Musk kept his word and published the code of Twitter's algorithm. This is the first time a major platform has revealed its algorithms. This appears to have been done to improve the effectiveness of the recommendations. I see this as a big threat and at the end I will show you how you can "hack the algorithm".
The article and algorithm analysis will focus specifically on the "For You" tab, i.e. the recommendation channel. Every time a user opens the tab, the system creates a list of the 1,500 most popular tweets. They are then split 50/50 between the accounts to which the user is subscribed and external sources. After filtering the content, a selection of 50 tweets is created that the user will see. And then it gets interesting.
Reach is influenced by the ratio of the number of subscribers to the number of outgoing subscriptions. This means that accounts with a high number of outgoing subscriptions and a small audience have less reach. Let's call them "mass followers", although this is incorrect.
Example: Profile A has 100 subscribers and 1000 outgoing subscriptions. Profile B has 1,000 subscribers and 100 outgoing subscriptions. Profile B is more important to the algorithm than profile A.
Twitter Blue users actually have priority coverage.
If the user subscribes to me, then the effect of the paid checkbox is x4, if the user does not subscribe, then x2 (relating to the user without Twitter Blue).
The algorithm extracts 4 types of pages (this code is no longer in the repository where it was originally located. It was either removed or added by mistake).
— experienced users (I don't know what this means, maybe they have been registered for a long time or have a large audience)
—Democrats
—Republicans
— Elon Musk (this is not a joke, it is a separate type of page)
Factors influencing whether a tweet will be displayed in the "For You" tab:
- if the user is likely to like/retweet this message
— opens your tweet and stays there for 2 minutes (why? Well, apart from reading replays, that explains the high thread coverage)
— goes to your account
But a photo or video in a tweet increases its visibility.
I found a mention in several sources that external links reduce visibility, but I didn't see it in the code and didn't understand it.
Additionally, words that the algorithm flags as "unknown language" receive a multiplier of 0.01, which is a ranking killer. This means that misspelled words and some new, complex terms can destroy your coverage. .
Additionally, in the same article we found references to a decline in the visibility of tweets about cryptocurrencies. I have not found.
What influences the decline in reputation and, consequently, reach:
- you are blocked
- you are muted
- complaints regarding violations
- spam complaints
- unsubscribes (but the impact is smaller than in the case of previous items)
There is an interesting detail in the code description. Twitter only counts unfollows for 90 days due to database limitations. And it does not give an eternal ban on shade.
How much do these factors affect coverage? VERY.
A user who clicks on your tweet and stays there for more than 2 minutes gains 22 times more weight than someone who simply liked your tweet. If they visit your profile via your tweet and like/reply to your tweet = 24 times more than a simple like. If they reply to your tweet? 54 times more than a like. If they reply to your tweet *and* you reply to their reply? 150 times more than a like.
I mean, I started a thread about SMM with about 20 points, users come to it, like it and read it. I respond to their responses = GIANT REACH. Simply amazing.
A tweet complaint has an effect 738 times greater than the effect of a like. It is not entirely clear whether all complaints are upheld or not.
Twitter's algorithm has a special condition for "state appeals." This means that it literally says that visibility can be changed at the request of the authorities.Algorytm zmniejsza trafność tweeta o 50% co 6 godzin.
How many tweets can you post effectively?
The system has a Tweepcred ranking parameter. It is the equivalent of Google PageRank, which is used to rank pages in search results. What he is doing?
Each Twitter account has its own score, or roughly speaking, the account's karma. It depends on the quantity and quality of its interactions with other users: the higher the score, the greater the influence in the heartless eyes of the social network that this account has. This means that its content is important and needed by others.
As far as I understand, the basic assessment is made based on simple parameters: account registration date, number of subscribers and device (xs).
The score is then adjusted based on the ratio of followers to followers.
The final score on a scale of 0 to 100 is the Tweepcred Score, which reflects a user's reputation on Twitter.
This score is used to determine which users you should recommend to follow or which user content you should highlight.
So what does tweet frequency have to do with it? And everything is simple, if Tweepcred is above 65, you can tweet as much as you want, everything will be covered. If it is below 65, the recommendation algorithm will only take 3 tweets per day.
The most important question: how to find your Tweepcred? No way. Maybe the craftsmen will cut out some bots for checking later, but they won't be able to check it now.
Why is revealing your algorithm bad?
We have reached the most interesting part. I was already against it when it was announced. What's the problem?
- Knowing exactly how the algorithm works, I can use it for my own purposes. My goals are noble, but someone else's are not so great.
- It turns out that it actually makes sense to recruit bots and then unfollow them en masse, it reduces the reach of the targeted account.
- We have to write threads. That was already clear, but now I would focus my entire Twitter content strategy solely on them. Gigantic reach.
Knowing the importance, knowing the influence of all factors, what prevents you from using them? Especially if my intentions are bad.
Comments