With the internet and smartphones having become ubiquitous, technology companies now have a greater amount of detailed customer information at their fingertips than ever before. They know how users surf the web, what products they are likely to buy, how they behave on social media platforms and even their location. This information enables advertisers to tailor their ads to extremely precise target groups. And that has revolutionized online advertising. Instead of developing broad advertising strategies aimed at reaching as many people as possible, companies have begun personalizing their ads and placing them only where they will have the greatest effect. To achieve the most accurate targeting possible, AI systems are used to analyze huge amounts of user data and produce detailed user profiles. Based on those profiles, customers are divided up into target groups, and for each of those groups, demand forecasts are generated, which will determine the advertising they are shown.

This practice has generated significant debate, with most of the discussion focusing on data protection and ethical concerns. The potential ecological risks, however, are less frequently addressed. According to estimates, internet use is responsible for the consumption of over 400 terawatt hours (TWh) of electricity each year. Current trends suggest that this total will continue to grow strongly in the coming years. Little research has been done into how the personalization of advertising using data analysis methods has impacted energy consumption. It seems fair to assume, however, that it has amplified existing trends.

The user data fed into such analyses comes from a number of different sources, including websites, social media platforms and mobile applications. Network infrastructure and data centers are required for the transfer of the data, all of which consume energy. And the data collected must be stored and managed for extended periods in data centers and on hard drives – and here too, energy is consumed. The next step involves using the data to train the machine learning models that are the backbone of personalized advertising campaigns, a process which requires (energy intensive) high-performance computer infrastructure with networked servers. And finally, servers must also be cooled during operation, which accounts for a significant share of a data center’s energy consumption.

But that’s not all: The determination as to how the advertisements are ultimately placed takes place in a process called “real-time bidding.” Advertisers bid in real-time auctions for ad real estate on websites that seem relevant to the target groups they are trying to reach. These auctions require extremely quick data processing so that advertisers, on the one hand, and those with advertising real estate, on the other, communicate with each other. The rendering of content such as the images and video that generally accompany advertisements when they are presented on end devices, also requires energy.

Simulations generate data about the amount of energy consumed by the deployment of personlized online advertising. But such studies tend to focus primarily on the energy consumed by the end device to which advertising data is transmitted. It is difficult to estimate by contrast, the amount of electricity consumed by the AI systems and data analysis procedures deployed in the personalization process or by the storage of the necessary data. Advertising companies are not always willing to share such information.

For our study, we used the data-protection tool OpenWPM to collect data about website visits, and we examined around 200 of the most-visited German internet domains. We installed an automated crawler on a laptop and had it repeatedly open websites while we collected data on the amount of energy used by the computer’s CPU and information about the cookies transmitted. We conducted the simulation both with and without an ad blocker to determine the effect that ads have on end device energy consumption.

The time it took to open the websites was 14 percent lower when the ads were blocked, which translated to 10 percent less energy being consumed by the end device’s CPU. Depending on the complexity of the website visited, the end device specifications and the browser used, opening a website generally resulted in energy consumption of 0.01 to 1 watt hours (Wh). The rendering of an advertisement by the graphic card produced average energy consumption of 0.005 Wh. An average of 155 cookies, with a mean size of 138 bytes, were transferred during each website visit, 87 percent of which came from third parties. During each website visit, an average of 0.2 MB worth of data was transmitted in the form of cookies. The rejection of non-essential cookies reduced the amount of cookie-related data transferred by 75 percent. The greatest savings were achieved with websites that fell into the “news and media” category, which are often financed through the serving of ads.

The amount of energy consumed by a single website visit and the related transfer of cookies is, to be sure, quite small. But just the 200 German websites analyzed for this simulation alone are visited 4.5 billion times per month. In our experiment, we only visited the homepages of the sites we included in the study, but generally, visitors also navigate to subsites as well. Ads are frequently served from those pages too, which translates to the additional transmission of data.

Our results show that the risks of personalized online marketing are not limited to privacy protection issues. It is estimated that the internet is responsible for 15 percent of total global electricity consumption – and rising quickly. Online advertising and the collection of user data are both big contributors. The flow and processing of data that is produced by online advertising can be limited through the technology deployed (“Privacy by Design”) and the settings chosen (“Privacy by Default”). Companies that use AI-based data analysis procedures to analyze user data for the purposes of personalized online marketing should face stricter requirements to take steps to reduce energy consumption and make available data pertaining to the energy efficiency of their systems so as to gain a better understanding of their ecological impact.