How Attention Is Allocated: Feed Ranking and the Science of What Spreads

Nothing in this course depends on which side of an issue you are on. The ranking systems and the diffusion research treat a water-conservation page, a property-tax-relief page, a library-levy page and a page opposing all three exactly the same way. The craft works for anyone willing to do it honestly, and the course says so here once. The beginner material on platform differences and post anatomy lives in the Journalism site's How the Platforms Differ and Anatomy of a Post. This lesson assumes you already run a page and want to know what the machine on the other side is doing.

What a feed is doing when it ranks

Meta's clearest public description is a January 26, 2021 Newsroom post by three of its Feed engineers, How Does News Feed Predict What You Want to See?. Strip away the example about a user named Juan and the mechanics are these. First comes the inventory: every post shared with the viewer by a friend, Group or Page since the last login, plus older unseen posts bumped back in and already-seen posts that have since drawn conversation. Integrity checks run on that pool. A lightweight model narrows it to roughly 500 candidates. Then the main scoring pass runs several separate models, each predicting one thing: the probability the viewer will like the post, comment on it, share it, and so on. Those predictions are combined into a single score "optimized for our primary objective," and the stated objective is long-term value, measured partly by asking people in surveys how meaningful an interaction was or whether a post was "worth their time." A final contextual pass applies rules such as not showing three videos in a row.

Two details in that document matter more than the rest for an advocacy page. One: the system scores each post for each viewer independently, so there is no single "reach" a post has; there are millions of per-viewer probabilities. Two: an action a given viewer almost never takes "automatically gets a minimal role in ranking" for that viewer. If your readers are people who read and never click like, the like prediction is nearly irrelevant for them and the comment or share prediction carries the weight.

Nick Clegg's June 29, 2023 post, releasing 22 system cards for Facebook and Instagram surfaces, repeats the frame with more candor: "no single prediction is a perfect gauge of whether a post is valuable to you," so the systems combine many, some from behavior and some from survey feedback. Meta also said it withholds some signals used to demote low-quality content so that people cannot game them. That is the honest limit of what you can know about Meta's ranking: the categories of predictions and the stated objective, never the weights.

Instagram's Adam Mosseri published a parallel account on May 31, 2023, Instagram Ranking Explained, which is useful because it separates surfaces. Feed predicts whether you will spend a few seconds on a post, comment, like, share or tap the profile. Explore weights likes, saves and shares most, and post information counts for much more there than in Feed. Reels predicts resharing, watching all the way through, liking and visiting the audio page. Mosseri also stated outright that low-resolution or watermarked reels are made less visible, which is a rule rather than a prediction and one of the few hard content penalties Meta has put in writing.

Surface	What the platform says it predicts	Source and date
Facebook Feed	Probability of like, comment, share and other interactions; weighted by survey answers about meaningful interactions and "worth your time"	Meta Newsroom, Jan. 26, 2021; system cards, June 29, 2023
Instagram Feed	Spend a few seconds, comment, like, share, tap profile	Instagram, May 31, 2023
Instagram Explore	Likes, saves and shares most important; post information outweighs relationship	Instagram, May 31, 2023
Instagram Reels	Reshare, watch all the way through, like, visit the audio page	Instagram, May 31, 2023
X For You (2023)	Ten probabilities: like, retweet, reply, good profile click, video played to 50 percent, reply engaged by author, good click, good click lasting 2 minutes, negative feedback, report	Heavy Ranker README, April 5, 2023
X For You (2026)	Engagement (favorite, reply, repost, quote, share, share via DM, share via copy link), clicks, attention (dwell, video quality view), follow author, and negatives (not interested, mute, block, report, not dwelled)	xai-org/x-algorithm README and param.rs, 2026
TikTok For You	User interactions (likes, shares, follows, comments), video information (captions, sounds, hashtags), device and account settings at lower weight; watching a longer video to the end is a strong signal	TikTok Newsroom, June 18, 2020
YouTube Home and Up Next	Clicks, watch time, survey responses on a 1 to 5 star scale, shares, likes, dislikes; "valued watchtime" counts only videos rated 4 or 5	YouTube blog, Sept. 15, 2021

Meta: what it has said about Pages, conversation and bait

Three Meta documents, read together, explain most of what an advocacy Page experiences on Facebook.

On January 11, 2018, Adam Mosseri, then head of News Feed, announced that Feed would "prioritize posts that spark conversations and meaningful interactions between people," predicting which posts a person might want to discuss with friends. He said plainly that "Pages may see their reach, video watch time and referral traffic decrease," and that "Pages making posts that people generally don't react to or comment on could see the biggest decreases in distribution." That sentence is the structural fact of Facebook for every organization since: content from Pages is ranked against content from friends, and it wins only when it starts conversation between people.

A month earlier, on December 18, 2017, Meta had defined engagement bait, posts that ask people to like, share or comment to gain distribution (its example: "LIKE this if you're an Aries!"), and said it would demote individual posts that use it and apply stricter demotions to Pages that do so repeatedly. Later updates extended the demotion to comments containing bait (November 13, 2018) and to videos with bait in the audio (January 30, 2019). So the 2018 instruction to generate conversation comes with a 2017 rule against asking for it. The craft lives in that gap, and the steps below are about working inside it.

The third document is the Widely Viewed Content Report, which Meta began publishing quarterly in August 2021. The introductory post reported that "fewer than 13% of content views were on posts with links" and that 57 percent of what people saw came from family and friends. The report lists the top 20 domains, links, Pages and posts by US Feed views each quarter. Treat it as a reality check on scale: the whole category of link posts is a small slice of Feed, and the top 20 lists show what kind of post actually reaches the most people in a quarter. The current quarter's figures render only in a browser on the Transparency Center page; the numbers quoted here are from the 2021 launch, and you should read the latest quarter yourself before citing a current share.

X: the only platform that published the weights

In March 2023 X (then Twitter) open-sourced its recommendation pipeline and described it in an engineering post: about 1,500 candidate posts pulled per request, "50% In-Network Tweets and 50% Out-of-Network Tweets on average," a "~48M parameter neural network" called the heavy ranker that "outputs ten labels" each representing the probability of an engagement, then heuristics and filters, then blending with ads and follow suggestions. The heavy ranker README gave the weights in force on April 5, 2023, with a warning that they "can be adjusted at any time."

In 2026 X published a rewritten system at xai-org/x-algorithm, with notable updates dated August 13, August 14 and September 18, 2026. The ranking model is now a transformer called Phoenix that reads the viewer's recent engagement history and predicts, per post, the probability of each action. The weights live in home-mixer/params/param.rs. As of the copy read on October 16, 2026, they were:

Predicted action	Weight, April 5, 2023 (heavy ranker)	Weight, 2026 (Phoenix, param.rs)
Like (favorite)	0.5	0.66
Repost (retweet)	1.0	1.0
Reply	13.5	5.0, plus a 15.0 boost when the viewer and author follow each other
Reply that the author then engages with	75.0	Not a separate label in the 2026 file
Quote	Not a label	5.0
Share (general)	Not a label	2.0
Share via direct message	Not a label	5.0
Share via copy link	Not a label	20.0
Follow the author	Not a label	4.0
Click into the post	11.0 ("good click"), 10.0 (stay 2 minutes)	0.3 (click), 0.4 per unit of click dwell time
Open a link	Not a label	0.2
Profile click	12.0 (then like or reply)	0.0
Video watched to 50 percent / quality view	0.005	0.0 (quality view), 0.02 (video open)
Dwell on the post	Not a label	0.05, plus 0.006 per unit of continuous dwell time
Not interested	Within -74.0 negative feedback	-52.272
Mute author	Within -74.0	-58.8
Block author	Within -74.0	-31.2
Report	-369.0	-234.0
Not dwelled (scrolled past)	Not a label	-0.02

Read the table with the caveat X itself added on August 14, 2026, after watching people misread it. The weights multiply the predicted probability that this viewer will take that action; they do not multiply raw counts. A report weighted 234 times a like does not mean one report cancels 234 likes. It means the baseline probability of a report is, in X's words, "more than 1000x lower than a Like," so the weight has to be large for the prediction to affect ranking at all. The same comment in param.rs says that mass blocking or reporting by a coordinated group mostly affects recommendations for users similar to that group, and that actions taken on a post someone opened directly, rather than one served in the Home timeline, have no ranking effect.

Still, the shape of the weights tells you what X wants. A post that makes people reply, quote, and copy the link to send somewhere else is worth many times a post that collects likes. Three adjustments follow the weighted sum: an author diversity decay (each post after an author's first in a session is multiplied down toward a floor), an out-of-network discount (posts from accounts the viewer does not follow, and replies and reposts from followed accounts, are multiplied by a factor below 1), and a new-author boost for authors with few impressions. A pre-scoring filter drops any post older than 48 hours from For You, so a post's recommendation life on X is two days.

TikTok and YouTube: watch time and satisfaction

TikTok's one public description, a June 18, 2020 Newsroom post, lists user interactions (videos liked or shared, accounts followed, comments posted, content created), video information (captions, sounds, hashtags) and device and account settings, with the last group weighted low because people do not choose them as preferences. It says a strong indicator such as "whether a user finishes watching a longer video from beginning to end" outweighs a weak one such as sharing a country with the creator. And it states that "neither follower count nor whether the account has had previous high-performing videos are direct factors." The feed avoids two videos in a row from the same creator or sound, mixes in some videos outside a viewer's stated interests, and does not recommend content under review, duplicates or spam. TikTok's recommendation description has not been materially updated in public since 2020; lesson 03 covers the 2026 US ownership change and what it claims about retraining.

YouTube's account, by engineering vice president Cristos Goodrow on September 15, 2021, is the most historically honest of the set. Ranking by clicks produced clickbait, so in 2012 YouTube switched to watch time and accepted "an immediate 20% drop in views." Watch time alone rewarded videos people regretted watching, so YouTube added surveys: viewers rate a video from one to five stars, a model predicts the ratings of people who do not answer, and "only videos that you rate highly with four or five stars are counted as valued watchtime." Shares and likes count as satisfaction, dislikes as dissatisfaction, and the weight of each varies by viewer. YouTube also demotes what it calls borderline content, using human evaluators to classify videos, and reported that after a 2019 change, US watch time on non-subscribed recommended borderline content fell 70 percent. The same post says YouTube "haven't seen evidence that borderline content is on average more engaging," which is a company claim you should weigh against the independent studies in the next section.

The science of what spreads

The platform documents tell you what is being predicted. The research tells you which content tends to win those predictions. Keep two distinctions in mind as you read: whether a study is observational (correlation in real posts) or experimental (random assignment), and whether its outcome is attention (shares, clicks) or belief (attitudes, knowledge). Almost all of what follows is about attention.

Arousal, not valence. Jonah Berger and Katherine Milkman, What Makes Online Content Viral? (Journal of Marketing Research, 2012), coded 6,956 New York Times articles from the home page between August 30 and November 30, 2008 and modeled which made the most-emailed list. Positive content was more viral than negative on average, and the real driver was physiological arousal. A one standard deviation increase in the anger an article evoked raised the odds of making the list by 34 percent; the same increase in awe raised it 30 percent. Sadness, a low-arousal emotion, reduced sharing. Three small lab experiments (141 participants in total) manipulated emotion directly and found the same pattern, with arousal mediating the effect. For an advocacy page the finding cuts both ways: anger works, and so does awe, and awe does not cost you your reputation.

Moral-emotional words. William Brady and colleagues, Emotion shapes the diffusion of moralized content in social networks (PNAS, 2017), analyzed 563,312 tweets about gun control, same-sex marriage and climate change. Each additional moral-emotional word in a tweet (words that are both moral and emotional, such as "shame" or "fight" in their lexicon) increased its retweets by about 20 percent. The spread was "bounded by group membership": the words moved messages within liberal and within conservative networks, much less between them. Correlational, large, and replicated in later work by the same group.

Out-group hostility. Steve Rathje, Jay Van Bavel and Sander van der Linden, Out-group animosity drives engagement on social media (PNAS, 2021), examined 2,730,215 Facebook and Twitter posts from US news media accounts and members of Congress. Posts about the political out-group were shared or retweeted about twice as often as posts about the in-group. Each term referring to the out-group raised the odds of a share by 67 percent. Out-group language was the strongest predictor in the model, about 4.8 times the effect of negative affect words and 6.7 times the effect of moral-emotional words, and it was the strongest predictor of "angry" reactions, the most common reaction in every dataset. The effect held on both platforms and for both parties and was larger for politicians than for news outlets.

Negative headline words, tested causally. Claire Robertson and colleagues, Negativity drives online news consumption (Nature Human Behaviour, 2023), used the Upworthy archive of headline A/B tests: 22,743 randomized trials covering about 105,000 headline variants, 5.7 million clicks and more than 370 million impressions. Because Upworthy randomly assigned headlines to visitors, this is causal. For a headline of average length, each additional negative word raised the click-through rate by 2.3 percent, and positive words lowered it. The effect is real and small per word, which is worth remembering when someone tells you negativity is the only thing that works.

Falsehood travels farther. Soroush Vosoughi, Deb Roy and Sinan Aral, The spread of true and false news online (Science, 2018), traced about 126,000 fact-checked stories tweeted by about 3 million people more than 4.5 million times between 2006 and 2017. Falsehood "diffused significantly farther, faster, deeper, and more broadly than the truth" in every category, most of all in politics. False stories were more novel and drew replies expressing fear, disgust and surprise; true stories drew anticipation, sadness, joy and trust. Bots spread true and false news at the same rate, so the gap came from people. For an advocacy page committed to accuracy this is the hardest finding in the lesson: the honest post is structurally handicapped against the sensational false one, and the compensating strategy has to come from elsewhere (novelty of true facts, documents, and the network you build in lesson 02).

Most "viral" things are broadcasts. Sharad Goel, Ashton Anderson, Jake Hofman and Duncan Watts, The Structural Virality of Online Diffusion (Management Science, 2016), measured a billion diffusion events on Twitter (news, videos, images, petitions) and found that structural virality is "typically low, and remains so independent of size." Popularity was mostly driven by the size of the largest broadcast, meaning one big account or outlet posting it, with person-to-person chains contributing less than the word "viral" implies. Practical reading: getting one large account to carry your post usually does more than any number of small reshares, which is why cross-promotion is the engine in lesson 06.

Study	Design	Sample	Outcome measured	Effect reported
Berger and Milkman 2012, JMR	Observational plus 3 lab experiments	6,956 NYT articles; 141 participants	Making the most-emailed list; stated sharing intent	+34 percent odds per SD of anger; +30 percent per SD of awe; sadness reduces sharing
Brady et al. 2017, PNAS	Observational	563,312 tweets, 3 issues	Retweets	About +20 percent per moral-emotional word; stronger within ideological group
Vosoughi, Roy and Aral 2018, Science	Observational	About 126,000 stories, 3 million users, 4.5 million tweets	Cascade size, depth, speed	False news spread farther and faster in every category, most in politics; bots neutral
Rathje et al. 2021, PNAS	Observational	2,730,215 posts, news media and Congress	Shares, retweets, reactions	+67 percent odds of a share per out-group term; out-group posts shared about twice as often
Robertson et al. 2023, Nature Human Behaviour	Randomized headline tests	22,743 trials, about 105,000 headlines, 370 million impressions	Click-through rate	+2.3 percent CTR per negative word at average headline length
Goel et al. 2016, Management Science	Observational	About 1 billion diffusion events	Structural virality	Typically low; popularity driven by largest broadcast
Milli et al. 2025, PNAS Nexus	Preregistered audit, engagement ranking vs chronological	X users and their timelines	Content amplified; stated preference	Engagement ranking amplified emotionally charged, out-group hostile political content that users said they did not prefer

What changing the feed did and did not do

In 2020 Meta gave a team of outside academics the ability to run experiments on consenting US users of Facebook and Instagram during the presidential election, with preregistered analyses and no Meta veto over findings. The first papers appeared in Science and Nature in July 2023. They are the best causal evidence available on the question most people argue about from intuition, and they point one way.

Andrew Guess and colleagues, How do social media feed algorithms affect attitudes and behavior in an election campaign?, assigned consenting users to reverse-chronological feeds for three months. Those users spent "substantially" less time on the platforms and did less. Their exposure changed: more political content and more content from untrustworthy sources on both platforms, less uncivil content and fewer slurs on Facebook, more content from moderate friends and ideologically mixed sources. And yet "the chronological feed did not significantly alter levels of issue polarization, affective polarization, political knowledge, or other key attitudes."

The same team's companion paper, Reshares on social media amplify political news but do not detectably affect beliefs or opinions, removed all reshared content from a random set of Facebook feeds for the same period. Political news exposure fell, including from untrustworthy sources; clicks, reactions and partisan news clicks fell; news knowledge fell within the sample. Polarization and individual attitudes did not change.

Brendan Nyhan and colleagues, Like-minded sources on Facebook are prevalent but not polarizing (Nature), cut exposure to like-minded sources by about one third for 23,377 users. Cross-cutting exposure rose and uncivil language fell, with "no measurable effects on eight preregistered attitudinal measures." The authors call the null results precisely estimated, which is the polite way of saying the effect was not hiding in noise.

Sandra Gonzalez-Bailon and colleagues, Asymmetric ideological segregation in exposure to political news on Facebook, used aggregated data on 208 million US users and found that segregation rises at each step from what could have been seen, to what was seen after ranking, to what was engaged with. That is the one result in the set that implicates ranking directly, and it is about exposure rather than attitudes.

Outside Meta's data, Smitha Milli and colleagues ran a preregistered audit of X in Engagement, user satisfaction, and the amplification of divisive content on social media (PNAS Nexus, 2025). Relative to a chronological baseline, X's engagement ranking amplified "emotionally charged, out-group hostile content that users say makes them feel worse about their political out-group," and users did not prefer the political posts the algorithm chose. Ranking by stated preference instead reduced angry and out-group hostile content, with a possible cost of more content that agreed with the viewer. This audit measures what is amplified and what people say they want, which is exactly the gap Meta's and YouTube's surveys try to close from the inside.

What to take from the set. The feed shapes what you see and how long you stay, and it rewards the emotional registers the earlier studies describe. It does not, on the evidence of the largest experiments ever run, convert people in a season. For an advocacy page that means attention is the thing ranking can give you, and the uses of attention (mobilizing people who already agree, setting the agenda, building the relationships that persuasion research says do the work) are a separate design problem. The persuasion evidence is in lesson 09; the theory of change that connects the two is in lesson 02.

Running a page that refuses to run on outrage

You now have the pieces. Ranking rewards predicted replies, shares and conversation, and penalizes predicted reports, mutes and "not interested." The research says out-group hostility and anger generate those predictions reliably. The research also says awe, novelty and usefulness generate them, that moral-emotional language works inside your own community, and that the platform satisfaction surveys (Meta's "worth your time," YouTube's four and five star ratings) are explicitly designed to pull ranking away from content people engage with and regret. A page can be built on the second set.

Meta's political content limits, 2024 to 2025

On February 9, 2024, Instagram announced that "we won't proactively recommend content about politics on recommendation surfaces" from accounts people do not follow across Instagram and Threads, covering Explore, Reels, In-Feed Recommendations and Suggested Users. Political content was described as "potentially related to things like laws, elections, or social topics." The change did not affect content from followed accounts, and users were given a control to opt back in. For an advocacy account that had grown through Reels and Explore, 2024 was the year non-follower reach fell off a cliff, and the cause was a documented policy rather than a mysterious penalty.

On January 7, 2025, Joel Kaplan's Newsroom post More Speech and Fewer Mistakes reversed course. Meta said it would "start phasing this back into Facebook, Instagram and Threads" with a personalized approach, ranking political content using explicit signals such as liking a post and implicit signals such as viewing posts, so that "people who want to see more political content in their feeds can," and would treat "civic content from people and Pages you follow on Facebook more like any other content." The same post ended the US third-party fact-checking program. Instagram's 2024 announcement now carries update notes dated January 7 and May 28, 2025 pointing to the Newsroom post and to the Transparency Center's Our Approach to Political Content page, which is where the current rules are maintained. Read that page before you plan a quarter; the policy has flipped once in eleven months and can flip again. The practical consequence for a page is that recommendation reach for civic posts in 2025 and 2026 depends on each viewer's own history of engaging with political content, which rewards the slow work of building a following that chooses to see you.

Questions and answers

Does Meta publish its ranking weights the way X does?

No. Meta publishes the categories of signals and predictions, through the 2021 Newsroom post and the 2023 system cards, and says it withholds some integrity signals so they cannot be gamed. X is the only major platform that has released the actual weights, in April 2023 and again in 2026 at xai-org/x-algorithm. Treat any specific Meta weight you see quoted as a guess.

If a reply is worth 5.0 and a like 0.66 on X, should I ask people to reply?

Asking is the problem. Meta's 2017 engagement bait policy demotes posts that ask for likes, shares or comments, and X's negative weights for "not interested" and reports are the mechanism by which readers punish posts that feel like bait. The weights reward predicted replies, so write posts that give people something to say: a question with a use, a document that invites interpretation, a claim specific enough to be checked. Then reply to them. The reply-engaged-by-author label was the single highest positive weight (75.0) in the 2023 X code.

Is awe really as strong as anger?

In Berger and Milkman's New York Times data, nearly: a one standard deviation increase in awe raised the odds of making the most-emailed list by 30 percent against 34 percent for anger, and their lab experiments showed arousal, not valence, carried the effect. The later field studies (Brady, Rathje, Robertson) focus on moral, out-group and negative language because that is what dominated political posts, and they found strong effects. The two sets of findings are compatible: anger is the easy route; awe and novelty are an available route that most political pages leave unused.

What does "the effect was bounded by group membership" mean for my page?

Brady et al. found moral-emotional words increased retweets within liberal and within conservative networks more than between them. Your morally charged posts will travel among people who already share your frame and stop at the boundary. If your theory of change (lesson 02) needs you to reach people outside that boundary, you need a second register for them, usually plain description, documents and local specifics.

Did the 2020 Facebook experiments show that algorithms do not matter?

They showed that three large changes to the feed over three months did not move polarization, knowledge or attitudes among consenting US users, while changing what people saw and how long they stayed. The authors and critics have both noted the limits: three months, a single election period, a platform already saturated with political information from many sources, and consenting participants. Gonzalez-Bailon et al. found ranking did increase ideological segregation of exposure. The fair summary is that the feed shapes attention measurably and attitudes much less than popular accounts assume.

Why did my Instagram non-follower reach collapse in 2024 and partly return in 2025?

Because of a written policy. In February 2024 Instagram and Threads stopped recommending political content from accounts people did not follow, with an opt-in control. In January 2025 Meta announced it would phase political content back into recommendations using each viewer's explicit and implicit signals. Your recommendation reach now depends on the political-content history of each potential viewer, which you cannot control, and on your follower base, which you can. The current policy lives on Meta's Our Approach to Political Content page.

TikTok says follower count is not a direct factor. Should I care about followers there?

TikTok's 2020 statement says neither follower count nor past high-performing videos are direct ranking factors, while conceding that accounts with more followers get more views because more people have chosen to see them. Followers on TikTok are a floor for distribution and a signal of fit, and each video still has to earn its own completion rate. The platform-by-platform treatment is in lesson 03.

How often do the weights and policies change?

X's 2023 README said weights "can be adjusted at any time," and the 2026 file shows that the like weight, report weight and the whole label set changed between releases. Meta's political content policy flipped between February 2024 and January 2025. Plan to re-read the four primary sources listed under Go further at least once a quarter, and treat any specific number in this lesson as dated October 2026.

Hard questions

Isn't learning the ranking signals just learning to manipulate people?

Learning how a system allocates attention is the same work a newspaper does when it studies which page the reader opens first, and the same work an opponent of your position is doing right now. The manipulation question is about the content, not the knowledge. The research in this lesson gives a clean test: the strategies that depend on out-group hostility and false novelty are the ones Rathje, Vosoughi and Milli identify as amplified beyond what users say they want, and the platforms' own satisfaction surveys exist to pull against them. A page that earns replies with usable questions, shares with useful documents, and awe with true surprises is working with those surveys rather than against the audience. Lesson 09 has an ethics code you can adopt that draws the line in writing.

Doesn't engagement optimization inherently reward outrage, so an honest page is doomed?

It rewards arousal, and outrage is one high-arousal emotion among several. Rathje et al. measured the out-group effect at 67 percent per term, which is large, and Milli et al. confirmed X's ranking amplifies hostile content users dislike. Against that, Berger and Milkman found awe nearly as powerful as anger, Vosoughi et al. found novelty was the mechanism behind false news (a true document nobody has seen is also novel), and every platform now includes satisfaction or negative-feedback predictions with large weights. An honest page is handicapped on the cheapest route to attention. It is not shut out, and the handicap is smaller on YouTube, where valued watch time counts only videos rated four or five stars, than on a chronological reply feed. Pick platforms accordingly in lesson 03.

If the feed does not change attitudes, why spend effort on social media at all?

Because the experiments measured attitudes, and advocacy mostly needs other outcomes: people who already agree showing up to a hearing, a reporter picking up a document, a legislator seeing that constituents are watching, a list of names that can be emailed next month. Bond et al.'s 61-million-person experiment in lesson 02 found a small but real effect on actual voting, carried through close friends. The feed is a distribution channel for mobilization and agenda setting, which is what lesson 07 and the measurement plan in lesson 08 are built around. Expecting conversion from a feed post is the mistake, and the 2023 experiments are the reason to stop making it.

X published its code, so can't a competitor just reverse-engineer reach?

Partly, and X's own August 2026 note explains why not fully. The weights multiply predicted probabilities that come from a model trained on each viewer's engagement history. Knowing that a report weighs -234 tells you nothing about the probability the model assigns to your post for a given reader, and X says coordinated reporting mostly affects recommendations for users similar to the reporters. The code also shows filters nobody can route around: a 48-hour age limit, author diversity decay, an out-of-network discount. The public weights are a guide to what the platform values, which is what this lesson uses them for, and no more.

Aren't these studies out of date? Twitter data from 2017 says nothing about 2026.

Some of it is dated and the lesson says so. Brady (2017) and Vosoughi (2018) used Twitter data from before the 2023 and 2026 ranking changes. Rathje (2021) covers Facebook and Twitter. Robertson (2023) used Upworthy headline tests from 2013 to 2015. What has held up is the mechanism: each later study, including Milli's 2025 audit of X and the 2023 Meta experiments, finds the same emotional and group dynamics in the amplified content. The weights in the 2026 X code reward the same actions (replies, shares, follows) that the older studies found hostile content produces. Until a well-designed study finds the pattern has broken, the conservative reading is that it has not.

Why trust platform documents at all? They are public relations.

They are, which is why the lesson pairs each with independent evidence and quotes them for what they commit to rather than what they boast. YouTube's claim that borderline content is not more engaging is a company assertion; Milli's audit of X and Rathje's cross-platform study are the counterweight. X's code is the exception: it is the running system, not a description of it, and anyone can read it. Where a platform statement is the only source, as with TikTok's 2020 description, the lesson tells you it is the only source. The habit to keep is the one in the reporting box above: ask which document, and then ask what independent study checked it.