X publicly released its entire recommendation algorithm yesterday, which has some nuances to explore beyond the well-known high-engagement scoring.

Instead of assigning a fixed total score to a post and pushing it out to everyone, the algorithm evaluates each individual reader to determine “what is this person most likely to do after seeing this post”: like, reply, quote, repost, privately share, copy the link, follow the author, or indicate “not interested,” mute, block, or report. The model predicts the probability of these actions and combines them into a ranking score based on different weights.

How Does This Algorithm Score Posts?

We can simplify it into a formula:

Base score of a post for a specific reader = Σ (predicted probability of the reader taking a certain action × weight of that action)

Furthermore, the system addresses issues such as excessive content from the same author, whether the post is from an account the reader follows, and the suitability of the content for display. In other words, a post may be very appealing to Reader A but may not necessarily have the same high score for Reader B.

In the publicly disclosed default parameters:
The weight for copying the link to share is 20,
significantly higher than the weights for replying and quoting at 5,
following the author at 4,
regular sharing at 2,
reposting at 1,
liking at 0.5.
clicking on the post is 0.4,
opening an external link is 0.2,
expanding images and watching videos are both 0.05.

This set of numbers is prone to misinterpretation. For example, an image does not automatically add 0.05 points; the model must first consider the reader's likelihood of expanding the image before including “probability of image expansion × 0.05” in the total score. Similarly, copying a link to share is not equivalent to twenty likes at once; it is a higher-weighted action with a typically lower occurrence probability. The algorithm always considers the product of probability and weight.

Therefore, simply pursuing likes is not the optimal strategy. Content that is worth sharing with colleagues, friends, or peers, or content that gives readers ample reason to write follow-ups, rebuttals, and references, aligns more with this ranking goal. The notion of “should” here is a content strategy judgment, but it directly stems from the weight attributed to sharing, replying, and quoting by the system.

In addition to the base score, the algorithm also reveals several easily overlooked realities. For the same round of candidate ranking, the first post from the same author competes normally, while subsequent posts suffer from author diversity decay: each additional post is multiplied by 0.5 by default, reaching a minimum of 0.25. Content from accounts the reader does not follow defaults to an external recommendation coefficient of 0.75; replies and reposts from accounts not followed are even more disadvantaged, with some being filtered out before ranking.

This explains a very common phenomenon: Posting five consecutive short threads on the same topic in a short period of time will not simply multiply exposure by five. What is really worth betting on is a complete, independent, and immediately understandable original main thread. This conclusion is an inference based on the content layer of public rules, rather than an additional system parameter.

Why “Useful” Is More Important Than “Engaging”

If the goal is to get readers to replicate a link or share via private message, then the most effective content is often not a general attitude, but something tangible: a judgment, a case study, a framework, a list, or a clear decision recommendation.

This is not because the algorithm gives extra points for the word "list." It's simply because a good list is easier to pass on to a specific person than a vague sentiment; a evidence-backed counterintuitive conclusion is also easier to quote and discuss than vague feel-good advice. In other words, content form is not the score, reader behavior is the score; content form only increases the likelihood of those positive behaviors occurring.

For example, the following two sentences express almost the same topic:

AI will transform the content industry.

I reviewed 30 AI content teams, and model capability is not important; breaking down topic selection, materials, and review into three separate stages is what matters.

The second sentence does not have any "data reward" from the algorithm. But it provides objects, samples, and debatable judgments. Those who have actually managed content teams can supplement with examples, present counterexamples, or pass it on to the person in charge of the process; these are the conditions under which replies, quotes, and shares are more likely to occur.

A high-exposure post is best positioned in the first sentence: It says who it's for, presents a specific judgment, and explains why the reader should keep reading. The body does not have to be long, but it should deliver enough information for the reader to gain value without leaving X. Links can serve as evidence or supplementary material, but the default public parameters do not show "external links lead to direct downranking"; a more cautious approach is to first deliver the core answer in the body, and then let the link provide the source and details.

The conclusion should also serve real discussion, not solicit engagement. "What do you all think?" easily leads to vague replies; "For those who have worked on B2B AI products, what is the most effective second-step activation action you have seen?" provides the right people with a clear entry point for a response. The weights of replies and quotes are both 5, which is public knowledge; that narrow questions are more likely to bring high-quality interactions than broad questions is a writing judgment.

Images, Videos, Links: To Use or Not to Use

Images, videos, and links are all forms of expression, not algorithmic talismans.

Both image expansion and video opening default weights are only 0.05. If a chart can make a growth trend, experimental results, or before-and-after comparison easy to understand at a glance, it is worth using; if it is just an unrelated visual to accompany text, it will not naturally receive a higher score. The same logic applies to videos: demonstrations, processes, comparisons, and on-site footage can help content be understood, but there is no evidence to suggest that simply posting a video will give exposure a boost.

The external link opening weight is 0.2, and there is no mention of an "external link penalty" in the open-source code. Therefore, there is no need to hide links to avoid so-called link weight reduction. More importantly, do not leave all the value of a post behind a link. A reader should be able to get conclusions, be willing to share or discuss within the body of the text; the link then becomes a supplement to deepen trust, rather than a reading threshold.

Behaviors That Will Lower Your Score

The default open-source weights are very harsh on negative feedback: not interested is -43.2, blocking the author is -31.2, muting the author is -58.8, and reporting is -234. These are also probabilities, not fixed deductions for each occurrence; however, they indicate that the algorithm would rather recommend fewer posts than push content that readers clearly dislike.

This leads to several practical writing consequences. If a title promises an amazing conclusion but the body of the text does not provide evidence or answers, it can easily make readers feel misled; attacking people to evoke emotions, jumping on unrelated hot topics, may trigger brief discussions but is more likely to lead to disinterest, muting, and reporting. Breaking down one piece of content into a series of posts without independent value also affects author diversity decay.

These "clickbait penalties" and "controversy for controversy's sake harming long-term exposure" are not verbatim rules in the source code but reasonable inferences based on negative feedback weights and diversity mechanisms. Their value lies in helping you determine what is not worth the risk rather than turning writing into a mechanical minefield.

Some Cases Do Not Result in Score Deductions, But Immediate Loss of Stranger Exposure

Beyond sorting, there are visibility filters. Open-source rules show that posts with junk content, malicious links, pornographic or explicit content, gory violence, hate speech, or severe insults may be directly removed when recommended to non-followers; accounts flagged as impersonation, compromised, restricted, or high-risk may also be removed. If readers have blocked or muted an author, the content may also become inaccessible in their recommendation stream.

This is not the same as losing a few likes. The former directly disqualifies a post from consideration. The repository also clearly states that some anti-spam rules are not public, so do not treat open-source code as a guide to circumventing detection. Activities such as inflating engagement, buying followers, participating in like-for-like groups, or automating interaction manipulation should not be attempted; the open-source code does not, and should not, be used to deduce which actions trigger certain tags.

There is another type of restriction that is not a security penalty but still limits the spread: the candidate process will remove posts older than 48 hours. Just because a piece of content did not find enough traction in its initial window does not mean it is worthless; however, it is not suitable to expect it to re-enter the For You feed widely days later. If necessary, present new developments, data, or insights in a new original post rather than reposting the old one verbatim.

So, How to Write a Good Post

The best approach is not to follow a "viral formula" but to accomplish four things in a standalone post: provide a specific conclusion, offer enough reasoning, deliver a takeaway for the reader, and finally pose a question worth answering by specific individuals.

For example:

I analyzed 30 AI product launch posts and found that what drives engagement is not "powerful features" but "which specific step it saves for whom."

Saying "We support Agent workflow" is hard to reshare.

Saying "The sales team used it to reduce customer research time from 2 hours to 12 minutes" is more likely to be passed on to colleagues.

A good product copy does not describe capability but illustrates the trouble a role is relieved from.

Have you encountered a product that excels in translating abstract capabilities into tangible benefits?

This example does not guarantee exposure since the algorithm also depends on the specific reader, account history, candidate competition, and experimental setup. However, it exhibits content characteristics aligned with the public goal: it is an independent original post; opens with a conclusion; provides a reshareable framework in the middle; and ends with an invitation for specific experiences rather than vague endorsements.

A truly sustainable strategy is not to chase after one viral hit but to repeatedly offer valuable insights worth resharing to the same group of people. The algorithm publicly predicts behavior based on reader personalization; consistent themes, genuine value, and healthy interactions will give the system more opportunities to place the next piece of content in front of those truly willing to respond, quote, and share.

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia