Twitter Algorithm 2023: How to rank well in For You feed

On March 31, Twitter's CEO Elon Musk made Twitter's algorithm open source. Everybody was surprised despite Elon tweeting so much about it. I also wrote a blog a few days back discussing this decision. In that blog, I mentioned that we would look deeper into the algorithm. So without any further ado, let's do it!

How does the algorithm work?

The Algorithm determines what should appear on your homepage, aka "For You" feed. It also calculates its sequence based on whether you like it.

The entire process is divided into three steps:

1. Fetch tweets that users might like from various sources.
2. Rank each tweet using a machine learning model.
3. Filter out tweets based on users' preferences, settings, etc.

First, it creates a pool of 1,500 tweets using candidate ranking. Half of these tweets include tweets from users you follow and interact with. The other half consists of tweets from your social graph and embedding spaces. Basically, Twitter collects tweets from users that overlap with your interests and the people you interact with. Additionally, it contains tweets from your interests and likings. It collects tweets that are similar to your interests.

The second step is ranking. A neural network with around 48 million parameters is used to rank tweets. The model analyzes each tweet individually and ranks it based on your interests, preferences, and previous interactions.

And finally, ranked tweets are filtered out by applying heuristics and filters. Twitter filters out tweets that don't follow Twitter's community guidelines, are spam, etc. It also removes tweets if you have blocked, muted, unfollowed, or reported the account. 

This is how you see your tweets on your homepage. Let's discuss factors that impact these decisions.

How to rank better on Twitter?

As the recommendation algorithm is open source, we can write better tweets that will end up in others' feeds.

Strive for "boosts"

The algorithm has hardcoded specific values that boost your tweet by predefined times. This depends on various factors like user interaction, negative factors, previous user history, etc.

• Likes: 30x
• If a user engages for more than 2 minutes: 22x
• Retweets: 20x
• If the tweet is from a trusted circle: 3x
• Twitter blue users: 2/4x
• Tweets with images and videos: 2x

Source

Use Twitter Blue

Now we understand that Twitter Blue was not a joke. The algorithm really favors tweets and users with an active Twitter Blue subscription.

If you want to grow on Twitter, Twitter Blue would significantly help you.

GitHub source

Don't get reported, muted, blocked, or unfollowed. 

If other users have a negative impression of you and your tweets, so does the algorithm! There are specific activities that negatively impact your reputation score. 

GitHub source

• Someone muted you when you appeared on their For You feed.
• Someone reports you for abuse or spam.
• Someone unfollows you 

Maintain your following-to-follower ratio

Twitter mentions reducing rankings of users who have less number of followers than their followers. It does this ranking through the "Tweepcred" page rank algorithm, which works similarly to Google's PageRank algorithm.

GitHub Source

Hence, work on those numbers if your follower count is less than the number of users you follow. This will increase your rank by a reasonable amount.

Include media, and avoid posting only links

It's now clear what works and what does not. Follow these methods:

1. Include images and videos in your tweets
2. Do not post anything with only links
3. Have a text included in your tweet, whether you are posting media or links
4. Create engaging content that users will actually read. 

Yet we are unsure of many things, like how posting external links affects its rank, whether bookmarking a tweet increases its boost, etc.

Closing Notes

Taking one of its most important algorithms and an ML model open source will help build trust in Twitter users. It would be interesting to see how users use this information to grow and engage on Twitter.