9,362 reads

The Death of App Attribution

by AlexApril 16th, 2019

Too Long; Didn't Read

This is the second chapter in the story of Branch, and describes how we built the next-gen solution for mobile attribution. The underlying problem is that the mobile attribution platforms we use today are chips off the same old block. We need a new toolbox to rebuild attribution in a way that can keep up with our rapidly-changing digital landscape. We’re going to explore Branch’s perspective on an ideal attribution 2.0 solution in five chapters: What does “attribution” even mean? How a persona graph provides reliable and accurate measurement everywhere.

Companies Mentioned

featured image - The Death of App Attribution

This is the second chapter in the story of Branch (we published Deep Linking is Not Enough, covering our rise to become the industry’s leading deep link and user experience platform, two years ago).

Attribution is deceptively difficult. Every mobile marketer considers it critical, and yet many people still feel there is vast opportunity for improvement.

How is this possible? For something that is so important to the modern marketing ecosystem, why are so many of the available options for attribution so disappointing?

The underlying problem is that the mobile attribution platforms we use today are chips off the same old block: technologies that were designed to passively measure a single channel or platform. This is a perfect example of “if your only tool is a hammer, everything looks like a nail.” We need a new toolbox to rebuild attribution in a way that can keep up with our rapidly-changing digital landscape.

The potential of “Attribution 2.0” is enormous. Done well, it is a strategic growth engine that actively helps you grow your business. However, the risks of getting it wrong are just as high: legacy attribution will strangle your growth with broken experiences, misleading data, and inaccurate decisions.

Today, we’re going to explore Branch’s perspective on an ideal Attribution 2.0 solution in five chapters:

What does “attribution” even mean? A brief history of marketing attribution, including offline, digital, and mobile.
How mobile attribution providers became blind. The reason why these platforms are rapidly losing the ability to do their job.
The future of attribution. How a “persona graph” provides reliable and accurate measurement everywhere.
Reviewing traditional attribution techniques. A deep dive into how measurement worked in the single-platform worlds of websites and apps.
The next generation: a persona graph. Why a persona graph works, and how we built one at Branch.

Chapter 1: What does “attribution” even mean?

This chapter is a tour of marketing attribution up until today, covering offline, digital, and the birth of the mobile attribution provider as a separate service. We will review the basic needs of an attribution solution, the evolution of attribution as channels have fragmented, and specific mobile challenges.

If you check the dictionary entry for “attribution,” you’ll find this:

“The action of regarding something as being caused by a person or thing.”

At the most basic level, every marketing attribution system in the world performs three tasks: 1) capture interactions between the user and the brand, 2) count conversions by the user, and 3) link those conversions back to any interactions that — in theory — drove them. When done correctly, this process allows you to figure out if your campaigns are worth the cost.

As we’ll discuss, there’s undeniably now a fourth task: 4) protect against broken user journeys. Attribution is an exercise in futility if anything blocks conversions from actually happening in the first place.

While this sounds straight-forward, it leads to an inherent conflict: tracking every activity and what caused it might sound like paradise to a marketer, but the proliferation of ad blockers, browsers with built-in tracking protections, and new privacy-focused legislation around the world clearly shows that many end users don’t share this perspective.

Attribution has existed for as long as marketing itself, and the techniques involved have changed dramatically over time. As we think about the future of attribution, it’s important to recognize that Attribution 2.0 doesn’t mean “x-ray vision to track everything.” What we need is responsible, secure, privacy-focused measurement that can reliably handle the technical challenges of a complicated digital ecosystem.

Attribution version 0: offline

Billboards, TV commercials, newspapers, and other mass-market campaigns all share one thing in common: everyone who sees them gets the same experience. These campaigns are not individualized, and they are not interactive.

This is a problem for accurate attribution, because it means there is no way to deterministically measure the relative influence of each activity. You could use proxy metrics (e.g., “how many people walked past this billboard last week?”), or try to infer attribution data (e.g., “did sales in a given city increase after my TV commercial?”). You could even try to tease out a few extra insights with workarounds like special discount codes or unique phone numbers for each campaign. But all of these techniques are analog and imprecise.

Attribution version 1: digital

Imagine discovering the electric lightbulb, after a lifetime of candles. That’s what happened to attribution when the digital ecosystem burst to life.

Imagine discovering the electric lightbulb, after a lifetime of candles. That’s what happened to attribution when the digital ecosystem burst to life. New technologies like hyperlinks and cookies made it possible for digital marketers to measure exactly which users encountered a marketing campaign, when and how they interacted with it, and what they did afterwards. Because insights like these are table stakes for attribution today, it’s hard to remember just how big of a breakthrough they were at the time.

In those early days, user journeys were confined to just a single place: the web, on a computer. This was a good thing, because measurement of single-platform customer journeys is a relatively manageable problem. Each marketing channel is responsible for its own attribution: email service providers measure email, ad networks track their ads, and so on. This worked because all channels still led directly back to a website, allowing marketers to string together a conversion funnel that went right down to events representing value (like sign ups or purchases).

The birth of the Mobile Attribution Provider

But then, in 2008, Steve Jobs opened Pandora’s Box by introducing the world to a brand new platform: native mobile apps. In those early days, many mobile marketers (especially in the gaming industry) found that using ads to drive app installs was a sure-fire path to positive ROI. So much so, that other channels and conversions were allowed to fall by the wayside because solving the technical complexity just wasn’t worth the investment.

However, ad install attribution comes with a few significant technical problems of its own: 1) matching, and 2) double attribution.

Matching. The iOS App Store and Android Play Store are attribution black holes. Between the ad click that takes a potential user to download and that user’s first app launch, marketers are completely in the dark. Since the basic definition of attribution is knowing where new users came from, it is critical to find a way around these black holes, in order to connect installs back to clicks that happened earlier.

Double attribution. With so many ad networks all vying for the same eyeballs, users often interact with multiple ads before successfully installing an app. No marketer likes being charged twice for the same thing, but this is exactly what happens when two different networks make claims for driving the same app install.

To solve these problems, a new type of company appeared: the Mobile Attribution Provider. Using a combination of device IDs and a probabilistic technique known as “fingerprinting” (which slurps up device data like model number, IP address, and OS version to create a signature that may or may not actually be unique), these companies provided “matching magic” to figure out which ad a new user had clicked prior to install.

By centralizing all this conversion data in one place, the mobile attribution providers were able to act as independent advocates on behalf of the marketer, ensuring the right ad network got paid (and only paid once).

Chapter 2: How mobile attribution providers became blind

This chapter discusses why mobile attribution providers are losing relevance in our multi-platform world, how this affects the companies that rely on them, and why user experience is now a critical piece of the attribution puzzle.

In the early years of mobile, getting a user to install an app was all that really mattered. Once users had your icon on their home screens, you’d won. And because the ROI of buying app installs with ads was reliably positive, there was no real need to invest in things like cross-channel acquisition or cross-platform re-engagement.

The DNA of these companies is so tied up in apps and ads that the words “app” and “ad” themselves often show up as part of the company name.

All of the traditional mobile attribution providers on the market today were born during this phase of simple and easy paid growth, which means they all suffer from the same foundational problems:

They’re black-and-white TVs in a Technicolor world. The mobile ecosystem has expanded, but the DNA of these companies is so tied up in apps and ads that the words “app” and “ad” themselves often show up as part of the company name (savvy teams recognized this shift years ago and made investments in rebranding).

They’re passive, third-party bean counters. Because these systems grew out of a single-platform, single-channel mindset, they are designed and built on top of an assumption that the only thing they ever needed to do was stand by and observe. They’re like bureaucrats who only care about one outcome (app installs), and only deal in one currency (mobile ads). The rest of the world does not matter. In our new, multi-platform reality, passive observation is no longer enough.

The ironic result of these issues is that legacy systems increasingly fail to deliver the one thing they were built to provide: accurate measurement.

Missed attribution already leads to real costs

Corrupted data and broken customer experiences can do measurable damage to digital businesses. Here’s a realistic possibility:

You want to buy a new pair of shoes.

Scenario 1: how it “should” work. While scrolling through Facebook, you see an app install ad for discounted shoes, download the app, and then proceed to purchase a pair of sneakers. Everything is fine, because this is the basic app install ad working the way it was designed to.

Scenario 2: how things actually happen in real life. While waiting in line at Starbucks, you start by searching the web for shoes. You’re a regular at this Starbucks, so your phone has automatically connected to the wifi. You see an app install ad and click it, but before the download can finish, your order is called and you walk out of the store without opening the app. You remember about the shoes later that evening and complete the purchase at your computer. Meanwhile, another Starbucks customer opens the same app a few hours later to buy a new hat.

Here’s where things get messy: because your ad click happened on the web, a traditional mobile attribution provider would be forced to use fingerprinting to match your install. And because both you and the unknown other customer have the same iPhone model and were using the same Starbucks wifi network, your device fingerprints will be identical. From the attribution provider’s perspective, a single user clicked the ad, opened the app, and purchased the hat. The web conversion, which was actually driven by the ad, gets tracked as a completely separate customer (if it is even captured at all).

While this particular example is an edge case, that’s the whole point: edge cases are no longer the exception to the rule — they are the rule. Businesses that aren’t equipped to handle this are pouring most of their attribution data down the drain without even realizing it.

Attribution and user experience are two sides of the same coin

In order to attribute a conversion, that conversion has to happen in the first place. On the web, single-platform user journeys were robust and relatively unlikely to break, but “The Internet” is no longer a synonym for “websites on computers.”

“The Internet” is no longer a synonym for “websites on computers.”

Today, if you aren’t able to provide the sort of seamless experience your customers want, the cost can be far more than just a lower conversion rate:

body[data-twttr-rendered="true"] {background-color: transparent;}.twitter-tweet {margin: auto !important;}

I just cancelled my @WSJ subscription after being a reader and subscriber for 39 years. Because they don't recognize me across my phone or tablet and force me to log in every time. The app works but links from Twitter or elsewhere don't open the app or deep link. Bye Bye.

— @chrisfralic

The explosion of new channels, platforms, and devices has fractured the digital ecosystem, but users don’t understand (or care) that this fragmentation causes technical headaches for you. They expect seamless experiences that work wherever they interact with your brand.

This means that these days, attribution is not just a marketer’s problem; it impacts every part of the business, and companies that provide attribution as a service need to take a far more active role than simply sitting on the sidelines, counting beans. Legacy systems that haven’t evolved fast enough are already a major business risk because of potentially missing data, and are silently becoming more of a liability over time as they fail to help improve the business metrics they are supposed to measure.

Chapter 3: The future of attribution

This chapter explores the new industry trend toward “people-based attribution” before introducing a truly comprehensive solution: a persona graph.

The digital ecosystem is quickly approaching a breaking point. For example, want to run an email campaign to drive in-app purchases? You’re out of luck with traditional mobile attribution providers; they’re from an older generation that can’t measure email. How about a QR-code campaign in an airport where everyone is sharing public wifi? The ambiguity of fingerprint matching — the only legacy methodology they can use to attribute a user journey like this — will kill you.

The future of mobile attribution isn’t just about apps and ad-driven installs; in fact, it isn’t even just about measurement.

The people-based attribution trend

A number of mobile attribution providers have recently begun jumping on the bandwagon of “people-based attribution.” In plain English, this means expanding scope to consolidate all the interactions and conversions of each user, regardless of where those activities occur.

This is a significant improvement — at least it shows the industry is beginning to acknowledge the problem! — but the devil is in the details: these “people-based” solutions aren’t all created equal, and most of them share the same critical flaws: they still rely on inaccurate matching methods, and they’re only built to provide passive measurement.

In other words, these systems may call themselves “people-based,” but it’s more like lipstick on a pig. The future of mobile attribution isn’t just about apps and ad-driven installs; in fact, it isn’t even just about measurement. Any system built on top of these two assumptions is fundamentally unsuited to the realities of the modern digital world.

The foundation of Attribution 2.0: a persona graph

Fragmentation isn’t a new problem for attribution. Even in the good old days of desktop web, a user might have two different web browsers installed. Or multiple computers. Or they might be using a shared computer at the public library. But this sort of fragmentation was a minor thing that could be filed away with all the other small, discrepancy-causing unmentionables (like incognito browsing mode) that are rarely worth the effort for marketers to address.

Things are different now. Like the frog that doesn’t realize it’s in a pot heating on the stove until it’s too late, fragmentation across channels, platforms, and devices is about to reach the boiling point. This is a data-sucking monster that costs customer loyalty and real money. No serious company can afford to ignore it.

The problem is that traditional attribution methodologies (things like device IDs and web cookies) are siloed inside individual ecosystem fragments. Existing attribution systems see each of these channel/platform/device fragments in isolation, as disconnected and meaningless points.

To fix this problem, what we need now is to zoom way, waaay out. We need a system that lives on top of all this fragmentation, stitching the splintered identity of each actual human customer back together into a cohesive whole, across channels and platforms and devices.

What we need is a Persona Graph. A shared, privacy-focused, public utility that serves the identity needs of everyone in the ecosystem.

This sort of collaboration is hardly a new idea (just think of any service that provides salary comparisons by aggregating the data submitted by individual users), but it has never before been applied as a solution to the challenge of accurate attribution.

Part 2: Building Attribution 2.0

The world of attribution is full of gnarly problems with no single correct solution: things like attribution windows (e.g., “is my ad really responsible for purchases that happened six weeks later?”) and attribution models (e.g., “how do I decide which interactions deserve credit when there are more than one?”) and incrementality (e.g., “did my ad campaign cause the customer to purchase, or would they have done it anyway?”). These lead to difficult questions for any system.

However, before we can even begin to discuss more sophisticated topics like these, the three basics have to be solid: capturing user <> brand interactions, counting user conversions, and linking interactions back to the conversions that drove them. In today’s fragmented digital ecosystem, it’s no longer safe to take that for granted.

In many cases, mobile attribution providers still rely on matching techniques that are essentially semi-educated guessing.

Here’s why traditional mobile attribution solutions fall short in all three areas:

They miss a lot of interactions. Attribution 2.0 needs to catch activity for every kind of campaign (whether owned, earned, or paid), which means reliably covering every inbound channel. Unfortunately, mobile attribution providers are still living in a world where ads are the only channel in town.

They also miss a lot of conversions. Attribution 2.0 needs to catch conversions everywhere businesses have a presence. Users download mobile apps, but they also convert on websites, inside desktop apps, on smart TVs, in stores, and more. Mobile attribution providers still treat all of these other platforms as second-class citizens…if they’re even covered at all.

They’re not very good at linking interactions and conversions. Attribution 2.0 needs to understand the connection between activity (cause) and conversion (effect), otherwise the only result is a mess of isolated event data. In many cases, mobile attribution providers still rely on matching techniques that are essentially semi-educated guessing.

Chapter 4: A review of traditional attribution techniques

This chapter describes the methods used to provide single-channel attribution for websites and apps — the same methods that are now falling short in a multi-platform world.

The ultimate solution to these problems is a persona graph. But before we get into the details of how it works, let’s revisit the world as it exists today; many of these techniques are still important pieces of the persona graph solution, even if they are no longer enough when used alone.

Traditional attribution techniques for the web

On the web, a variety of techniques make attribution possible, including URL decoration, the HTTP referer (yes, it really is spelled that way in the official specification), and cookies.

**URL decoration**Everyone who has ever clicked a shortened URL (e.g., https://branch.app.link/jsHNKjzIeU) or wondered why the address of the blog post they’re reading has an alphabet soup of nonsense words at the end (utm_channel, mkt_tok, etc.) is already familiar with this technique. URL decoration is simplistic and often requires manual effort, but it has survived because it just works: encoding attribution data directly into a link the visitor will click anyway is a robust and surefire way to make sure it gets passed along. This is why you’ll often encounter URL decoration in mission-critical attribution situations where durability is key, such as search ads or links in an email campaign.

**The HTTP referer**When you click a link, your browser often tells the server where you were right before you clicked. This technique has a number of limitations that make it less robust than URL decoration (notably that it can be faked or manipulated by users, and the origin website can intentionally block it), but the biggest advantage for attribution is that it’s automatic. This makes the HTTP referer a popular choice for “nice-to-have” measurement, like tracking which social media sites send you the most traffic.

**Cookies for basic identification**Techniques like URL decoration and the HTTP referer let you determine how a visitor arrived on your website, but they disappear after that initial pageview. This makes it impossible to rely on either of them alone for attributing conversions back to campaign interactions. Fortunately, there is a solution for this: cookies.

Today, even casual internet users know what cookies are: little pieces of data that browsers remember on behalf of websites. They have many uses, but one of the most common (and the most important for attribution) is storing a unique, anonymized ID. These IDs don’t contain any sensitive info, but the effect is much like sticking a name tag on each visitor: they make it possible to recognize every request by a given browser — including down-funnel conversions like purchases — and attribute them back to the original marketing campaign.

**Advanced cookies**Pretty much every web-based measurement or analytics tool on the market today uses cookies to some degree, and basic identification was just the beginning — cookies have long been used for other, more sophisticated purposes. One of the most common is cross-site attribution, which works like this:

For obvious security and privacy reasons, browsers restrict how cookies can be set and retrieved. After all, no one would be happy if Coca-Cola had the power to mess with Pepsi’s cookies. To prevent this, cookies are scoped to individual domains, and web browsers only give cookie permissions to domains that are involved in serving the website. This means that unless pepsi.com tries to load a file from coke.com, Pepsi’s cookies are secured against anything devious taking place [attempts to defeat these protections are part of a large infosec topic known as “cross-site scripting attacks,” or XSS for short].

Cookie security is a necessary and good thing, so the web ecosystem has figured out a number of creative ways to perform cross-site attribution within these limitations. For example, if Pepsi wants to run ads on both www.beverage-reviews.net and www.cola-lovers.org, then everyone agrees to allow a neutral third-party domain (in the world of web-only attribution, often owned by an ad network) to place a cookie that is accessible across all three of these websites. The end result is that the third-party ad network can recognize the same user across every site involved, and leverage that data to provide attribution for their ads. To help increase coverage, it’s even become standard industry practice for these third-parties to share their tracking cookies with each other (a process called “cookie syncing”).

However, the tide is starting to turn against cookie-based attribution networks. Due in part to end-user outrage triggered by “creepy ads,” major web browsers have implemented restrictions on cookies: ITP on Safari, ETP on Firefox, and even Chrome is reported to be working on something similar. Third-party ad blockers and privacy extensions pick up where the built-in functionality stops, and new privacy-focused legislation around the world (such as GDPR) continues to restrict what companies can implement.

Traditional attribution techniques for apps

Mobile attribution providers rely on two techniques for matching installs back to ad touchpoints: device IDs, and fingerprinting.

**Device IDs**Every mobile device has a unique, permanent hardware ID. In the early days, it was common practice for app developers (including, by extension, attribution providers and ad networks) to access these hardware IDs directly, and one of the common use cases was ad attribution.

However, while “unique” is a good thing for attribution accuracy, “permanent” leads to obvious privacy concerns. Apple recognized this in 2012, and closed off developer access to these root-level hardware IDs. As a replacement, app developers got the IDFA (ID For Advertisers) on iOS. Google quickly followed with the GAID (Google Advertising ID) on Android. The IDFA and GAID are still unique to each device, making them a good solution for attribution, but give additional privacy controls to the end-user, such as the ability to limit access to the ID (“Limit Ad Tracking”) or reset the ID at any time, much like clearing cookies on the web.

Device IDs are a “deterministic” matching method. This means there is no chance of incorrect matching, because the device ID on the install either matches the device ID on the ad touchpoint…or it doesn’t. No ambiguity. Because of this guaranteed accuracy, device IDs remain the attribution matching technique of choice, whenever they are available.

Unfortunately, device IDs are not always available. This issue crops up in many situations, but here’s the big one: device IDs are off-limits to websites. This makes them a single-platform matching technique — they only work for attribution when the user is coming from an ad that was shown inside another native app.

This left the mobile ecosystem with a problem: since device IDs are siloed inside apps, and cookies are equally limited to just the web, how to bridge the gap and perform attribution when a touchpoint happens on one platform and a conversion happens on the other?

**Fingerprinting**To solve this problem, the mobile attribution industry turned to a technique known as “fingerprinting”. While fingerprinting had long existed as a niche solution on the web (often used to help fight fraud), app attribution took it mainstream.

By now, most marketers — and even many savvy consumers — are familiar with how fingerprinting works: various pieces of data about the device (model number, OS version, screen resolution, IP address, etc.) are combined into a distinctive digital signature, or “fingerprint.” By collecting the same data on both web (when the ad or link is clicked) and app (after install), the attribution provider is theoretically able to identify an individual user in both places.

While this solves the immediate challenge of tracking a user from one platform to another, there are two important catches:

Fingerprinting is a “probabilistic” matching method. No matter how confident you may be that two fingerprints are from the same user, there’s always a chance that you’re wrong. There’s always an element of guesswork involved.

Fingerprints go stale. Much of the data used to generate fingerprints can change without warning, which means they begin going stale as soon as they’re created. This degradation is exponential, and most mobile attribution providers consider a fingerprint-based match to be worthless after 24 hours.

In the early days of app attribution, most marketers saw the ambiguity inherent in fingerprinting as a manageable risk (and it was certainly better than the alternative, which was no attribution at all). However, this ambiguity has become harder and harder to ignore over time: today, there are simply too many people with the latest iPhone and the most recent version of iOS, all downloading apps via the same AT&T cell phone tower in San Francisco.

Chapter 5: The next generation: a persona graph

This chapter explains how a persona graph works, addresses common concerns around user privacy and data security, and goes in depth on how we built Branch’s persona graph. It ends by comparing the older generation of mobile attribution providers with what is possible with a persona graph.

The problem with traditional attribution techniques is they are either probabilistic (meaning there’s a chance the data is wrong), or siloed inside a single platform (web or app). A persona graph provides the best of both worlds.

Imagine the game of Concentration (for those who haven’t played this in a few years, it’s the one where you flip two random cards over, hoping to find a match). The chances of discovering a pair on your first turn are extremely low, but over time (and time is the critical element here), you learn where everything is. Eventually, assuming you have a good memory, you’re uncovering matches on almost every round.

Now, let’s take the metaphor one step further: instead of you flipping cards to learn where they are, imagine a hypothetical situation where you get to join a game in progress, where every card on the table has already been turned face up by other players before your first turn. It wouldn’t be much of a game, but you’d be guaranteed to find a match every time.

Like a Concentration game where all the cards have already been flipped before your first turn, a persona graph allows you to accurately match users that YOU haven’t seen before, but someone else in the network has.

That’s the concept behind a persona graph: by sharing matches between anonymous data points, everyone wins. Like a Concentration game where all the cards have already been flipped before your first turn, a persona graph allows you to accurately match users that YOU haven’t seen before, but someone else in the network has.

The elephants in the room: privacy, security, and confidentiality.

For a persona graph to survive, there are a couple of critical things that must be guaranteed: 1) privacy and security of user data, and 2) confidentiality.

User privacy and data security. A persona graph makes it possible to recognize a given user in different places, but it does not tell you anything about WHO that user is. If the user wants you to know that information, then you already have it in your own system — the persona graph simply closes the loop by telling you that you’re seeing an existing customer in a new place. And like cookies or device IDs, the user can reset their connection to the persona graph on demand.

In other words, the persona graph must take the same approach to privacy as the postal service. Our letter carriers need to know our physical location in order to deliver mail, but they’re only concerned with the address, not the addressee. We trust that they won’t open our letters and won’t sell information about what we buy to the highest bidder.

At Branch, we feel so strongly about user privacy that we have made a number of public commitments about it. The short version can be expressed as three points in plain English: 1) we proactively limit the data we collect to only what is absolutely necessary to power the service that we deliver to our customers, 2) we will only ever provide our customers with data about end-user activity that happens on their own apps or websites, and 3) we do not rent or sell end-user personal data, period (not as targeting audiences to other Branch customers, not via cookie-syncing side deals with identity companies, not via an “independent” subsidiary — we just don’t do it).

In addition, we rigorously and proactively follow best practices to purge sensitive data and protect our platform against bad actors.

Confidentiality. The only data that is available via a persona graph is knowledge of the connection itself. Not where or how the connection was made, or by which company’s end user. A persona graph must guarantee that it will never allow Pepsi to purchase a list of Coke’s customers.

Said another way, the Swiss have avoided every war in Europe for over 500 years, because everyone recognizes that they are (and always will be) neutral. A persona graph must maintain the same unimpeachable reputation.

A peek inside the Branch persona graph

When we set out to build Branch in 2014, there was already a well-established industry of mobile attribution providers. All of them were competing with each other for the low-hanging fruit of measuring ad-driven app installs. If you work in the mobile industry, you’re likely familiar with their names already (Branch acquired the attribution business of one last year).

Even though the Branch platform might resemble a traditional attribution provider on the surface, the engine underneath is something fundamentally, radically different.

We decided to take a different approach: we realized the app install ad was a bubble that would eventually deflate, and we also knew that seamless user experiences would become increasingly important as marketers began to care about other channels and conversion events again. So we started by solving the more difficult technical problems that everyone else was ignoring (this is the story we told two years ago in Deep Linking is Not Enough).

The result: through solving the cross-platform user experience problem at scale, for many of the best-known brands in the world, we created a persona graph that allows Branch to provide an attribution solution that is both more accurate and more reliable than anything else available.

Here’s how it works today:

Step 1: Collect deterministic IDs

Believe it or not, this is actually the relatively easy part. User activity occurs in fragments across platforms, and the goal is to have a deterministic ID for each of them. Since Branch’s customers invest most of their marketing resources into websites and mobile apps, these are the platforms where we’ve focused the majority of our effort so far. But the same principle applies anywhere.

To create deterministic IDs on the web, we use a javascript SDK to set first-party cookies. Inside apps, we offer native SDKs to leverage device IDs.

We’ve also built SDKs for desktop apps on macOS and Windows, and custom OTT (Over The Top) device integrations. We will continue adding support for new platforms as customers request them.

Step 2: Create persona matches

Once we have an ID for an identity fragment, we use a layered system of cross-platform matching techniques to tie it back to a persona record on the persona graph. Here are a few examples:

Deep links. When a user clicks a link to go from one place to another, that is an ideal time to make a connection. This is our primary method for matching fragments that exist on the same device (e.g., Safari, Facebook browser, native apps), and one of the most reliable because it’s driven by the user’s own activity.
User IDs. When a user logs into an account, they’re providing a unique ID that can then be matched if the same user signs in later in another place. We only use this signal to a limited extent today, because there are a number of tricky problems related to shared devices, but we’re actively working on solutions and see a lot of promise in this method. As a side note, this is the only matching method we’ve seen competitors use when they talk about “people-based attribution.” Given the shared device challenges mentioned above, or the fact that (depending on the vertical) the vast majority of visitors never log in, this is certainly an area to question if you’re currently working with one of them.
Google Play referrer. Google passes a limited amount of data through the Play Store during the first install. Branch uses this one-time connection to create a permanent match back to the persona graph.
Fingerprinting. This is one cross-platform matching method we don’t use to build the persona graph, but it deserves a mention because it is so commonplace in the attribution industry. Branch sometimes has to fall back on fingerprinting when the persona graph can’t provide a stronger pre-existing match, so we’ve invested in an IPv6-based engine that greatly increases accuracy over traditional mobile attribution providers that still rely exclusively on IPv4.

Because of Branch’s massive, worldwide scale, we can also use machine learning to uncover connections between different personas that likely belong to the same user, and just haven’t yet been deterministically merged. We call these “probabilistic matches” because they’re not 100% guaranteed on each end, but they’re still useful and helpful when combined with the high degree of confidence that we get from observing other deterministic patterns.

Here’s how probabilistic matching compares to fingerprinting:

Fingerprinting. Fingerprinting has to happen in real time. In other words, it requires a guess to be made based solely on whatever data is available at the exact moment a user does something. That user might be sitting alone at home (high accuracy situation), or they might be sharing public wifi with several thousand other people while walking around a shopping mall (very low accuracy situation). With fingerprinting, the system has only two choices: 1) it can take a gamble and make the match, or 2) it can throw away the match and say no attribution happened. All of the fancy “dynamic fingerprinting” systems offered by traditional mobile attribution providers are really just trying to decide when to choose option 2.

Probabilistic matching. Because the persona graph is persistent, Branch can afford to be patient. We don’t have to play roulette in real time when the conversion event occurs; instead, we’re able to preemptively store “prob-matches” when the system detects no ambiguity (e.g., when the user is alone at home) to use later (e.g., when the user is inside a crowded shopping mall). For example, the algorithm might create a prob-match if it notices that persona A and persona B have matching fingerprints, were both active on the same IP within 60 seconds of each other, and no other activity occurred from that IP within the last day.

When making these prob-matches between different personas, our system records a “confidence level.” This allows us to move linked personas in and out of consideration depending on the use case. For example, a “match guaranteed” deep link used for auto-login would obviously require a confidence level of 100%, but the industry expects ad installs to be matched with a confidence level usually between 50–85% (the persona graph allows Branch to hit the top end of this range without being forced to accept lower-confidence matches).

Today, Branch dynamically sets the confidence level required for each use case, but this is a configuration we could expose directly to our customers in the future.

Step 3: Scale the network

It’s impossible to just “build a persona graph” because — in the beginning — there is no reason for anyone to sign up.

Why? The value of a persona graph increases for everyone as more companies contribute to it, which means the benefit of joining an existing persona graph is enormous, but there is very little incentive to be one of the best participants in a brand new persona graph — it would be like giving up that already-flipped Concentration game for a new one where you’re playing all by yourself.

Because Branch started out by solving cross-platform user experiences, our persona graph scaled as a natural side-effect of other products that provide independent value at the same time. This approach allowed the Branch persona graph (which now covers over 50,000 companies) to reach critical mass. However, while basic deep linking was a hard problem to solve back in 2014, it is now well on the way to commoditization. Today, it would be almost impossible to get a persona graph off the ground using basic deep links, let alone ever reach a similar level of coverage.

Step 4: Use the match data

What can Branch do with these cross-platform/cross-channel/cross-device personas? Here are a few examples:

Solve attribution ambiguities. This is the obvious one, of course. The persona graph makes it possible to correctly attribute the complicated user journeys we’ve been discussing, such as when you and the other Starbucks customer were both using the same shopping app, and traditional fingerprint-based attribution methods couldn’t tell the difference.

Provide data for true multi-touch reporting. Using multi-touch modeling to better understand user activity is the Promised Land of attribution: every marketer wants it, and everyone has a different idea of what it should be. But there’s one thing everyone should agree on: multi-touch attribution is only as good as the data you feed it, and bad data compounds the problem.

The persona graph allows Branch to consolidate data from across channels and platforms. Legacy mobile attribution providers completely miss this data, which means their “multi-touch attribution” is really just “multi-ad app install attribution.”

Protect user privacy. Fingerprinting has long been a necessary evil for mobile attribution, but inaccurate measurement isn’t the only cost — when fingerprinting matches the wrong user, this also introduces user privacy issues because it means the system believes it is dealing with someone else. The persona graph allows Branch to dramatically reduce the risk of incorrect matching (we even offer a “match guaranteed” flag to enforce it), better protecting the privacy of end users.

Go beyond measurement. Attribution is only possible if the conversion happens in the first place. The persona graph allows Branch to provide the seamless cross-platform user experiences that make this more likely, improving the performance of all your marketing efforts.

For example, if a user lands on your website, even though they already have your app installed, Branch can use the persona graph to detect this and show that user the option to seamlessly switch over to the same content inside your app, where they’re much more likely to complete a purchase.

Comparing persona graph attribution with previous-generation alternatives

To wrap up, let’s revisit the three core tasks of an attribution system, and compare the capabilities of a persona graph-based platform with the traditional alternatives.

1. Capture interactions

Mobile attribution providers started with ads, and have struggled ever since to retrofit their systems in a way that accommodates other channels.

A persona graph is able to support ads, but also support email, web, social, search, offline, and more.

2. Count conversions

Mobile attribution providers are optimized to capture app install events, and aren’t set up to handle non-install conversions that happen on other platforms. Many of them are now rushing to figure out how to perform basic web measurement, a problem that was solved years before apps entered the picture.

A persona graph can attribute app installs, and also captures other down-funnel conversions on websites, desktop apps, OTT devices, and more.

3. Link conversions back to interactions that drove them

As described in part 2, mobile attribution providers have two matching methods available: they default to device IDs, and fall back on fingerprinting.

A persona graph-powered system can also use device IDs for single-platform user journeys (app-to-app), and has device ID <> web cookie pairs for cross-platform (web-to-app) user journeys. It may occasionally have to fall back on fingerprinting when a matched ID pair is not yet available, but this is a far less frequent situation.

What comes next

Fragmentation in the digital ecosystem is a hornet’s nest that can’t be un-kicked, and the challenge of attribution between web and app is just the beginning — it’s going to get worse (just imagine what it will be like when you need to attribute between your toaster and your car!)

Web and app is just the beginning — it’s going to get worse. Just imagine what it will be like when you need to attribute between your toaster and your car.

Attribution based on a persona graph makes it possible to handle this fragmentation, and a persona graph built on user-driven link activity is even more powerful because it leads to a virtuous circle: links are the common thread of digital marketing, which means they’ll always be the natural choice for every channel, platform, and device. These links help build the persona graph, and the result is increased ROI, comprehensive measurement everywhere, and more reliable links.

No other platform-specific attribution solution is even in the same league.

At Branch, we see attribution as one part of a holistic solution that provides far more than app install measurement. Our true mission is to solve the problem of content discovery in the modern digital ecosystem. Deep linking was one critical part of this mission. Fixing attribution is another. But the real win is yet to come…stay tuned!

Appendix: FAQ & Objections

What if device manufacturers try to limit the persona graph?

Device manufacturers have a duty to protect their users. They also need to ensure their ecosystems allow companies to be commercially viable. A privacy-conscious, third-party persona graph is an excellent fit for both of these requirements.

Branch works closely with a number of device manufacturers. They are aware of our platform, and supportive of the solution we’ve built.

Doesn’t a persona graph allow companies to steal their competitors’ proprietary data?

No, it does not, because the only data available via a persona graph is knowledge of the connection itself. Not where or how the connection was made, or by which company’s end user. A healthy persona graph contains thousands of participants, ensuring no single company is disproportionately represented, and to survive, a persona graph must guarantee that it will never allow any company to access data it hasn’t independently earned.

Persona graphs sound problematic for user privacy…

A persona graph makes it possible to recognize a given user in different places, but it does not tell you anything about WHO that user is. And like cookies or device IDs, the connection is resettable on demand.

Branch feels so strongly about user privacy that we’ve adopted the Branch Guiding Privacy Principles. Here they are in full:

We limit the data we collect. We practice data minimization, which means that we avoid collecting or storing information that we don’t need to provide our services. The personal data that we collect is limited to data like advertising identifiers, IP address, and information derived from resettable cookies (the full list is below in our privacy policy). We do not collect or store information such as names, email addresses, physical addresses, or SSNs. Nor do we want to. In fact, our Terms & Conditions prohibit our customers from sharing with Branch any kind of sensitive end-user information. We will collect phone numbers if a customer uses our Text-Me-the-App feature — but in that case, we will collect and process end user phone numbers solely to enable the text message, and will delete it within 7 days afterwards.

We will only provide you with data about actual end-user activity on your apps or websites. Our customers can only access “earned” cookies or identifiers. This means that an end user must visit a customer’s site before our customer can see the cookie; and an end user must download a customer’s app in order for Branch to collect the end user’s advertising identifier for that customer. In short, the Branch services benefit customers who already have seen an end user across their platforms and want to understand the relationship between those web visits and app sessions.

We do not rent or sell personal data. No Branch customer can access another Branch customer’s end-user data. And we are not in the business of renting or selling any customer’s end-user data to anyone else. To enable customers to control their end-user personal data, they can request deletion of that data at any time, whether in bulk or for a specific end user. These controls are available to customers worldwide, although we designed them to comply with GDPR requirements as well.

How is a persona graph different from “identity resolution” or “people-based marketing” products?

While these products may have similar-sounding names and seem comparable on the surface, they are very different underneath. Here are three major contrasts:

How they are built. The data for these products is typically purchased in bulk from third-parties, and then aggregated into profiles. The Branch persona graph is built from directly-observed user activity, and does not incorporate any personal data acquired from external sources.

What they contain. The user profiles available via these products typically contain sensitive personal data like name, email address, age, gender, shopping preferences, and so on. The Branch persona graph contains only anonymized, cross-platform identifier matches, and has no use for sensitive personal data — we don’t even accept it from customers.

How they are used. A major use case for these products is selling audiences for retargeting ads. This is a fundamentally different objective than the accurate measurement and seamless user experiences that Branch exists to provide.

What about fraud?

Fraud is a never-ending game of cat-and-mouse: as long as there is value changing hands (the literal definition of an ad), fraud can never be truly solved because savvy fraudsters will always find a way through.

The realistic objective of a mobile attribution provider is to block “stupid fraud,” and make fraud hard enough that fraudsters will go somewhere else. The best way to do this is by weeding out anything that doesn’t reflect a realistic human activity pattern. A persona graph has vastly more sophisticated data to use for this assessment than any single-channel, single-platform system.

What about when the persona graph doesn’t cover a user?

Even with a network size of Branch, there are still situations when the persona graph isn’t available. As just a few examples: the first time seeing a new device, browser cookie resets, ITP in iOS, etc.

In those situations, the system has to fall back on the next-best matching technique available. In Branch’s case, this is still as good as (and usually better than) what is available via legacy attribution providers.

What about cross-device attribution?

Cross-device is a surprisingly complicated problem. In theory, the persona graph can connect data across devices, just like it does across channels and platforms.

Some mobile attribution providers have recently begun leaning on cross-device tracking as their entry into “people-based attribution.” Essentially, they merge activity based on a customer-supplied identifier such as an email address or username — if you sign in with the same ID on two devices, then they consider these to belong to the same person for attribution purposes.

This sounds logical on the surface, and it works for these providers because they’re still approaching measurement from a siloed, one-app-at-a-time perspective. Branch already does similar cross-device conversion merging based on user IDs on an app-by-app basis, in addition to the persona graph.

Here’s where things get complicated for cross-device as part of a persona graph:

It’s fairly rational to assume the majority of activity on a single mobile device is from a single human. Sure, people let their friends make a phone call, or check the status of a flight, and this has the potential to muddy attribution data somewhat, but the impact is pretty limited. However, if a user lets a friend sign into their email account on a laptop to print a flight confirmation, and the attribution provider then uses that as the basis to merge identity fragments across the entire persona graph network, the cascading effects could lead to massive unintended consequences.

Our customers ask us about cross-device attribution regularly, and our research team has made good progress. We feel data integrity is the most valuable thing we can offer, so we haven’t rushed because we want to make sure we get this right.

Why are deep links so important?

Some legacy mobile attribution providers feel that deep links aren’t critical to attribution. And from a certain perspective, they’re right: it’s perfectly possible to be a bean counter without also being a knowledgeable guide. At Branch, we feel this is an extremely shortsighted perspective, because the ongoing fragmentation of our digital ecosystem means that without working links, eventually there will be nothing left to measure.

Let’s illustrate this with an example from the offline world:

Imagine a billboard for your local car dealership. Driving down the highway, on the way home from the grocery store, you see this billboard advertising the newest plug-in hybrid. You don’t really need a new car, but your old one has been leaking oil all over the garage floor for months and the Check Engine light came on last week, so you decide (on the spur of the moment) that you want to stop in for a test drive.

You’re excited. You can almost smell “new car” already, and you’re all set to take the highway exit for the dealership…but the off-ramp is blocked by a big orange sign: “Closed for Construction.” You’d have to go five minutes further up the highway to the next exit, and then spend ten minutes figuring out how to drive back on local roads. And besides, that milk in the back seat is going to spoil if you leave it in the sun. You give up and go home.

A week later, you happen to be driving by the dealership again. The highway exit has reopened, and that new car smell has been following you around everywhere for the last few days. But the billboard is now advertising your local bank, and that ad you saw a week earlier has completely faded from your memory. When the salesperson asks, “What caused you to come by today?”, you say, “Oh, I just happened to be in the neighborhood.”

Now, the dealership has two problems:

You might never have come back after the first broken journey. You might even have gone to another dealership instead, because all new cars smell pretty much the same.

The dealership has no idea that the billboard is the real reason behind your visit. Because you don’t even remember yourself. If you end up buying, the billboard was a worthwhile investment…but they’ll never know this, because the highway construction interrupted your journey and broke the dealership’s attribution loop.

It’s not much of a stretch to replace “car” with “app,” “billboard” with “install ad,” and “highway exit” with “link.”

The reality is that in the digital world today, links are the customer journey. If your links don’t work, then even the best measurement tool in the world can’t help you attribute conversions that never happened.

Bottom line: if you find an attribution system that claims to provide measurement without also solving for links that work in every situation (and proving with verifiable data that their links don’t break), be very, very skeptical. It’s likely you’re dealing with a legacy system that hasn’t adapted to changes that have happened in the ecosystem over the last few years.

What if another company creates a persona graph?

This is always a possibility, but due to the nature of network effect, it would be extremely challenging for any other company to reach the critical mass necessary to compete with the Branch persona graph.

What about Self-Attributing Networks (SANs)?

SANs like Facebook, Google, Twitter, and so on hook their walled gardens into the ecosystem via the device ID. The difference is that instead of allowing attribution providers to observe all of the user’s interactions, the SAN just responds with a Yes or No when asked the question “hey, this device ID just did something…did you see that user in the last X days?”

The SAN approach has advantages (fraud is an almost non-existent problem) and disadvantages (it’s a black box that provides very little visibility), but it’s a reality of the ad ecosystem.

Since most walled gardens already connect theirs users across platforms through a user ID or email address, there’s no reason why the SAN can’t start reporting activities by that user on other devices/platforms. This sort of connection gets incorporated into the persona graph automatically through the associated device ID.

What about Limit Ad Tracking (LAT) on iOS?

When LAT is enabled, iOS sends the IDFA as a string of zeros. Currently, it appears that around 20% of iOS have this setting enabled. Without an IDFA, Branch is unable to connect that user to the persona graph, but we are still able to perform attribution via fingerprinting or the IDFV (an alternative device ID that is available even with LAT enabled, but scoped to a single app/vendor).

<a href="https://medium.com/media/3c851dac986ab6dbb2d1aaa91205a8eb/href">https://medium.com/media/3c851dac986ab6dbb2d1aaa91205a8eb/href</a>

L O A D I N G
. . . comments & more!