Artificial intelligence or “AI" tracking has been all the rage since marketers, business operators, and business owners were desperate to fill the iOS14.5+ shaped hole in their hearts back in 2021.
We won’t bury the lead so here goes.
We all need to slow down. It’s concerning to see many of us in ecommerce willing to accept data prima facie from alleged “AI” data providers when we don’t fundamentally understand how that data is collected or have any sense as to how their technology works.
As an owner or marketer, you make important business decisions based on that data. You better have at least a general understanding of how it’s collected and what its limitations and underlying assumptions are.
Imagine if you took the same approach to your finances. For example, if you’re relying on your income statement, balance sheet, or cash flow statement to make important business decisions, you and the personal who prepped your financial statements would need to agree on definitions and accounting methodology. Before a deep analysis, you better know things like:
All of the things mentioned above can greatly impact your assessment of business performance, your profitability, and even your pricing strategy. Accurate data is important, but as an operator you need to understand the context around your data in order to understand and interpret it.
Okay, enough analogy.
We’ve seen an abundance of attribution software solutions spring up recently and new terms like AI tracking being thrown out at us haphazardly.
There is no such thing as AI tracking. It’s hogwash. Poppycock. And also huh-uh.
It’s just a term used to market to marketers.
It may be a hard pill to swallow. Many of you may be using AI tracking software that we're gearing up to throw some dirt on. But do you understand how it works? Are you sure it’s working the way you think it is? The question isn't whether or not there are assumptions and limitations to the data that's being reported to you, the question is whether or not you know what they are.
Facebook iOS14.5+ brought a number of challenges, and we want to define the context we’re specifically speaking to in this gentle tirade. The biggest and most annoying challenge is the loss of cookie data after 7 days. Yes, there are other issues (like the opt-out prompt), but the biggest concern we would argue is the loss of data beyond 7 days for iOS/Safari users (which is technically part of Safari’s Intelligent Tracking Prevention).
Ok, so the context of the conversation today has been clearly stated – data loss for iOS14.5+ users and users using Safari (or in-app Safari browser) beyond 7 days. We’ll refer to this as the iOS14.5+ problem for the remainder of this written reproach.
When attribution software companies first started popping up over a year or so ago, WE ALL were promised that AI was the answer to the iOS14.5+ problem.
Of course, if you were press these companies on what they mean by “AI” and how their tech works, you get a variety of answers like:
“Sorry, we can’t divulge that. That’s our secret sauce.”
“It’s AI magic.”
“We measure a bunch of ‘signal’ from the user that our AI can leverage.”
Alright. What the h*ck does that mean? How exactly is AI being used by these types of companies?
The Tl;DR version is that…they aren't. The iOS14.5+ problem is not something that can currently be solved with AI.
Artificial Intelligence in a business analytics context, is almost entirely synonymous with machine learning (neural networks are used fairly often, too). Machine learning (ML) is at its core statistics. It can do some incredible things, but at the end of the day it’s not magic; it’s statistics.
Machine learning depends on a good training data set. The whole idea with ML is to develop an algorithm from a training data set either through supervised, unsupervised, or reinforcement learning. Any good data scientist will tell you, the training data set is very, very important.
If you want to develop an algorithm that identifies an image of a dog, for example, then you better have lots of images of what are clearly dogs with which to train your model. No aardvarks.
To restate the important question: How are "AI" tracking companies using this process (or anything that could be considered artificial intelligence) in order to solve the iOS14.5+ problem?
They're not. There's a glaring chicken and egg problem. You seen, in order to train an algorithm to recognize users beyond 7 days, one would require a lot of good data that didn’t have a 7 day limitation so the model could "learn" what these users looked like. Otherwise, it would be akin to feeding images of aardvarks to your dog image model. It just doesn’t make sense. What’s more is that if there were a way of procuring the type of data we would need in order to train a good algorithm to begin with, we would have no need for machine learning at all. If such an ability existed, it would mean we've solved our original problem. No AI/machine learning necessary.
Ultimately, the iOS14.5+ problem is not a problem that has an application in AI. Instead, we’ve lost our data. We can’t regenerate it; we can’t analyze it. It’s gone. To be clear, machine learning offers many wonderful benefits on good data you already have – especially in the area of predictive analytics. But, machine learning doesn’t magically replace lost data.
If you understand one thing, understand this. The issue of iOS14.5+ hinges on the ability to retain key pieces of data tied to a user i.e. a unique, persistent identifier over a long period of time. This is a technical limitation that must be overcome and is not something that can be solved with AI.
BUT WAIT! I hear you scream at me as you shake your fist at the screen, my attribution software measures unique signals from each user! They most certainly to do not to a meaningful degree. At least, not for iOS14.5+ users.
These “data signals" consist of data that can be extracted from the device and browser with javascript. The idea is that a user will have unique settings or configurations in their device and browser that can be leveraged in order to create a unique identifier – like a fingerprint. It’s a nice idea and one that may have worked pretty well 10-15 years ago. Today, especially if we’re looking at iOS14.5+ you just can’t fingerprint in any meaningful sense. Safari ITP has taken action against this and have changed their browser API’s in such a way that does not allow for fingerprinting. You can refer to this documentation if your curious to learn more: https://webkit.org/tracking-prevention
For the sake of fairness, we concede that you can combine fingerprint data with other data (like an ip address) to come up with a more accurate fingerprint. However, at that point, the ip address is really the unique identifier and driving influence behind the fingerprint. The fingerprint from a modern iOS user is essentially useless. They all look about the same. Again, this would not require the use of AI.
We say all of this to point out that there are still significant challenges on tracking iOS14.5+ users. It’s very difficult to get any of these data providers to admit that.
To a certain degree, we understand why many of the big names in “AI" tracking want to keep the technical details behind their business underwraps. It's understandable. The big issue, though, is when these same companies over-promise on what they can deliver or misrepresent what they actually do and then balk or defer to “AI magic” when pressed about their claims.
So do these AI tracking solutions offer benefits? Some, absolutely. Are they just making up data? We certainly hope not. But, data collected and processed even with the best of intentions would be significantly more beneficial to you if you understood how the tracking worked worked along with its inherent limitations and assumptions.
Simple. We cut the bull. We work closely with each of our partners so that you can understand how our algorithms work and what assumptions (if any) are made with the attribution done to each of your purchases. We consider ourselves part of your team and expect that our data be scrutinized with a careful eye. We will be here to answer any questions that may come up.
For example, if you’re looking at Kluvos reporting and don’t understand why Kluvos selected the 6 sessions it did for attribution for order #56122 we will tell you. We can breakdown the logic behind each attribution decision. In this way you will start to understand how your data is collected and processed, and understand any assumptions or limitations that might be present.
We can even tune our algorithms on a store by store basis to get an attribution algorithm that works best for you.
Sound interesting? Nerd. JK! We hope so. You can sign up free today :)
FREE Access!