Personalization of Web Search During the 2020 US Elections

TL;DR

This study uses synthetic internet users to measure how search results during the 2020 US Elections vary based on user preferences and behavior, revealing biases in personalization.

Contribution

It introduces a novel experimental framework with synthetic users to causally analyze personalization effects in political search results.

Findings

Search results vary significantly across users.

Google favors previously visited and local news sites.

No strong bias towards user ideology in search rankings.

Abstract

Search engines play a central role in routing political information to citizens. The algorithmic personalization of search results by large search engines like Google implies that different users may be offered systematically different information. However, measuring the causal effect of user characteristics and behavior on search results in a politically relevant context is challenging. We set up a population of 150 synthetic internet users ("bots") who are randomly located across 25 US cities and are active for several months during the 2020 US Elections and their aftermath. These users differ in their browsing preferences and political ideology, and they build up realistic browsing and search histories. We run daily experiments in which all users enter the same election-related queries. Search results to these queries differ substantially across users. Google prioritizes previously visited websites and local news sites. Yet, it does not generally prioritize websites featuring the user's ideology.