Reviewed and updated with current information and practical guidance.
Facebook and X, the service formerly called Twitter, use ranking and recommendation systems to help decide which posts people see. Understanding their broad operation can improve planning, but public documentation is not a complete formula for obtaining reach. A post can be accurate and useful without being widely recommended, and an interaction count does not tell you why the system distributed it.
This article originally discussed product announcements from 2015 and 2016. Those are useful history, but old settings such as “Show me the best Tweets first” should not be presented as current instructions. Start with the surface and account you are actually using, then distinguish official descriptions from marketing guesses.

What a feed ranking system is trying to do
A feed has more candidate content than a person is likely to inspect. Systems select and order some of it using available signals and predictions. Content eligibility, personal choices and the particular surface all matter. A recommendation is a prediction about possible interest, not a factual judgment that a post is true or suitable for every viewer.
Avoid talking about “the algorithm” as a single public score. Facebook Feed, Reels, Stories and other experiences can use different systems. Paid distribution has its own campaign objectives and auction processes. A change in organic reach does not automatically reveal a change in an ad system or a platform-wide rule.
How Facebook describes ranking today
Meta’s explanation of its AI system cards describes multiple systems and models. They can consider information about the content, the viewer’s previous activity and predicted interactions. The description includes examples of signals, not a set of stable numerical weights a marketer can optimize mechanically.
The practical implication is to make content useful for the intended task and evaluate it on the relevant surface. A helpful product demonstration, a conversation in a group and a status update serve different purposes. Do not assume that copying a high-view Reel into a link post reproduces its distribution.
What the 2016 Facebook article actually tells us
The historical announcement by Moshe Blank and Jie Xu discussed time spent viewing articles. It showed that visible likes or shares were not the only information Facebook considered. It does not establish the current weight of dwell time or a tactic that guarantees recommendation.
Contemporary coverage of reaction buttons belongs to that historical context too. Do not assign a universal multiplier to a reaction or infer that provoking anger is a sound strategy. A response can reflect disagreement, confusion or a service problem rather than approval.
Separate organic distribution from advertising
The older Facebook relevance-score announcement describes an advertising concept from an earlier period. It should not be treated as a current public score governing every organic post. Use the current advertising account’s documented metrics for the chosen objective, and keep them separate from Page content analysis.
For an ad review, record the campaign objective, audience settings, creative, destination and cost. For an organic review, record the publishing surface, content purpose and available distribution metrics. Combining them into one engagement total can hide the fact that additional exposure was purchased rather than earned through audience response.
A business should also distinguish relevance from suitability. Someone interacting with a broad topic may have no need for a particular product. Review qualified inquiries, complaints and completed actions so a high interaction rate is not automatically celebrated as success.

How X distinguishes For you and Following
The current X timeline guide distinguishes For you, which includes recommendations, from Following, which focuses on accounts the person follows. People can move between these views. A business cannot assume that all followers open the same view or see all posts in the same order.
X’s recommendation explanation describes multiple signals and says no one signal has a permanently assigned higher weight. The supported conclusion is that recommendations are personalized and multifactorial. It is not evidence for a fixed posting-time trick, reply quota or universal engagement formula.
Keep the older Twitter announcements in context
The original article linked to Mike Jahr’s profile, a 2016 timeline announcement and an older timeline help route. These are historical references rather than a verified menu path for the current app.
“While you were away” and “best Tweets first” explained earlier product choices. Do not instruct users to refresh away a batch of recommended posts or toggle a legacy option based on that description. Inspect the current account controls and current help before giving a colleague operational instructions.
Plan content around a useful audience question
Start with something the audience needs to understand or do. A hypothetical outdoor equipment store might demonstrate a tent’s setup and limits, then provide accurate product information. The task can be evaluated even if the clip never receives broad recommendation. A mystery caption that attracts irrelevant replies may perform worse for that purpose.
Use clear visual material, a readable explanation and an honest next action. The existing guide to X content and conversation ideas can help generate treatments, but ideas still need an audience purpose. For a service announcement, check whether a reader can identify the change, the people affected and the next step without opening another thread. A poll or video is optional; choose it only when it explains the information better. Check links and destination pages before distribution; more views cannot repair a broken booking route.
Make videos understandable in their actual context. Review captions, important visual information, audio quality and mobile readability. Use the current upload requirements rather than a copied historical chart. Accessibility and clarity help people use the content; they should not be sold as guaranteed ranking boosts.
Investigate changes without inventing a ranking explanation
When reach changes, record the date, surface, format, topic, paid support and measurement definition. Compare several relevant posts rather than one unusually successful example. A change in audience interest, competition, seasonality or campaign distribution could contribute to the result. The available report rarely isolates the ranking system as the cause.
Check account notices and recommendation or content-status information where available. Eligibility and actual distribution are different: being permitted to appear does not promise an impression. A policy issue needs the appropriate correction or appeal route, not a burst of extra posting to “reset the algorithm.”
Choose a bounded experiment
Test one meaningful creative question when practical, such as whether showing a real setup before explaining the materials improves understanding. Keep the task and distribution conditions reasonably comparable, and record the limitations. Do not conclude that the platform rewards every opening of that type from a small uncontrolled sample.
Review the quality of replies, completed customer tasks and staff effort alongside reach. Preserve content that remains useful, and update it when product details change. Keep a short decision record containing the creative question, observed result, alternative explanations and the next change. This makes learning usable by another editor without turning one person’s anecdote into a platform rule. The aim is a reliable communication process under uncertain distribution, not a claim that the feed has been fully decoded.
What marketers can control
You can control the accuracy of the claim, the quality of the explanation, the rights to the material and the working destination. You can choose a manageable publishing and reply schedule, and evaluate whether the audience found the content useful. You cannot control an individual’s feed or promise that a follower will receive a post.
Use current official descriptions to understand broad constraints and settings. Treat historical announcements as history, and treat observations from your own account as bounded evidence. That combination supports better decisions without turning an uncertain recommendation system into a set of invented laws.