Attribution

Before someone buys something online, they usually see several ads. First maybe a video on YouTube, later an ad on Instagram, then also a promotional email. In the end they search for the shop on Google and place an order. Now the question arises: which of these ads actually caused the purchase? Attribution is the attempt to answer exactly that and to distribute the success across the individual touchpoints. The term comes from the Latin “attribuere”, meaning “to allocate”.

Why companies fight over every attributed dollar

Large companies spend billions on advertising every year. They must decide how to distribute this money across channels. Does TikTok get more budget, or rather the search engine? This decision depends directly on which channel the sales are credited to. Attribution is therefore not an academic pastime, but shifts real flows of money.

A famous saying from marketing goes roughly: Half of my advertising spend is wasted, I just don’t know which half. Online advertising promised to solve this problem, because every click can be measured. In practice, however, attribution remains difficult. People switch between phone, laptop, and TV, and part of their journey simply isn’t visible.

For stock markets and financial news, this topic is relevant because the revenues of Meta, Google, or Amazon depend on it. If advertisers can attribute fewer sales to a channel, they book less there. When Apple restricted cross-app user tracking in 2021, part of Meta’s measurement basis collapsed. The stock price reacted sharply at the time.

From the last-click rule to the calculated model

The simplest method is called last click. The last advertising contact before the purchase gets all the credit, while all earlier ones get nothing. This is easy to measure, but obviously unfair. The search ad through which someone finally orders benefits from the attention that the video generated weeks earlier.

That’s why there are models that split the success across multiple contacts. Some distribute it evenly, others give more weight to the first and last contact. Data-driven models go further: a program compares a very large number of purchase paths with each other and estimates how strongly each contact increased the probability of purchase. Such methods are the point where machine learning comes into play, meaning software that derives patterns from large amounts of data.

A common mistake is confusing attribution with causation. A model only sees that a contact took place and a purchase followed afterward. Whether the person would have bought without the ad, it does not know. Anyone who searches for a brand on Google was already determined to buy anyway. To test real causes, companies use experiments: a randomly selected group does not receive the ad, and then both groups are compared.

Attribution in ad accounts, data protection, and AI debates

Anyone who opens an ad account with Google or Meta can set the attribution model themselves there. Website analytics programs also show which path visitors came through. Small online shops notice this at the latest when two platforms both claim the same sale for themselves and the sum of reported revenues is higher than the actual amount.

Closely connected is the debate about data protection. Classic attribution relied on cookies, small files that recognize a browser. Because browsers increasingly block these files, providers work with estimation models and with methods that only output group values instead of individual persons.

The word also appears in a completely different context. With AI systems, one likewise asks about attribution when it comes to which source or which training data shaped an answer. News agencies also speak of attribution when they assign a cyberattack to a specific perpetrator. What all these meanings have in common is the basic question: to whom or what is a result to be credited?