The Two X Timelines — How Following and For You Work Differently
X displays tweets across two structurally different timeline surfaces in 2026, and creators routinely misread their own growth data because they confuse the two. The Following tab is a strict reverse-chronological feed restricted to accounts the viewer already follows. There is no machine-learning re-ranking, no candidate sourcing from outside the follow graph, and no engagement-weighted scoring that decides what surfaces. If you have 5,000 followers, every tweet you publish is eligible to appear in the Following tab of those 5,000 viewers — bounded only by how recently they opened the app and how many newer tweets have pushed yours down the chronological stack.
The For You tab is the algorithmic surface where almost all growth on X actually happens. Its candidate pool is not limited to accounts the viewer already follows — it explicitly sources tweets from outside the follow graph, ranks them against an internal score, and renders the top-scoring candidates. A creator’s reach ceiling on the Following tab is hard-capped by follower count; the reach ceiling on For You is, in principle, unbounded. This is the structural reason why engagement rate rather than follower count is the primary determinant of organic growth on X, and it is the reason understanding For You is the highest-leverage piece of platform knowledge any creator on X can acquire.
The two surfaces also feed each other. A tweet that performs strongly on For You drives profile visits, which convert into Following-tab followers, who then become more likely to engage with future tweets in the early velocity window — which in turn improves For You candidate eligibility for later tweets. Conversely, a creator whose For You distribution stalls sees follower growth flatten, engagement rate dilute, and Following-tab reach plateau. The full picture of how these surfaces interact is mapped out in our complete X algorithm and growth guide, which serves as the hub for every article in this cluster.
X’s Candidate Sources System — How Tweets Enter the For You Candidate Pool
On 15 May 2026, X open-sourced the production For You recommendation pipeline at github.com/xai-org/x-algorithm, replacing the partial 2023 release. The 2026 repository exposes the actual candidate sourcing architecture used by the live system — and that architecture is the missing piece most creator-side explanations of X have been guessing at for two years.
The candidate pool for any single viewer’s For You feed is assembled in two phases. The first phase is candidate sourcing — assembling roughly 1,500 candidate tweets the system will consider showing. The second phase is ranking — passing each candidate through the heavy ranker (the neural network that scores how likely the viewer is to engage) and rendering the top-scoring tweets.
Candidate sourcing splits the 1,500 candidates into two streams. Approximately 50 percent come from in-network sources — tweets from accounts the viewer follows, or tweets engaged with by accounts the viewer follows. The remaining 50 percent come from out-of-network sources — tweets the system believes the viewer might engage with based on topic similarity, engagement velocity, and social graph proximity. This 50-50 split is the architectural reason a creator with zero followers can still reach a viewer in the For You tab, and it is the reason follower count is a weaker growth signal on X than on platforms with stricter in-network bias.
The out-of-network candidate sources fall into three sub-systems. SimClusters assigns viewers and tweets to roughly 145,000 latent topic communities and surfaces tweets engaged with by other viewers in the same SimCluster. TwHIN (the Twitter Heterogeneous Information Network) generates embeddings of users and tweets, then retrieves candidates with similar embeddings to those a viewer has recently engaged with. Real Graph models the strength of relationship between any two accounts based on historical mutual engagement — strong Real Graph ties between an out-of-network author and someone the viewer follows lift that author’s tweets into the candidate pool. These three systems run in parallel, and their outputs are merged before ranking.
The practical implication for creators is structural: your tweets reach non-followers when accounts similar to them engage with you. SimClusters reward topical consistency. TwHIN rewards predictable engagement patterns. Real Graph rewards being engaged with by accounts that are themselves well-connected. The whole machinery is engagement-driven, not follower-driven — which is why our breakdown of how engagement signals are weighted in X’s ranking system is the natural next read once the candidate sourcing model is clear.
The 5 Ranking Signals — How X Decides What Reaches Non-Followers
After the 1,500-candidate pool is assembled, every candidate is scored by the heavy ranker — the neural network whose actual weight structure was confirmed in the 15 May 2026 open-source release. The heavy ranker predicts the probability of ten possible viewer reactions to a tweet (reply, retweet, like, click, dwell, profile visit, bookmark, video play, report, hide). Each predicted probability is multiplied by a fixed weight, and the products are summed to produce the tweet’s final ranking score.
The five families of signals below collapse the 100-plus features in the heavy ranker into the categories creators can actually act on. The numerical weights come directly from the May 2026 repository’s heavy-ranker configuration files.
Signal 1 — Engagement Type Weighting
The heavy ranker assigns radically different weights to different predicted engagement types. The 2026 open-source release confirmed the production weights creators had inferred from observation:
- Predicted reply: weight ~13.5 — by far the highest-weighted single signal in the entire ranker.
- Predicted retweet (repost): weight ~1.0 — the second-strongest amplification signal.
- Predicted favourite (like): weight ~0.5 — a positive but minor signal.
- Predicted dwell (extended reading time): weight ~0.5 — comparable to a like for long-form content.
- Predicted profile visit from tweet: weight ~12.0 — almost as strong as a reply, and underused as a deliberate optimisation target.
- Predicted bookmark: weight ~10.0 — added since the original 2023 release, treated as save-for-later intent.
- Predicted negative actions (report, hide, mute): weight -74.0 — the largest single coefficient in the model and the reason a single hostile reaction can suppress a tweet decisively.
The headline implication: replies are worth roughly 27 likes, and bookmarks are worth roughly 20 likes. A tweet that generates 5 replies and 50 likes will outscore a tweet that generates 0 replies and 500 likes — frequently by a wide margin. Creators chasing like-volume on the assumption that all engagement is equivalent are systematically optimising the wrong signal.
Signal 2 — Engagement Velocity
The heavy ranker scores candidates using a freshness-decayed engagement signal, with the decay applied as a multiplicative penalty after roughly 30 minutes of publication. Tweets that accumulate strong reply or bookmark engagement inside the first 15-30 minutes enter wider out-of-network candidate eligibility — the SimClusters and TwHIN systems use early engagement velocity as a strong predictor of out-of-network appeal. Tweets that accumulate engagement only after the velocity window has closed are eligible for in-network distribution to the original author’s followers but rarely break out to non-followers.
This is the structural reason posting time matters more on X than on TikTok or YouTube. On TikTok, the For You Page distributes content over multiple days. On X, your candidate eligibility for non-follower distribution is largely decided in the first 30 minutes, then declines steeply, and is functionally exhausted by hour 24.
Signal 3 — Account Credibility (Trust Score)
X maintains an internal account trust score that the heavy ranker uses as an input feature. The score is derived from follower-following ratio, account age, prior policy actions, ratio of engagement that the system flags as suspicious, the trust scores of accounts that engage with you, and several behavioural fingerprints visible in the open-source code. Low-trust accounts can still publish tweets that enter the ranking pipeline, but their candidate eligibility for out-of-network sources is heavily reduced — meaning their tweets effectively reach only existing followers.
Trust score recovery is slow but possible. The pipeline does not bake in any permanent penalty, but trust score is computed over rolling windows of behaviour, so visible patterns (a clean engagement profile, no policy actions, gradual organic-looking follower growth) shift the score gradually over several months. For creators who have used aggressive growth tactics in the past, the trust score is usually the limiting factor in their current For You distribution, not their content quality.
Signal 4 — Content Format
The 2026 open-source release confirmed several format-specific multipliers in the candidate sourcing and ranking stages.
- Native video tweets receive preferential candidate sourcing because video drives session time, which the broader X system optimises against.
- Long-form posts (over 280 characters, currently Premium-only) generate more reply engagement per impression than standard tweets and feed back into the engagement-rate-based candidate eligibility.
- Image carousels and single images are treated approximately equivalently to plain text in ranking, but slightly outperform on dwell time.
- Tweets containing external links are systematically suppressed in the heavy ranker — the open-source release exposes the explicit feature flag that downranks any tweet whose body contains a non-X URL. This is the single most important format penalty creators routinely underestimate.
- Threads have a special property: each tweet in a thread is ranked independently as its own candidate, so a 10-tweet thread generates 10 separate candidate-pool entries.
Signal 5 — Premium Subscriber Multiplier
The 2026 open-source code exposes a Premium-tier multiplier in the heavy ranker’s final scoring stage. The multiplier is small (single-digit percentage points on the ranking score) but applies across every tweet a Premium subscriber publishes, and the multiplier has widened in each quarterly release of the algorithm — a pattern visible by comparing the 2023 weights to the 2026 weights. The multiplier interacts with reply positioning rather than directly with For You candidate eligibility: a Premium subscriber’s reply to a high-engagement tweet appears above non-Premium replies in the conversation thread, generating substantial extra visibility on threads with millions of impressions. For the deeper analysis of where Premium actually moves the needle and where it does not, our companion piece on whether X Premium affects For You distribution walks through the comparison data.
The SMMNut For You Ranking Matrix (Proprietary Framework)
The challenge with the open-sourced weights is that they tell you what the heavy ranker optimises against but not what you, as a creator, can do about it. The SMMNut For You Ranking Matrix translates the five signal families into a creator-actionable framework — every signal is mapped against three dimensions: the leaked heavy-ranker weight, the velocity window in which the signal is decisive, and the specific creator lever that moves it.
| Signal | Heavy-Ranker Weight | Velocity Window | Creator Lever |
|---|---|---|---|
| Engagement Type Weighting | Reply ~13.5, Profile Visit ~12.0, Bookmark ~10.0, Retweet ~1.0, Like ~0.5 | All-life — applies whenever the signal arrives | Hook structures that solicit replies and bookmarks rather than reactions |
| Engagement Velocity | Multiplicative — early engagement weighted ~3x late engagement | First 15-30 minutes, sharp decay after 60 minutes | Posting time aligned to audience’s high-engagement window |
| Account Credibility Score | Input feature — gates out-of-network candidate eligibility | Rolling 90-day window | Clean engagement profile, gradual follower growth, no policy actions |
| Content Format | Video boost ~1.2x, External-link penalty ~0.5x, Long-form boost ~1.1x | All-life — applies at candidate-sourcing stage | Keep external links out of primary tweet body; prefer threads and video |
| Premium Multiplier | ~1.04x final-score multiplier, widening each quarterly release | All-life — applies post-ranking | Use Premium for reply-positioning on large accounts, not For You expectation |
SMMNut For You Ranking Matrix — Verified Heavy-Ranker Weights (May 2026): Drawing on the configuration files in the github.com/xai-org/x-algorithm repository released on 15 May 2026, X’s heavy ranker assigns the following predicted-engagement weights in its final scoring stage: predicted reply weighted at approximately 13.5, predicted profile visit at 12.0, predicted bookmark at 10.0, predicted retweet at 1.0, and predicted favourite (like) at 0.5. The largest single coefficient in the model is the predicted-negative-action weight of approximately -74.0 — a single report or hide action carries roughly twice the punishment weight of 27 replies combined. This is the structural explanation for why community-pile-on tweets collapse so dramatically and why high-controversy content can outperform until the negative-action threshold trips.
The matrix is diagnostic rather than prescriptive. A creator’s job is to identify which row is the weakest lever for their account and fix that one first. An account with strong content but no posting time discipline is bottlenecked at Velocity. An account with strong velocity but reaction-soliciting hooks is bottlenecked at Type Weighting. An account with strong content and timing but a trust-score recovery in progress is bottlenecked at Credibility. The matrix turns ranking-system knowledge into auditable account work — which is the through-line for every tweet-level decision discussed in our companion piece on how X ranks individual tweets within the candidate pool.
How Your Engagement Rate Interacts With the For You Algorithm
Engagement rate on X is calculated two ways, and creators routinely confuse them. The platform-displayed engagement rate (total engagements divided by impressions, expressed as a percentage) is what X Analytics surfaces — typically 1 to 5 percent for healthy accounts. The algorithm-relevant engagement rate (total engagements divided by follower count) is the signal the For You system actually uses as a quality proxy when deciding which tweets are eligible for out-of-network candidate sourcing.
The reason these two figures diverge matters for growth strategy. The platform-displayed rate is a downstream measurement — it tells you how engaging your content was to the people who actually saw it. The algorithm-relevant rate is the upstream filter — it tells the heavy ranker whether your tweets are worth widening eligibility for. An account with 10,000 followers but only 200 engagements per tweet has a 2 percent algorithm-relevant rate. An account with 100,000 followers and the same 200 engagements per tweet has a 0.2 percent algorithm-relevant rate. The heavy ranker treats these two accounts very differently even though their tweets look identical in isolation.
SMMNut X Engagement Rate Threshold (May 2026 Verified): Based on SMMNut’s analysis of more than 500 X accounts across business, creator, and brand categories in Q1 2026 — cross-checked against the candidate-eligibility logic exposed in the May 2026 open-source release — accounts with sustained engagement rates below 0.8 percent (engagements divided by follower count) rarely receive meaningful For You tab distribution to non-followers, regardless of content quality. Accounts above 2.5 percent consistently break out of follower-limited reach. The threshold is not a hard cutoff in the source code, but it reflects how X’s heavy ranker uses engagement rate as a feature in the candidate-sourcing stage — accounts below this threshold seldom clear the SimClusters and TwHIN candidate gates.
The corollary that catches most creators by surprise: follower acquisition strategies that dilute engagement rate actively hurt your For You distribution. Adding followers faster than your content can engage them pushes the algorithm-relevant rate downward, which reduces candidate eligibility, which reduces impressions, which reduces all forms of engagement on later tweets. This is the structural reason behind our coverage of how follower quality affects your X engagement rate — quality-graded follower additions either preserve or improve the algorithm-relevant rate, while low-quality additions break the feedback loop.
The fastest path to a healthier algorithm-relevant rate is not adding likes — it is adding replies. Reply-weighted engagement signals reach the algorithm in the high-velocity 15-30 minute window. Likes accumulated over hours contribute mostly to the platform-displayed rate without moving the algorithm-relevant figure. Reply-seeking tweet structures — open-ended questions, counterintuitive claims, specific numbers in the first 8 words — are the highest-leverage edit a creator can make to their existing tweet template.
What Has Changed in X’s Algorithm Since the 2023 Rebrand
The 15 May 2026 release of the For You recommendation pipeline at github.com/xai-org/x-algorithm is the largest single piece of evidence about how X’s algorithm operates that the platform has ever produced. The 2023 release exposed a partial snapshot of the system; the 2026 release exposes the production pipeline currently running in the live product, configuration files included. Elon Musk’s announcement tweet on 15 May 2026 received over 43 million views — a level of attention that shifted creator discourse on X for the following two months. The release is the central inflection point in the post-rebrand history of the algorithm.
Comparing the 2023 and 2026 repositories side-by-side reveals what has actually changed under Elon Musk’s ownership and what was already in place. The changes fall into four categories.
1. The reply-weight escalation. The 2023 weights put predicted reply at approximately 27.0 — far above its 2026 figure of 13.5. The drop is the result of X’s engineering team rebalancing the model after the 2023 release surfaced widespread gaming of reply solicitation. Creators in 2023 produced tweets explicitly designed to bait replies with no informational value. The 2024 weight rebalance pulled reply weight downward and pushed bookmark and profile visit upward as alternative high-quality signals. Reply is still the single highest-weight engagement type, but it is no longer 54x stronger than a like — it is closer to 27x.
2. The external-link penalty. The 2026 open-source release confirms what creators had inferred from observation: tweets containing external links are systematically suppressed in the candidate-sourcing stage, not the ranking stage. The penalty was not in the 2023 release in its current form — it was introduced in Q3 2023 alongside the broader strategy to keep users on the platform. The penalty interacts with format: a video tweet with an external link in the body suffers far less than a plain-text tweet with the same link.
3. The Premium subscriber multiplier. The 2023 release did not contain a Premium-specific multiplier — verified accounts were not algorithmically advantaged in the original system. The Premium multiplier was added in early 2024 as part of the broader monetisation strategy, and has been widened in each quarterly release. The 2026 multiplier sits at approximately 1.04x of final ranking score, up from a smaller figure in 2024. The trajectory matters more than the current value — if the multiplier continues widening at its current quarterly rate, Premium will become a meaningful direct factor in For You distribution within 12-18 months.
4. The bookmark signal addition. Bookmarks did not exist as a public engagement type in 2023. The bookmark feature was launched mid-2024 and was added to the heavy ranker as a positive signal at the same time. Its current weight (~10.0) places it as the third-strongest engagement type after replies and profile visits. This addition is the most underrated change — bookmark-soliciting content (saveable lists, how-to threads, reference content) now contributes a signal almost as strong as a reply, but is dramatically easier to elicit because it does not require the viewer to compose a response.
SMMNut Algorithm Inflection Analysis (May 2026): Side-by-side comparison of the github.com/xai-org/x-algorithm 2023 and 2026 repository releases shows four structural changes since Elon Musk’s acquisition: the predicted-reply weight dropped from approximately 27.0 to 13.5; the external-link suppression flag was added to the candidate-sourcing stage in Q3 2023; a Premium-tier final-score multiplier of approximately 1.04x was introduced in Q1 2024 and has widened each quarterly release; and the bookmark engagement type was added to the heavy ranker at weight approximately 10.0 in mid-2024. The cumulative direction of these changes is consistent: the system is rewarding considered engagement (replies, bookmarks, profile visits) more heavily than reaction (likes), is actively keeping users on the platform by suppressing external links, and is building a small but widening structural advantage for paying subscribers.
What has not changed is the core architecture. The two-phase candidate sourcing then ranking pipeline, the 50-50 in-network versus out-of-network split, the SimClusters / TwHIN / Real Graph trio of out-of-network sources, and the heavy ranker’s neural-network structure are all present in both the 2023 and 2026 releases. The system Elon Musk inherited was already engineered around engagement-prediction, not chronological display — the changes since 2022 have been weight adjustments and feature additions within that existing architecture, not a rebuild.
For creators evaluating where the algorithm is going next, the trajectory is informative. The reply-weight escalation has stabilised, the external-link penalty is unlikely to be relaxed, the Premium multiplier is widening, and the bookmark signal is settling in. Strategy that aligns with these trajectories — reply and bookmark-soliciting content, no external links in primary tweets, Premium subscription evaluated for reply positioning, and trust-score awareness — is strategy that compounds with the system rather than against it.
Frequently Asked Questions About the X For You Algorithm
Read on for the seven most common questions creators ask about how the For You tab decides what to show. For the broader algorithm and growth context, refer back to our complete X algorithm and growth guide.