What Marketing Attribution Looks Like in the Post-iOS14 Era

Executives have always dreamed of a “single source of truth” to measure marketing performance and avoid wasted ad spend. Then the wakeup call came with iOS14, when Apple let users opt out of tracking. Even before the “privacy apocalypse,” brands like HelloFresh, Monday․com and Harry's (where one of the authors led the Marketing Science team) had started to explore “untrackable” channels like Billboards, Podcasts, and TV, and could no longer rely on digital tracking alone.

In the post-iOS14 era, marketing attribution is about combining methods to calibrate your model of the business, so you know what levers you can pull to drive performance and hit growth goals.

In this post we'll walk through the modern attribution stack, and how each part works together to support budget allocation decisions.

You’ve probably had this experience: you spend some time on the whiteboard mapping out your marketing funnel and growth loops, and you’re really fired up ready to grow your business, but then the ultimate question hits you like a sledgehammer: “How are we going to track this?”. (This is an age-old question in marketing: Hopkins was running split tests using mail order coupons back in 1923, a quarter century before the first published randomized control trial in medicine.)

A few years ago, the marketing industry relied heavily on digital tracking for growth – through first-touch, last-touch, or multi-touch tracking, marketers were able to tie every order and every new customer acquisition to a series of marketing “touches”, generally down to the creative level. This worked pretty well when there was less competition on Facebook and Google and DTC brands could get huge just through being diligent about CPA monitoring and in-platform experimentation. While this was never “easy”, the changes to the industry make us look back on those idyllic days fondly.

When Apple rolled out iOS14 and users opted out of tracking in droves, tracking started to seem like an impossible task. Shake off that sinking feeling – this is an opportunity!

Whenever an industry gets more complex, it opens up space for smart, motivated people to gain an edge over the competition.

We’re going to guide you on your journey through the post-apocalyptic world of marketing attribution since iOS14.It’s a journey we were lucky enough to take ourselves – pre-apocalypse – as we started to uncover the flaws in digital tracking, and the startups we worked with branched out into hard to measure channels like TV, Billboards and Podcasts.

You’re going to learn about:

The Attribution Stack: A visual map of the modern marketing landscape
The 4 Pillars of Attribution, and how to prioritize different methods
6 ways digital tracking was already broken, before iOS14
How to combine attribution methods and optimize your tools, strategy, and rituals
The return of Marketing Mix Modeling (MMM)

You’ll see that although everything is a lot messier and more complicated, there’s still a surprising amount you can do to understand what’s driving your company’s growth.

The Attribution Stack: A Visual Map of the Modern Marketing Landscape

What attribution methods are available to modern marketers, where do you begin, and what does the gold standard implementation look like? How do you decide between different methods based on their strengths and weaknesses?

Being in the media measurement space, our inboxes and DM’s have been full as marketers at modern consumer brands wake up to the need to skill up in this area. However it’s a messy, murky world full of confusing technical and statistical terms. We do our best to guide marketers on their journey, but they often don’t have the same map in their head that we had from years of experience with these methods. Some have no idea how many methods are available, or what their relative strengths and weaknesses are.

Time and time again we see people over-sold on the latest attribution tech, only to be disappointed when it isn’t the “one tool to rule them all” they were promised. They have the (false) impression that they just need to “fix” tracking and then it will be “done” (attribution, like growth, is never done). We even see this with our own clients, who are interested in marketing mix modeling but struggle to understand where it sits in the overall attribution hierarchy.

So we built the Attribution Stack as a visual map of the attribution landscape. You can use it to see what major methods companies actually use to measure the performance of marketing, and plot your own course as you navigate its complexities. We’ve ranked each in terms of sophistication, so you can start small at the bottom with something simple and reliable, then look ahead to what you need to explore as your company scales.

We’ve categorized each method into one of the four pillars of media measurement: Surveys, Tracking, Modeling, and Experiments. Most companies employ more than one of these methods, as it helps to triangulate the truth from different angles. We generally advise working from left to right as you grow. For example in the early days you can get most of the intel you need just talking to customers, and it would be a mistake to attempt running A/B tests without a lot of traffic.

These are our core beliefs:

There are many more methods than most people realize
Each attribution method has its own strengths and weaknesses
Start with one or two methods and evolve with your needs
You’ll use most of these methods eventually as you scale
Tracking is never done: build a culture of continuous improvement

For deeper attribution insight, unlock the Reforge Emerging Channels Worksheet where you can assess emerging channels that ladder up to the customer problem your product solves. Just submit your email and let us know where to send it.

The 4 Pillars of Attribution, and How to Prioritize Different Methods

Now let’s run through a broad overview of what each attribution method does, so you can understand how each of these techniques works. We’ll cover the strengths and weaknesses of each approach, so you can understand if it makes sense for the specific attribution challenges you face in your business.

Surveys

The advice in the startup community is consistently “talk to your users,” because that’s the best way early on to learn more about who you’re building for, and how they heard about you. Truly great companies keep this customer obsession up as they scale, graduating from asking customers directly to tracking their responses in panel surveys.

Sales / Support Calls: Most companies start here in the very early days, when the company is just a vision in the founder’s head. They ask customers directly how they found out about them on sales calls and in customer support emails.
Customer Interviews: Later as the company starts conducting more formal user interviews to improve the product, this question gets standardized and the information lives somewhere other than the founder or sales team’s heads.
HDYHAU (How Did You Hear About Us) Surveys: If the company is in ecommerce, it’s fairly standard practice to add a HDYHAU survey on checkout, or on signup for subscription products. This data is imperfect, because people often misremember or even lie on surveys, but it can be an important source of information for comparing and validating other attribution methods.
Panel Surveys: Later, larger companies may invest in brand recall surveys, in which a panel of people are asked questions over time, so you can track increases in brand awareness.
3rd-party Consumer Data: Finally you can purchase 3rd-party transaction level data, which shows you precisely how many customers you’re losing or gaining from competitors, both offline and online.

Tracking

Everyone starts as a last-click marketer, but nobody stays there for long. Soon you begin to notice discrepancies between your ad platform reports and analytics, and realize the need for independent measurement. For hard-to-track channels this usually means a discount code, though the gold standard is stitching user sessions together on the server with your own first-party data (for example if they’re logged in).

Tracking Pixels: If you work at a startup or a new brand, initially you trust what the ad platforms are telling you, because nobody really knows your brand other than the people seeing the ads, so any conversions you get are likely incremental.
URL Parameters: As you expand channels you move up to tracking via analytics which uses URL parameters to note which channel & campaign the user came from. Counted within this method are vanity URLs, bought as a destination for a specific campaign.
Discount Codes: If you experiment with affiliate marketing or harder to track channels like podcast advertising, you will be familiar with the humble discount code. Give somebody a code to enter on checkout (or a QR code) to tie them back to a source.
Server-Side Tracking: A more recent technological development is server-side tracking, which allows advertisers to circumvent a lot of the tracking prevention methods employed by users. This is at the leading edge of the arms race against ad blockers, as everything happens on the server where nobody can see what you’re doing with the data (morally dubious if you ask us!).
Customer Matching: Finally customer matching is the gold standard wherever you have logged in users: you can stitch their sessions together across devices based on their unique user id and truly attain a single customer view.

Modeling

Just because somebody clicked on or saw an ad, it doesn’t mean it was the ad that caused them to buy. There’s only a probability: a strong one if that was their only interaction with your brand before buying, but as you scale cases like this become rare. When you expand to multiple channels — including harder to track channels like TV, PR and Billboards — you have to start modeling conversions not simply counting them.

Device Fingerprinting: Device fingerprinting is controversial, as many users aren’t aware of all the data points their browser or phone is sharing, and how that can be stitched together to create an almost unique profile of their behavior across the web.
Behavior-Based Attribution: Behavior based approaches are growing in popularity: you might know them as ‘data-driven attribution’ (what attribution isn’t data-driven?). This technique models at the user level using machine learning to fill in the gaps in tracking.
Competitor Activity / Macro Trends: With global pandemics, economic recessions and armed conflicts dominating news cycles, marketers are starting to investigate how external forces drive performance. Simple linear regression is unreasonably effective.
Causal Inference: Many data scientists are experimenting with causal inference after a popular framework for it won the Nobel prize, but applications are limited to situations where you can emulate a controlled experiment. Regression discontinuity is used when there’s an abrupt change in policy across a continuous variable, like time or age (e.g. if over 65s are charged a difference price you can isolate the effect of price by comparing 64-and-a-half year-olds with those that are 65), or you can look for natural experiments (SuperBowl host cities are chosen as good as randomly assigned before ads are purchased, so a local uplift in sales may be attributed to ads).
Marketing Mix Modeling: Marketing Mix Modeling has been popular for over 40 years. Nowadays Econometricians have been rebranded to ‘Data Scientists’, but the general principles remain the same: collect the data you have, clean it, and attempt to predict past observations. Once you can predict the past, you can tell the future (assuming you did things the right way).

This is the pillar we fall into at Recast, as our tool uses Marketing Mix Modeling (MMM) to eliminate wasted ad spend. We have seen a resurgence of interest in MMM thanks to the fact that it’s privacy friendly (no user-level data needed) and many of its weaknesses are being addressed with the application of modern Bayesian techniques.

Experiments

Randomized Controlled Trials are at the top of the evidence hierarchy: they’re the only reliable way to control for all of the variables and establish that one thing caused another. They’re also often impossible to set up in the messy world of marketing, where we don’t control enough of the ecosystem to meet all of the requirements. A/B testing is where most marketers first encounter the scientific method, though the most consequential tests usually need to happen on the server. Where user-level testing is impossible, geo-testing can be a good solution, as well as the classic “switch it off and switch it on again” (deprivation testing).

A/B Testing: this is the one everyone starts with, as popular tools like Google Optimize, Optimizely and VWO make it easy to set up experiments within a drag-and-drop interface. Usually tests are limited to surface level changes.
Server-Side Testing: Server-side testing has become popular for channels like SEO, where these other methods don’t work, and you need to randomize the application of changes at the page template level. Server side is also required when testing changes to backend systems like recommendation algorithms.
Geo Experiments: Many channels can’t be easily A/B tested, so testing by splitting by geographical region is common. For example if you increase TV ad spend in 10 randomly chosen U.S. states, you can see the relative impact on sales.
Deprivation Testing: This is for when you can’t run a true Randomized Control Trial, but still need to be scientific. The simplest method is to switch it off and switch it on again and calculate the difference (commonly used to test the impact of branded search ads). More advanced methods include running an ad for an unrelated company (ideally a charity) to the control group and counting how many conversions would have happened anyway compared to those that were incremental.
Randomized Controlled Trials: methodologies differ but typically will involve assigning users to control and test variations, then only serving ads to the test group to see the relative difference. In ad platforms this is typically called Lift Testing.

6 Ways Digital Tracking Was Already Broken, Before iOS14

Digital tracking has been of enormous benefit to the internet economy, supporting the almost half a trillion dollars in ad spend going through digital channels, now around 60% of the market. The ability to optimize performance in real-time has allowed “the Math Men” to take over advertising from “the Mad Men,” as The New Yorker cleverly (if gender-exclusively) put it. The real winners have been small business owners, who typically couldn’t afford more sophisticated attribution techniques.

However convenient, digital tracking has always been problematic. Think about what you’re measuring – conversions from people who saw or clicked on an ad – versus what you really want to measure: people who converted because of that ad. It might sound like a subtle difference, but it can mean giving way too much credit to digital channels.

Click Windows

Click windows are usually the first complication you run into. If someone clicked on an ad 29 days ago, then visited the site 15 times through different channels before purchasing, which channel deserves the credit? The ad platform will claim 100% of it. If you’re running brand advertising campaigns, they must have some effect on your performance, but you’ll find that very rarely do they get clicks.

View Windows

What is an ad impression worth? We know that there is some value to ad impressions: afterall TV ads and billboards work despite not being able to click them. However consider the following thought experiment: you show an ad on Facebook today to every person in America. You may be surprised to learn that Facebook would claim 100% of your sales that day – even the sales you would have gotten anyway. The platform counts all conversions that occur within 7 days of a click on an ad, but also within 24 hours of someone viewing an ad (imagine how many ads you scroll past every day).

Correlation vs Causation

The view window problem is just one example of the difference between correlation and causation: Attribution is an attempt to find out how many incremental sales you got from marketing, but digital tracking can’t distinguish between what sales it generated versus what would have happened anyway. In fact optimization algorithms (like The Vickrey-Clarke-Grove auction function Facebook uses to decide what ads to place in the newsfeed) are designed to target people who are already likely to buy.

Ad Blockers & Privacy Legislation

Digital attribution is complicated enough without considering that not all digital activity is trackable. Globally 42% of internet users use ad blockers, which stop your tracking and analytics pixels from firing. Browsers like Firefox, Safari and Brave have built-in tracking prevention features: for example Safair clears all cookies every 7 days, making it difficult to track repeat visits over time. Governments are waking up to the fact that their constituents don’t like the idea of being tracked. Privacy measures like GDPR in Europe and U.S. state laws in California, Virginia and Colorado, are governing in what circumstances user data can be recorded, and how it can be used – they’re responsible for those annoying “do you accept cookies” banners that started popping up everywhere.

Multi-Channel Attribution

Have you ever looked at your Google Analytics data and found that it doesn’t match what your ad platforms are reporting? Ad platforms all report on performance differently. Google doesn’t share data with Facebook, and neither share data with your analytics platform. So you often find that multiple channels claim the same conversions. Ideally you wouldn’t trust these platforms to grade their own homework anyway, but as you scale to more channels, some independent source of truth becomes essential.

This gets exponentially harder as your company matures. As you start building a brand and seeing increasing word of mouth spread, it gets less likely that the person you showed the ad to hasn’t already heard of you. Once digital channels get saturated and you hit diminishing returns, eventually you need to start mastering harder to measure channels like TV, billboards and radio. These channels will show up as ‘direct’ or ‘organic’ traffic in your web analytics: how do you tell which offline channel is working?

That’s assuming all of your sales occur on your website: what if you sell on Amazon, or via a registry, or via a physical retail partner like Walmart? You won’t get as much data from them as you do on your own website. If you have physical retail stores, how do you track users who saw an ad online and then purchased in store? Or walked past your storefront which reminded them to search for your brand term and buy online. Does your Google brand campaign really deserve 100% of the credit, or is this good justification for paying expensive commercial rents? It’s impossible to make informed budget allocation decisions without getting answers to these questions.

What Broke with the iOS14 Rollout

In June 2020 Apple released iOS14, which gave users the ability to opt out of tracking within apps. If a marketer advertising within an app wants access to the user’s IDFA, the unique identifier for the user’s device, the app must first ask permission – and only 80% of users opt out. With iOS commanding about half of the U.S. market, advertisers suddenly had a big hole in their digital tracking.

The IDFA was used by analytics tools and ad platforms to match users who take actions within the app back to the individual ad they clicked. This allowed advertisers to measure the performance of their campaigns and justify billions of dollars of ad spend, particularly in the lucrative mobile gaming sector.

Even non-app companies were impacted, as the Facebook and Instagram apps were the source of a significant amount of digital ad inventory. If a user opts out of tracking, Facebook can no longer tie their data back to the ads they clicked on, hurting targeting and tracking of performance: in aggregate costing Facebook an estimated $10 billion in ad revenue. In a telling move, Facebook quickly deprecated its industry-leading Lift testing functionality as it could no longer guarantee assignment of users to test and control groups with no unique device ID.

Vendors rapidly began innovating to replace the IDFA with more sophisticated methods:

Device Fingerprinting: building a profile from unique attributes of each device when clicking an ad, then matching it to app users on the other side
Customer Matching: linking customer data across apps by PII (Personally Identifiable Information) like email, IP, or physical addresses.
Server-Side Tracking: matching marketing and conversion data on your own server rather than within the app code, so Apple can’t see what you’re doing

Mobile attribution providers are now in the unenviable position of being in an arms race with the world’s most valuable company, who’s sitting on $265 billion in cash and is aggressively positioning iPhone as the privacy-friendly option. Apple has since followed up with statements expressly forbidding customer matching and device fingerprinting, and have developed features like private relay, which make stitching together user-level data increasingly impossible. Even Google, who relies on advertising unlike Apple, has seen the writing on the wall: they plan to deprecate the use of cookies in Chrome by 2023, as well as roll out other changes as part of its Privacy Sandbox initiative.

Let’s take another look at the attribution stack, but with the attribution methods afflicted by iOS14 shaded in red so we can see what was lost.

The Attribution Stack: After iOS14, these methods are no longer viable: customer matching, server-side tracking, tracking pixels, device fingerprinting, and randomized control trials — Source: “The Attribution Stack”, recast.com

The remarkable thing is that so many methods are still available to us! Most of the pillars are still intact, and most of the important methods that have been in use for decades still work. It’s important to note that digital tracking isn’t completely dead, it still works on non-iOS platforms and for iOS users who opt in. However gone are the days where you can solely rely on it.

At Recast we’ve seen interest in marketing mix modeling (MMM) dramatically increase, and Facebook has stepped into the space, generating huge interest in Robyn, their open-source MMM tool. Vendors who relied more heavily on user-level data have been seeing early signs of hope with behavior-based attribution, but across the board we’re seeing marketing organizations experimenting with many more techniques than they used to, which is a good thing! Accessibility is the biggest challenge, as most of these techniques still require a data science team to implement and interpret, but the next generation of martech vendors is being born to solve these issues.

How to Combine Attribution Methods and Optimize Your Tools, Strategy, and Rituals

With so many competing attribution methods, how do companies combine their insights to make decisions? For example if Google is telling you that your search ads drive 50% of sales, but your marketing mix model says it’s 30%, and analytics says 40%, which do you believe?

Here’s where you can start optimizing your stack holistically:

Commit to Different Methods for Different Levels of Decision Making

For example, trust Google and Facebook on a daily basis when they tell you a new campaign is performing better than another, and use them to make optimization decisions when testing ad creative. Marketing mix models are unlikely to give insights on such a granular level, and you can’t justify setting up a formal experiment for every small change. However when it comes time to set monthly or quarterly budgets, bias towards more scientific methods.

How do you factor in incrementality findings? It’s simpler in practice than you may think. If you trust your model and it says Google drove 600 sales compared to the 1,000 that Google claimed, then that channel is 60% incremental (40% or 400 of the sales would have happened anyway). When you calculate performance metrics, simply multiply your conversions or revenue by 0.6 to be directionally right when optimizing. If you spent $100 to get 10 conversions today, keep in mind that Google deserves credit for only 6 of them (10 x 0.6 = 6). So your cost per acquisition is in reality closer to $16 ($100 / 6 = $16) not $10. This can save you from doubling down on a campaign that actually won’t be worth it when you run the numbers again at the end of the month.

Use Each Method to Calibrate the Others and Triangulate the “Truth”

For example if analytics significantly disagrees with your marketing mix model, that’s important information for your data science team to test the model against, to make sure they’re building a model that’s plausible. Or go the other way: if you find that a significant proportion of your search ads conversions come from your brand term, that could explain why your marketing mix model is identifying it as not incremental: time to set up a Lift Test or other experiment.

Remember Marketing Attribution Isn’t About What to Spend, it’s About What to Do

If your attribution techniques aren’t feeding information back into your strategy, that disconnect will hurt your performance. Every piece of analysis and report produced should be teaching you something you didn’t know about your customer and their typical behavior. Survey responses should give you ideas for more experiments, surprising ad platform performance should inform your testing schedule, test results should be used to calibrate your marketing mix model.

Build Consistent Rituals Around Your Attribution Methods

You’ll have to figure out the right checkpoints for your own team. But here’s what a typical daily, weekly, monthly and quarterly schedule might look like in companies that are at the top of the marketing attribution game:

Daily Optimization

Check ad platform performance and make small changes
Look at analytics dashboards to validate ad performance
Investigate severe anomalies or unexpected results

Weekly Tactics

Make pause / play decisions on campaigns
Check KPIs against performance goals
Ad hoc analysis into unexpected results

Monthly Strategy

Plan testing schedule and pause poor performers
Deep dive analysis into bigger strategic questions
Share key insights with team and democratize data

Quarterly Resourcing

Strategic review of what worked (or didn’t) and what’s next
Forecasting and planning for next quarter or year
Set budgets & KPIs based on progress and opportunity
Budget optimization across core and emerging channels
Calibrate your model to ensure consistency in findings

The Return of Marketing Mix Modeling (MMM)

Marketing mix modeling has the potential to tie all of these methods together and offer more real-time granular insights than were possible in the past:

MMM is the only method that works across every channel at scale
Bayesian MMM lets you calibrate your model with other attribution methods
The challenge is to make it real time and granular enough

In the last few years media mix modeling has been making a comeback. This has been driven first by the shift to a truly “omni-channel” world where most brands are utilizing both online and offline marketing channels and a blend of trackable and un-trackable distribution channels, and then subsequently by the rise of ad blockers and privacy regulation and policies. As stated in Privacy-Centric Digital Advertising: Implications for Research (Johnson, Runge, Seufert 2022): “When advertisers lose user-level data… [they] must then fall back on macro-level ad measurement: aggregating by time, market, cohort, and/or channel… Advertisers will therefore return to media mix modeling”.

We’ve been building Recast as a next-generation MMM tool that has all of the advantages of traditional MMM while still being useful for modern performance-driven marketers. We think MMM is an incredibly powerful tool to have in the toolbox, but we recognize that MMM is really only one part of the puzzle.

Marketing mix modeling has some important limitations:

It can take months to build a model and cost tens of thousands of dollars
It requires complex statistical techniques, which are vulnerable to human error and bias
It doesn’t deliver the same campaign / creative level granularity as digital tracking

We’re attempting to solve these problems at Recast, and teach people how to solve them at Vexpower, but we’re not going to lie, they’re difficult problems to solve. Modern data pipelines and Bayesian methods are already a huge leap forward, so models can be built and refreshed automatically in real-time, but we’ve got a long way to go. Even when we solve some of these problems, attribution won’t be solved: we’ll never have a single source of truth.

The truth is we’re in a post-truth world, where more than one opinion is valid. You should never trust just one single source of information, without validating what it's telling you, and thinking about the potential motivations and incentives behind that conclusion.

It’s important to have multiple weapons in your arsenal, and know how to use them together to see through the fog of war. We all have to be marketing scientists now: we have no choice.

Organizations that navigate this transition will have an edge against those that don’t. Chaos is a ladder.

This Emerging Channel Worksheet can help you measure the value of investing in emerging channels as they relate to your customer problem and the larger goals your company sets. For more resources on attribution and overall marketing strategy, apply today!