Flocking to Mastodon: Tracking the Great Twitter Migration

We collect a global list of Mastodon instances from instances.social, which contains a comprehensive index of Mastodon instances. We compile a set of 15,886 unique instances. We then collect all tweets containing a link to these Mastodon instances using Twitter’s Search API.¹¹1https://api.twitter.com/2/tweets/search/all Additionally, we collect all tweets containing the following list of keywords related to the migration from Twitter: ‘mastodon’, ‘bye bye twitter’, ‘good bye twitter’; and hashtags #Mastodon, #MastodonMigration, #ByeByeTwitter, #GoodByeTwitter, #TwitterMigration, #MastodonSocial, #RIPTwitter. In total, we collect 2,090,940 tweets posted by 1,024,577 users between October 26, 2022 (i.e. a day before Musk’s takeover) and November 21, 2022. Figure 2 shows the temporal distribution of these tweets.

We next search for Mastodon usernames in these tweets and the accompanying metadata of any account that posted a tweet (i.e. display name, location, description, URLs, pinned tweet text). Mastodon usernames take the form @[email protected] and https://example.com/@alice, where alice is a username and example.com is an instance. To map a Mastodon handle to a Twitter account, we do this search in a hierarchical fashion: We first look for Mastodon usernames in user metadata (e.g. bio) and create a mapping between Twitter account & Mastodon account if one is found. If the search is unsuccessful at the first step, we then look for Mastodon usernames in the tweet text. To ensure mapping accuracy, we only map a Twitter account to a Mastodon account identified from a tweet text if both the Twitter and Mastodon usernames are identical.

Using this methodology, we identify the Mastodon accounts of 136,009 Twitter users, which are created across 2,879 unique Mastodon instances. We find that 72% of Twitter users that migrated created a Mastodon account with the same username that they use on Twitter. 4% of the Twitter users who create a Mastodon account, have a (legacy) verified status (i.e. authentic, notable, and active) on Twitter, suggesting that even well-established users have been migrating.

While we track and analyze the migration of a large number of Twitter users (136,009), the takeover of Twitter by Musk, is likely to have pushed even more users to migrate who the above methodology cannot identify. Indeed, on November 12, 2022, Mastodon announced that over 1 million users had registered since October 27, 2022 [16], significantly more than our methodology identifies. To understand the wider activities on the Mastodon instances, we cross-check the new registrations on the 2,879 instances by crawling their weekly activity from Mastodon’s Weekly Activity Endpoint.²²2https://docs.joinmastodon.org/methods/instance/#activity Figure 3 shows the weekly number of registrations, logins and statuses. We notice a large increase in all three activity metrics after the Twitter acquisition. Of course, we cannot confirm that all these users migrated directly from Twitter. However, given the timeline of registrations, we believe that it is very likely that a large share of these new users migrated from Twitter.

We next crawl both the Twitter and Mastodon timelines of the migrating users identified in the previous section. We use Twitter’s Search API and Mastodon’s Account’s Statuses Endpoint.³³3https://docs.joinmastodon.org/methods/accounts/#statuses For each user, we crawl all tweets/statuses from October 01, 2022 to November 30, 2022. In total, we gather Twitter timelines for 94.88% of the users. The rest were suspended (0.08%), deleted/deactivated (2.26%), or the tweets were protected (2.78%). We crawl the timelines of 79.22% of Mastodon users: the rest have either not posted a single status (9.20%) or their instances were down at the time of crawl (11.58%). In total, we gather 16,163,600 tweets, and 5,746,052 Mastodon statues.

We also crawl the user’s followees for both Twitter and Mastodon accounts. We use the Twitter Follows API⁴⁴4https://api.twitter.com/2/users/:id/following and the Mastodon Account’s Following Endpoint⁵⁵5https://docs.joinmastodon.org/methods/accounts/#following respectively. Due to the rate limitations of the Twitter’s API we crawl a sub-sample of 10% of the migrated users. For representativity, our sample relies on the followees distribution takes 5% from above the median value and 5% from below. In total, we gather followee data for 13,068 users. This covers 11,453,484 followee relationships.

The datasets in this paper include both user and post information, and therefore might have privacy implications. To overcome any data mishandling, we have exclusively collected the publicly available data following well-established ethical procedures for social data. We have obtained a waiver from the ethics committee at the author’s institution.⁶⁶6anonymised for double-blind submission We anonymize the data before use and store it in a secure silo. Upon acceptance of the paper, anonymized data will be made available to the public, which we hope will help further works.

We first analyze the size of the social network (i.e. number of followers & followees) that the migrated users have on both Twitter and Mastodon. Figure 7 plots the CDF of the number of followers and followees of migrated users on both platforms. The median followers and followees that migrated users have on Twitter are 744 and 787, respectively. Just 152 users (0.11% of total migrated) have no Twitter followers, and 465 (0.35% of total migrated) have no Twitter followees. In contrast, on Mastodon, 6.01% of users have no followers, and 3.6% do not follow anyone. The median followers and followees on Mastodon were 38 and 48, respectively. Interestingly, 1.65% of migrated users gained a median of 33 more followers on Mastodon than their Twitter followers. This confirms that these new users are yet to bootstrap a significant social network on Mastodon. However, we emphasize that the median age of migrated accounts on Twitter is 11.5 years, in contrast to just 35 days on Mastodon. Hence, due to these disproportionate ages, the size of the social networks on the two platforms are not directly comparable.

We next conjecture that a user’s (Twitter) social network may have an impact on their likelihood of migration. For example, if a user’s friends migrate to Mastodon, this may encourage them to do the same.

To inspect this, we analyze the followees data from both Twitter and Mastodon for 10% of the migrated users (see §3.3). Figure 8 shows CDFs of the fraction of Twitter followees of each migrated user that (i) moved to Mastodon (blue); (ii) moved to Mastodon before the user (orange); and (iii) moved to the same Mastodon instances as the user (green). We notice that just 5.99% of each user’s followees also migrate (on average). In fact, for 3.94% of the migrated users, none of their Twitter followees move to Mastodon. Thus, the majority of the social network of the migrated users seems indeed reluctant to migrate, and sometimes they are the first in taking this step.

To better understand this, we compare the date on which each migrated user joined Mastodon with that of their Twitter followees who migrated as well. We find that, out of their social network (i.e. their followees), 4.98% of the migrated users were the first and 4.58% were the last to migrate from Twitter to Mastodon. On average, 45.76% of the followees of a user migrated to Mastodon before the user actually did.

We are also curious to understand if users select the same Mastodon instance as their social network. We therefore compare the instance of each migrated user with that of its Twitter followees. On average, 14.72% of each migrated user’s followees (that move to Mastodon) join the same instance. With 15K+ Mastodon instances, this is a considerable proportion, suggesting a clear network effect. However, we also notice that this average is highly impacted by one flagship instance: mastodon.social. This is the largest instance available, and is probably the best known. Of all the migrated users whose Twitter followees move to the same instance, 30.68% are on mastodon.social. That said, we also find small instances that attract significant proportions of a given user’s Twitter followers. For example, 4.5% of the migrated users whose Twitter followees join them on the same instance are on mastodon.gamedev.place (a Mastodon server focused on game development and related topics)

A unique feature of Mastodon is that users can easily ‘switch’ instance. This involves migrating their data from one instance to another. We are curious to see if this is also driven by network effects. Overall, 4.09% of the users have switched from the Mastodon instance they initially created an account on (hereinafter first instance) to a new instance (hereinafter second instance). Curiously, 97.22% of these switches happened after Musk’s Twitter takeover. This suggests that users may join initial instances, but migrate to a more suitable one once they are more experienced.

Figure 9 shows the chord plot of switches from each user’s first Mastodon instance to their second. A common pattern across these switches is that users move from general purpose/ flagship instances (e.g. mastodon.social, mastodon.online) to more topic specific instances, e.g. sigmoid.social (a Mastodon instance for people researching and working in Artificial Intelligence) and historians.social (a Mastodon server for people interested in history).

Interestingly, we notice a strong social network influence behind these switches. Figure 10 shows the CDFs of the fraction of Twitter followees of each switched user that (i) moved to the first instance (blue); (ii) moved to the second instance (orange); and (iii) moved to second instance before the user (green). On average, 46.98% of each user’s followees (who moved to Mastodon) at some point also join the second instance. In contrast to just 11.4% who join the first instance. Interestingly, 77.42% of each switching user’s followees (on average) joined the second instance before the user. This suggests that the users switched from the first instance because a large fraction of their Twitter followees moved to the second one.

We first analyze the timelines of migrated users from both Twitter and Mastodon. Figure 11 shows the number of tweets on Twitter and the number of statuses on Mastodon posted by migrated users each day from October 01, 2022 to November 30, 2022. We observe a continuous growth in user activity on Mastodon after the acquisition of Twitter. However, the activity of migrated users on Twitter do not decrease in parallel, i.e. our migrated users are using both their Twitter and Mastodon accounts simultaneously.

We next check if people are generating identical content across both platforms or are, instead, projecting multiple ‘personas’. Figure 14 plots the CDFs of the fraction of each migrated user’s Mastodon statuses that are identical or similar to its tweets. We consider the Mastodon status similar to a tweet if the cosine-similarity of their sentence embeddings [20] is greater than 0.7. Surprisingly, just 1.53% of each migrated user’s Mastodon statuses are identical. On average, just 16.57% of each user’s Mastodon statues are similar to their tweets. Instead, 84.45% of the migrated users use the two platforms to post completely different content. This suggests a mix of users, some of whom create different personas on the two platforms, and a smaller subset who mirror all their content.

A potential explanation for the latter is the use of cross-posting tools. Such tools allow users to automatically mirror their Mastodon statues on Twitter, and vice versa. To examine this, we compare the number of tweets posted via different sources before and after Musk’s takeover in Figure 12. Naturally, the majority are posted by official Twitter clients such as the Twitter Web App. The two sources that increase most dramatically, however, are two well-known cross-posters, Mastodon-Twitter Crossposter and Moa Bridge — by 1128.95% and 1732.26%, respectively. Of all migrated users, 5.73% use one of the two cross-posters at least once. This suggested such users see both Twitter and Mastodon as viable platforms, and have limited intention of creating multiple ’personas’. Figure 13 also plots the number of users using cross-posters over time. We see that their usage increases rapidly after Musk’s takeover. The downward trend towards the end of November is likely a result of the posting issues that cross-posters faced after their posting rate limit was revoked by Twitter [21].

Given that 84.45% of the migrated users post completely different content on the two platforms, we next inspect the hashtags used. This gives us a flavour of the parallel discussions taking place on Mastodon and Twitter. Figure 15 presents the top 30 most frequent hashtags used over the two platforms by the migrated users. We notice that users discuss more diverse topics on Twitter such as Entertainment (#NowPlaying, #BBC6Music), Celebrities (#BarbaraHolzer), and Politics (#StandWithUkraine, #GeneralElectionNow), whereas Mastodon seems dominated by Fediverse related discussion (#fediverse) and the migration to it (#TwitterMigration). We conjecture that we might see more diverse discussions on Mastodon once the migrated users make themselves familiar with the platform.

Moderation on Mastodon has received significant attention in recent months [5, 4]. This is because the administrators of Mastodon instances do not universally have the resources to moderate malicius content. To shed light on this, we study the extent to which toxic content is shared by migrated users on both platforms. To do this, we label all tweets and statuses using Google Jigsaw’s Perspective API.⁷⁷7https://www.perspectiveapi.com For a given post, Perspective returns a score between 0 and 1 for its toxicity (0 $=$ non-toxic). Specifically, we use the API’s TOXICITY attribute that defines toxicity as “a rude, disrespectful, or unreasonable comment that is likely to make people leave a discussion”. In the literature, 0.5 is the most common choice to threshold the perspective scores [5, 22, 17], however, higher values such as 0.8 are also used [2]. Here, we use 0.5 as a threshold and consider a post to be toxic if its toxicity score is greater than 0.5 (and vice versa).

Figure 16 shows the CDFs of the fraction of each migrated user’s toxic posts on Twitter and Mastodon. Overall, just 5.49% of tweets are toxic. Mastodon is substantially less toxic, with just 2.80%. On average, each user posts 4.02% toxic tweets on Twitter vs. just 2.07% toxic statuses on Mastodon. Even though the discourse is non-toxic over both platforms, we notice that 14.26% of migrated users post at least one toxic post on both the platforms. While this may not be problematic for Twitter which has its own moderation team, it might present challenges for Mastodon, where volunteer administrators are responsible for content moderation [11].