How Does X’s Algorithm Respond During Electoral Campaigns? An Experimental Approach to Elections in Argentina

What a social media user sees (or doesn’t see) on their phone screen largely depends on recommendation algorithms. According to the latest Reuters Institute report, more than half of respondents in Argentina use social networks to get information, so the content that platforms prioritize in their feeds not only contributes to amplifying certain narratives but can also influence the public conversation agenda.

In the context of Argentina’s 2025 legislative elections, and based on the research by Global Witness (GW), at Chequeado we explored the content that X’s algorithm shows to new users, especially those interested in politics but without an explicit partisan leaning.

We decided to experimentally analyze X’s behavior because it is the platform chosen by influential actors in public debate to communicate viewpoints and public policy decisions; and because messages circulating on that social network are often amplified later by news media and reach a wider audience.

Following GW’s methodology, we created 6 users with different demographic profiles, each showing explicit interest in the 2 main political forces in the elections in the City of Buenos Aires (CABA) and Province of Buenos Aires —La Libertad Avanza (LLA) and Fuerza Patria (FP)— and in political topics in general. To reflect this interest, each account began by following the lead candidates of both forces in CABA and PBA, as well as the official accounts of each force. Then, we analyzed what content appeared in those users’ feeds during their first 20 minutes of exploring the “For You” section, where the platform algorithmically shows recommended posts. The observation was conducted on October 9 (for more details, see the Methodology section).

For each tweet with political content that appeared in the “For You” section, we analyzed whether the post expressed support for or rejection of a political party, a policy measure, or a political stance associated with them. With this information, we classified each tweet according to whether it was favorable to the LLA, to FP, to both forces, to another political force, or to neither.

During the observation, we found more posts favorable to La Libertad Avanza, even when the user’s initial interest was balanced. Additionally, the vast majority of total content collected came from accounts that users did not follow (i.e., recommended by the algorithm) and, mostly, from verified accounts.

What Did We Find?

In this experiment, which has limitations and whose conclusions cannot be generalized, 55% of the posts that appeared in the “For You” section of the 6 analyzed users were favorable to La Libertad Avanza, while 39% favored Fuerza Patria, — that is, 41% more posts favored LLA than FP..

The majority of content favorable to both forces came from accounts that users did not follow, meaning they were algorithmically recommended. Of those tweets favorable to LLA, 70% corresponded to accounts recommended by the algorithm that users did not follow, while 30% were tweets from followed users. Regarding tweets favorable to FP, 89% corresponded to accounts that users did not follow, while 11% were from followed accounts.

If we focus on all tweets from accounts the users did not follow, we find a relative balance between those favoring LLA and those favorable to FP, with those favorable to LLA being only 11.9% higher.

However, we noticed a marked difference in the characteristics of unfollowed accounts: 97% of these posts that favor LLA are from verified accounts, compared to 67% of those favoring FP. That is, among accounts the users did not follow, almost all those supporting LLA are verified. 

Regarding the general behavior of posts appearing in the analyzed users’ feeds, we found that 79% come from accounts that were not followed. And of those, 84% correspond to verified accounts. This would suggest that X’s algorithm prioritizes verified accounts with blue checkmarks for recommendation through its algorithm. This verification checkmark is a paid platform feature.

Methodology

We followed the methodological guidelines of Global Witness research.

1. We created 6 new users on the web version of X. Before conducting the experiment with each of them, we deleted cookies and browsing history so this would not affect the content shown by the algorithm. Each user was assigned a profile distinguished only by age and gender variables. Thus, 3 users were configured as women aged 16, 25, and 50 years, and 3 were configured as men aged 16, 25, and 50 years.

2. With each user, we followed 8 accounts  relevant to the upcoming electoral process: the official accounts of La Libertad Avanza and Fuerza Patria, and the lead candidates in CABA and PBA for the categories voted in each district. Data collection took place on October 9, when the head of the list of LLA was Karein Reichardt, by a judiciary decision. 

3. In each of the followed accounts, we interacted with their 5 most recent posts in the “Posts” tab of the profile. In cases where posts included videos, we played the first 30 seconds of each. When they involved threads, we scrolled to the end of them. When they included images, we viewed each of them.

4. Next, we navigated for 20 minutes in the “For You” tab, which presents algorithmically defined content. There we interacted with tweets with “political content” in the same manner described in the previous step. We defined “political” tweets as those referring to parties and alliances, candidates, representatives, proposals, government actions, and evaluation of administrations corresponding to any sector.

5. The collected posts were classified according to:

• Whether they corresponded to an account that was followed or not.
• Whether they were retweeted by an account that was followed or not.
• Whether they expressed support for or rejection of the main forces competing in the elections.
  • We also recorded cases where support or rejection was expressed for a third force, and those where it was not possible to determine a defined inclination.
  • We considered tweets critical (or accusatory) of candidates or public figures from these forces as “rejection” of a force.
• Whether they expressed support for or rejection of positions or actions associated with the main forces competing in the elections.
  • We also recorded cases where support or rejection was expressed for a third force and those where it was not possible to determine a defined inclination.

From this coding, we determined which posts were favorable to LLA, how many were favorable to FP, and how many were not clearly favorable to either of the two.

Limitations

The experiment was conducted with a small number of cases—6 new accounts—and does not seek to represent all users. Even so, the results show a possible part of the experience that some people may have on the platform, which allows identifying certain patterns and better understanding how X’s recommendation system works. The results should be read as a starting point: an initial exploration that offers clues and raises questions for broader studies or future tests with greater scope.

It could be argued that part of the reason we saw a greater amount of content favorable to LLA is that this sector has more popularity on X or publishes more frequently than other parties. The fact is that we don’t know how these algorithms work and that, despite the user showing equivalent interest in both political forces, they end up seeing more content favorable to one. We are not claiming that X deliberately favored content favorable to LLA over FP, but rather what we point out is that its algorithms produced that result under the conditions established for this exploration.