebook, it will probably open in the App’s built-in browser, rather than in the user’s usual browser. If the user moves from there to your site using his default browser, probably some of the tracking information will not migrate with them. Thus, if an item was added to user cart on an in-app browser, the purchase info will not be sent to the ordinary browser. This makes the functionality of cookies and tracking trickier, and results in broken sessions and other issues. On top of that, the privacy policy changes on Apple devices, starting iOS 14 makes it even trickier.

(3) Attribution of Off-line Channels & Branding Activity

If, for example, you advertise on TV and radio, billboard ads, booths at conferences or even pay to TikTok influencers, in addition to your measurements online advertising activities, the effects of your offline ads are hard to measure. If you have a strong brand, you cannot exactly say how a temporary suspension of paid promotion will affect it. It might have marginal product, but it also might not.

(4) Non-direct Influence

We do not live in a vacuum. People talk about your brand, websites mentions it, a YouTuber just did a nice review on your product. Your activity may trigger many different responses beyond your direct control which might have surprising effects on your site performance.

(5) Data Quality

Data in sufficient quantity and quality are hard to gather, since they are not always fitted with proper tracking systems, missing tags, low-quality implementation, missing UTMs or simply lack of volume. Do not waste your time and resources on advanced attribution models if you have 10 conversions a week. Even Google Analytics data driven model requires at least 600 monthly conversions. A plenty of data minimizes noise, smoothes statistics and stresses seasonal changes.

Lets summarize the attribution challenges:

1. Multiple online marketing channel, each with its own attribution windows and conversion count model.
2. Different analytics platforms show different data based on different algorithms.
3. Users delete cookies or don’t give their consent to be tracked.
4. Long conversion times, up to several months or even more.
5. The possible effects of offline channels. Some call it Implied Attribution model, which examines the effects of these channels on direct on organic traffic. Such can be the effects of the beginning or ending a radio campaign, your presence in a conference which temporarily increased the website visits. This model is only partially effective, since an activity focused on branding has no immediate effect.
6. Not all online or offline channels’ effects must be weighed.
7. You don’t know which model to choose.
8. The effects of cannels beyond your control.
9. The effects of your branding activity, which is practically non-measurable.
10. The effects of Cross-Device browsing.
11. Technical issues, measurement problems, improper measurements, etc.

The Attribution Model Premise

1. No model is perfect
2. You’re not alone here. It’s everyones problem.
3. A model’s popularity indicates nothing. Many choose a U shape model (40% of the attributions goes to the first channel used, 40% to the last channel, and the other 20% splits evenly between all the in-between channels), but it may not work for you.
4. Do your own testing and trust nobody’s recommendation (including mine.) What works for company A may not work for company B.
5. Even the best of models is only based on correctly gathered data. It may be affected by noise, corrupted data or missing data.

How To Choose The Right Attribution Model

First of all, let me assume that all your campaigns are properly tagged, your pixels work perfectly, and your analytics display conversions accurately. During the early stages of your campaigns you might prefer a wider attribution window in order to let the platforms get to know your users and build target audiences, even at the cost of letting networks attribute your conversions to themselves. Later on, if you think it’s necessary, you can narrow it.
In addition, you should have not only hard-core conversions, such as a sale or receiving a lead, but also micro conversions, like “Add to cart”, “Product view”, newsletter sign-up etc.

Questions Needs To Be Asked

How long is your sales cycle?
How long does it take user to convert from the moment of exposure?
How long does it take to decide to buy?
In some products and services, it doesn’t take long. Suppose you sell a car insurance policy. People already know what “insurance” means, and just want to hear a price-offer since their policy expired and they want a new one. They may even know your phone number due to an intensive radio campaign in the past, so they just call you. In this case, the sale will not be attributed to any channel. They might also Google “car insurance” and click on some results. This will probably take no more than a couple of days to convert. In this case, the attribution model should be close to Single Touch.

If you wish to attribute the conversion to the first ad the customer saw, select First Click. This is, if you suppose that without that ad, you would’ve lost that customer. In other words, if you know your product and target audience, you can figure out your customer-journey until it reaches you, and select the suitable attribution model.

If the customer journey is a several months long, it involves completely different setup.
For example, you’ve developed a new SaaS for DevOps. You have already given webinars for people in the industry and regularly show them remarketing ads on Facebook and Google. At the same time, you also start an your sales people target potential customers with private messages, and content activity in order to increase your organic traffic and even designed a couple of ambassadors users which answer questions on Stack Exchange; your sales people target potential customers using LinkedIn InMails, and you present your system on a convention, so that in a couple of months, somebody, on some channel, will conclude a deal. In this case, it’s better to use a Multi Touch model, the aforementioned U-shaped model, or even a Time Decay model which attributes the sale to all channels, but mostly to the channels used closer to the end of the journey (it is more effective for short sales cycles).

Attribution Model Selection Factors

As you can see in my examples above, I tend to focus on these factors:

1. The number of channels affecting the conversion
2. The time it takes a user to convert, for the specific conversion you defined. It may take another period of time or another conversion type, and therefore, it requires another model.
3. Brand strength.

Here are some examples:

• The longer the customer journey ➡️ the more suitable Multi Touch models are.
• The shorter the customer journey➡️ the more suitable Single Touch models are;
• A short customer journey with multi channels journey ➡️ a Time Decay model with an increasing weight of the channel closer to the moment of conversion.
• A long customer journey with multi channel journey ➡️ require a Position-Based model with considerable weight given to the first and last channels used (the U-shaped model).
• The less channels up have, the more suitable it is to use the Single Touch models. If there is no branding activity, it is better to start with Last Click model in order to see which channel can generate sales, and, at the same time, select compare it with a First Click model, to see which channel brings good users to the beginning of the funnel
• A strong brand with multi channel journey ➡️ a Multi Touch model. It is better to start with some Position-Based models, such as the U-shaped, and, if technically possible, a Data-Driven model. I also recommend comparing the results of these two models, since they might have the same results.

Eventually, different channels have different weight for different conversion types. Try adjusting the models to your specific conversion type and see what each channel contributes to it. This requires a careful mapping of the strengths and weaknesses of your marketing system. Some channels attract good users, but other channels convert them. Other channels generate many sign-ups for webinars, but fail to make the webinar-watchers to do anything further. And there are also channels which can make users conclude deals quickly – provided you have strong brands.

Either way, it involves intensive preparation, careful decision making and most importantly, practicing what you have learned.

Bonus: MMM – Marketing Mix Modeling

Marketing Mix Modeling (MMM) is a statistical method that uses historical sales data and marketing data to understand the relationship between marketing activities and sales performance. It aims to identify the most effective marketing strategies and tactics and determine the optimal budget allocation for different marketing activities. It typically includes the analysis of various elements of the marketing mix, such as product, price, promotion, and distribution (also known as the “4Ps”).

How do we start a MMM project?

Define the project objectives

Clearly define what you hope to achieve with the MMM project, such as identifying the most effective marketing strategies, determining the optimal budget allocation for different marketing activities, or forecasting future sales.

Gather data

Collect and organize data on sales, marketing activities, and other relevant variables. This may include data on product, price, promotion, and distribution (the “4Ps”), as well as data on the external factors that may influence sales.

Data cleaning and preparation

Clean and prepare the data for analysis, including checking for errors and missing values, and transforming the data as necessary to make it suitable for modeling.

Model development

Develop the MMM model using appropriate statistical techniques, such as regression analysis, time-series analysis, or structural equation modeling. This may involve testing different model specifications and selecting the one that best fits the data.

Model validation

Validate the model by checking its performance on out-of-sample data and comparing it to other models.

Analysis and interpretation of results

Analyze and interpret the results of the model, drawing conclusions about the relationships between marketing activities and sales performance and identifying the most effective marketing strategies and tactics.

It’s worth noting that MMM projects may require a specialized team, including statisticians, data scientists, and marketing experts, to ensure the quality and accuracy of the data, the development of the model and the interpretation of results.

About two years ago, Facebook launched Robyn, an experimental, ML-powered and semi-automated Marketing Mix Modeling open source package, and later on Google deployed their own solution called Lightweight.

Closure

That’s it.
I hope it wasn’t too tedious, and although there’s so much more to write on attribution models, I think I covered most of the basic. Good luck finding your perfect model 🙂