X released the current For You recommendation code on August 14, 2026. Within hours, screenshots turned its model weights into slogans such as “one copied link equals forty likes” and “one report erases hundreds of likes.” The code does not work that way. It predicts how likely one viewer is to take several actions, multiplies those probabilities by weights, applies filters and adjustments, then ranks candidates for that viewer.
That distinction matters for builders. The useful question is not how to manufacture the highest-weight action. It is how to publish original work that the right person is likely to read, share, discuss, and follow—while giving the profile a clear promise when curiosity turns into a visit.
The released algorithm is personalized. A weight of 20 multiplies one viewer’s predicted probability of an action; it does not add twenty points every time someone acts. Small accounts do have a code path for limited new-author exploration, but the released defaults make eligible original posts—not replies or reposts—the useful unit to test.
For 30 days, keep the experiment simple: a clear profile promise, one useful original post on each active day, a small number of substantive replies in relevant conversations, and one weekly funnel review. Track distribution, profile curiosity, follows, and return signals separately so a view spike cannot disguise a conversion problem.
Contents
- The four-stage growth loop
- What X Actually Released—and What It Did Not
- Read the Ranking Formula Before You Read the Weights
- What a 20-Post Public Discussion Sample Revealed
- The Small-Account Opportunity Is Real—but Narrow
- Build the Follow Promise Before You Chase Reach
- Write Original Posts That Create a Useful Next Action
- Use Replies to Build Relationships, Not as an Original-Post Substitute
- Run This 30-Day Builder Growth Experiment
- Measure Follower Growth as a Funnel, Not a Screenshot
- Build From Public Sources, Not an Algorithm Swipe File
- What to Stop Doing After the Algorithm Release
- The Bottom Line
- FAQ
Find the stage that leaks before you add volume
A code-backed growth loop for builders: earn an original-post test, create a useful next action, convert the profile visit, and measure the stage that leaks.
-
Earn a candidate test
Can the post enter a relevant person’s For You candidate set? Publish fresh original work; small-author exploration applies to eligible originals, not every reply.
-
Earn a useful action
Is this viewer likely to reply, quote, share, or follow? Create proof, a reusable lesson, or a decision people genuinely want to pass on.
-
Earn the follow
Does a profile visit answer what the person will keep getting? Align the bio, recent posts, and pinned proof around one recurring promise.
-
Earn the return
Do new followers respond to the next post in the lane? Repeat the topic-job pair that creates downstream actions, not only views.
What X Actually Released—and What It Did Not
On August 14, X Open Source published Open-sourcing the For You timeline and refreshed the public xai-org/x-algorithm repository. The release covers candidate retrieval, action prediction, scoring, filters, author diversity, out-of-network adjustment, and selection. X also introduced an Under the Hood transparency pilot for eligible accounts.
The official For You documentation remains the right plain-language frame: the feed mixes followed accounts with recommended posts and ranks relevance from each person’s interests and interactions. The repository adds detail, but it does not publish a universal posting quota, guaranteed timeline, or raw points balance for creators.
This guide uses the August 14 commit listed in the references. Defaults and experiments can change. Treat the code as a dated window into one recommendation pipeline, not a permanent contract for every X surface.
Read the Ranking Formula Before You Read the Weights
The released ranking scorer uses a weighted sum of predicted actions. For each candidate and viewer, the model estimates the probability of a like, reply, repost, quote, share, follow, negative feedback, and other actions. The scorer multiplies each prediction by its configured weight and adds the results.
That is why “a copied link equals forty likes” is false. A rare action can have a larger weight and still contribute less than a common action if its predicted probability is much lower. These released defaults are not exchange rates and not instructions to beg for a particular click.
For a builder, the actionable reading is behavioral: a post that creates a useful next action can outperform one that earns frictionless approval. Coordinated sharing, duplicate replies, or manufactured engagement collide with X’s policy and teach you nothing about audience demand.
- Favorite: 0.5
- What the scorer uses: Predicted probability that this viewer will like. Builder reading: A like is useful feedback, but it is not the whole objective.
- Reply: 5
- What the scorer uses: Predicted reply probability; an eligible mutual-follow original reply gets a separate boost. Builder reading: Make a claim specific enough to discuss. Do not bolt on an empty question.
- Quote: 5; follow author: 4
- What the scorer uses: Predicted quote and follow probabilities. Builder reading: Publish a decision others can contextualize and make the follow promise coherent.
- Copy-link share: 20
- What the scorer uses: Predicted probability of a comparatively rare copy-link action. Builder reading: Create something worth sending to a teammate. One share does not equal forty likes.
- Not interested: -43.2; report: -234
- What the scorer uses: Predicted negative feedback from this viewer. Builder reading: Audience mismatch and bait are expensive. A report does not literally cancel 468 likes.
Weights multiply predictions; they do not count reactions.
When a growth post turns one released weight into an equivalent number of likes, check whether it preserved the per-viewer probability term. Most viral arithmetic does not.
What a 20-Post Public Discussion Sample Revealed
On August 16, we manually reviewed 20 distinct public X posts across four searches: the official announcement plus four high-visibility quote posts, nine algorithm explainers or rebuttals, and six near-identical builder-connection templates. This is a transparent qualitative sample, not a claim about the total number of posts on X.
The official announcement alone showed 663 replies, 2.4K reposts, and 5.2M views when captured. Those figures establish substantial attention, but they are engagement events—not a unique headcount of people discussing the algorithm.
The strongest discussion separated predictions from raw counts and warned that ranking is only one stage. The weakest copied weights into “how many likes is this worth?” tables, treated the small-author path as guaranteed reach, or repeated the same “Hey algorithm, connect me with builders” wording across accounts.
Useful counterexamples said the viral table was reading the code backwards and showed why a weight of 20 is not forty likes. The valuable contribution was a correction that made the primary source harder to misuse.
For builders, that is a content opportunity: translate technical changes into decisions your audience can verify. A precise correction with a source is more follow-worthy than an unexplained weights screenshot.
- Typical useful take
- Explains one mechanism, links the code, states what it cannot prove, and turns the result into a bounded test.
- Typical viral misread
- Drops the probability term, converts weights into like equivalents, and presents an experiment default as platform law.
- Typical low-value builder post
- Tags the algorithm and copies a broad networking prompt without proof of work, a point of view, or a reason to follow.
The Small-Account Opportunity Is Real—but Narrow
The released new-author cold-start scorer gives small accounts a reason to test original work. In the August 14 defaults, an eligible candidate is an original post—not a reply or repost—from an author with no more than 1,000 followers, with no more than 1,000 impressions, and no more than 24 hours old. The selector can lift an eligible candidate toward a target position so the system can learn from it.
That does not mean every small-account post receives a boost, that a position is guaranteed, or that the defaults are identical in production. It means the released system contains an explicit exploration mechanism. Your job is to give that test something worth learning from.
The released parameter file also contains a default 48-hour candidate age filter. That is not a platform-wide expiration date. It does mean a builder should measure the first day or two instead of waiting a week to diagnose the initial For You test.
- Publish an original unit
- A decision note, teardown, test result, code lesson, or customer observation gives the exploration path something distinct to evaluate.
- Make the audience legible
- “What broke activation for our first 30 SaaS trials” is clearer than “a lesson from onboarding.”
- Measure the first 24–48 hours
- Record distribution and downstream actions while the candidate is fresh, then compare with your own recent originals.
- Do not confuse eligibility with entitlement
- A cold-start path creates a test opportunity. It does not make generic content competitive.
Build the Follow Promise Before You Chase Reach
A ranking model can earn an impression. It cannot make a confused profile convert. A profile visit asks: “If I follow this account, what will I keep getting?”
X lets you set a name, bio, website, profile image, header, and one pinned post. Treat them as one answer. The bio names the reader and recurring value. Recent posts prove the promise. The pinned post gives a visitor the best place to start.
A builder does not need to sound like a generic creator. Product decisions, failed tests, customer language, and operating constraints are already differentiated raw material.
- Name the reader
- “Bootstrapped SaaS founders shipping onboarding” is clearer than “builders and dreamers.”
- Name the recurring value
- “One teardown from our activation experiments each week” gives the follow a future payoff.
- Pin proof
- Use the best teardown, postmortem, resource, or result—not a résumé rewritten as a greeting.
- Make recent posts agree
- A precise bio cannot convert if the visible timeline looks like six unrelated accounts.
Write Original Posts That Create a Useful Next Action
Do not design a post around a weight. Design it around what a relevant reader can do next. Builders have three strong raw materials: proof from real work, reusable decisions, and informed disagreement.
Rotate the three jobs below during the first month. The mix is small enough to compare and varied enough to reveal whether the audience values evidence, instruction, or judgment.
- Proof
- “We cut onboarding from eight steps to three. Activation did not move. The actual blocker was the eleven-minute wait for the first import. Here is the change that fixed it.” The surprise creates curiosity; the operating detail makes it credible; the result is easy to share with a product team.
- Reusable decision
- “Before adding an AI summary, check three things: can the user verify the source, recover the raw record, and tell what the model inferred?” A compact decision rule gives a reader something worth bookmarking or sending.
- Informed disagreement
- “More signup questions did not improve lead quality for us. They moved uncertainty from sales into onboarding. We now ask one qualifying question and measure time to first value.” A counterpoint works when it includes the condition and receipt, not when it is contrarian theater.
- Weak copy-paste networking
- “Hey algorithm, connect me with builders interested in AI, SaaS, startups, growth, and tech.” It identifies almost everyone, demonstrates nothing, and gives the profile visit no specific reason to become a follow.
Would someone send this to one specific teammate?
If you can name the teammate and decision the post helps them make, it has a real share path. If the answer is only “because the algorithm values shares,” it is not ready.
Use Replies to Build Relationships, Not as an Original-Post Substitute
The durable part of older X growth guides is human: a specific reply in a relevant conversation can earn attention, a profile visit, and a relationship. The new code adds an important qualification to the reach story.
In the released For You pipeline, the out-of-network reply and repost filter removes replies and reposts from authors the viewer does not follow. That makes “reply under big accounts for algorithmic distribution” too broad. Replies can still be found inside conversations, seen by network participants, strengthen relationships, and create profile curiosity. They should not replace original posts that prove why someone should follow.
Choose a small set of people whose readers overlap with your audience. Reply when you have evidence, an edge case, a clarifying question, or a respectful counterpoint. Leave when the only move is applause.
- Weak
- “Great point. Consistency is everything.” It can sit under almost any post and creates no reason to inspect the author.
- Useful extension
- “We saw the same drop, but only for teams importing more than 10,000 rows. The fix was background validation, not a shorter form.” The reply adds a boundary condition the original audience can use.
- Useful question
- “Did the result hold after the novelty week? Our first launch cohort converted, but week-three activation returned to baseline.” It exposes a missing decision instead of adding a generic CTA.
Hide the parent post.
If the reply still contains a useful claim, example, question, or counterpoint, it may earn a profile visit. If it collapses into applause, rewrite it or skip it.
Run This 30-Day Builder Growth Experiment
This plan creates comparable evidence; it does not promise a follower count. On each active day, publish one useful original and write a small number of considered replies where you have something real to add. Five replies is a workable ceiling, not an algorithm quota.
Keep one row per original post with its URL, audience, post job, topic, visible time, views or impressions, profile clicks when available, follows when available, replies, reposts, bookmarks, and one sentence about what happened.
-
Days 1–3: Fix the conversion surface
Write one follow promise. Align the bio, recent posts, and pinned proof. Record follower count and a baseline from recent originals.
-
Days 4–10: Test original-post jobs
Publish proof, a reusable decision, and informed disagreement. Keep the profile stable and record the first 24–48 hours.
-
Days 11–20: Repeat the useful lane
Choose the two topic-job pairs with the strongest downstream actions. Create adjacent examples from new evidence; do not copy wording.
-
Days 21–30: Turn the winner into a series
Name a repeatable weekly format, keep one exploratory lane, and stop the lane that creates empty views without profile curiosity or follows.
Measure Follower Growth as a Funnel, Not a Screenshot
X’s Post Activity Dashboard documentation describes impressions, profile clicks, and follows. Supported developer analytics also expose post-level fields such as replies and bookmarks. Availability varies, so use the fields you have.
Do not confuse views with people. X says view counts are not unique; repeats and the author’s own views may count. A view spike is a distribution event, not proof of audience growth.
Compare each post with your own recent baseline after metrics settle. Identify the leaking stage, change one variable, and keep the rest stable.
- Originals receive little distribution
- Likely bottleneck: Candidate relevance or opening clarity. Next test: Name the audience and problem earlier; keep profile and topic stable.
- Views rise, profile clicks stay weak
- Likely bottleneck: Curiosity. Next test: Add a specific result, tradeoff, or unresolved decision.
- Profile clicks rise, follows stay weak
- Likely bottleneck: Profile conversion. Next test: Rewrite the follow promise and pin proof matching the visit source.
- Follows rise, the next post is ignored
- Likely bottleneck: Return value. Next test: Publish an adjacent post in the same lane before changing direction.
- Replies get views but originals stay flat
- Likely bottleneck: Borrowed attention without owned proof. Next test: Turn the strongest reply insight into an original post with evidence.
- Curiosity rate
- user-profile clicks ÷ impressions. Use only when both fields are available.
- Follow conversion
- follows from posts ÷ user-profile clicks. A falling ratio points back to the profile promise.
- Return evidence
- response to the next post in the same series. Review replies, bookmarks, reposts, and follows separately; this is not a ranking formula.
Build From Public Sources, Not an Algorithm Swipe File
A good research note keeps the post URL, author, visible date, claim, evidence, and your interpretation together. That lets you learn without copying wording or repeating a context-free screenshot.
Use CarryFeed’s public X/Twitter search to find discussions by account, phrase, date, link, or media clue. When an example matters, open the public source and keep its final URL with the note.
Use AI to cluster notes, compare openings, or challenge a draft. Keep source selection, judgment, lived evidence, and publishing with a person. A model can help interpret a code path; it cannot turn an unsupported growth claim into evidence.
- 01
- Public source URL
- 02
- Author and visible date
- 03
- Exact claim being evaluated
- 04
- Primary source that confirms or contradicts it
- 05
- What a builder can test
- 06
- What the source cannot prove
- 07
- Capture date and visible metrics, if relevant
- Find public examples without losing the source.
- Search public X/Twitter postsSearch by account, phrase, date, link, or media clue, then keep the public result URL attached to the note.
What to Stop Doing After the Algorithm Release
X’s authenticity policy prohibits bulk, duplicative, irrelevant, unsolicited, and coordinated manipulative activity. Its automation rules restrict automated unsolicited replies. Actions that fake demand also make your experiment unreadable.
- Stop converting weights into like equivalents
- The calculation drops the predicted probability and misstates the scorer.
- Stop tagging the algorithm with copied networking prompts
- Six near-identical examples appeared in a small public sample. Repetition is not positioning.
- Stop manufacturing the golden hour
- Engagement groups and secondary-account amplification cross policy boundaries and obscure demand.
- Stop replacing originals with replies
- Replies build relationships; originals provide the clearest current small-author test and profile proof.
- Stop worshipping views
- A large non-unique count with no profile clicks, follows, or return activity is distribution without growth.
- Stop changing six variables after one miss
- One post is an event. A stable series of comparable tests is evidence.
The Bottom Line
The new X algorithm does not give builders a cheat code. It gives them a better model of the problem. A post has to enter a candidate set, look useful to a particular viewer, survive filtering and adjustment, earn a profile visit, and make a coherent follow promise.
Small accounts have a real but bounded exploration path for fresh original posts. Use it to publish proof from the work, not weight arithmetic. Run the 30-day experiment, preserve the source behind every lesson, and fix the stage that leaks before adding volume.
Frequently asked questions
How do I grow Twitter followers after X’s new algorithm release?
Publish fresh original posts for a specific audience, make each post useful enough to discuss or share, align the profile around one recurring promise, and measure impressions, profile clicks, follows, and return activity separately. The released code supports a limited small-author exploration path, not a guaranteed shortcut.
Does one copy-link share equal forty likes on X?
No. The released weight multiplies the predicted probability that a particular viewer will take that action. It is not a raw engagement score, so weights cannot be converted into equivalent counts of likes.
Does X boost accounts with fewer than 1,000 followers?
The August 14, 2026 code contains a cold-start path whose defaults make certain fresh original posts from authors with no more than 1,000 followers eligible for exploration. Eligibility is not a guarantee, experiments can differ, and replies or reposts do not qualify as the original candidate.
Are replies or original posts better for small accounts?
They do different jobs. Originals provide the clearest code-backed small-author test and prove what the account publishes. Replies can create relationships and profile curiosity, but the released For You pipeline filters out-of-network replies and reposts, so replies should not replace originals.
How often should I post on Twitter/X to grow?
X does not document a universal creator quota. Start with one useful original per active day during a controlled 30-day experiment. Add volume only when ideas remain specific and downstream metrics hold.
What should I measure besides followers?
Track views or impressions, profile clicks, follows from posts when available, replies, reposts, bookmarks, and response to the next post in the same series. Views are not unique people.
Can AI write my X posts and replies?
AI can organize public source notes, compare openings, edit drafts, and flag unsupported claims. Keep the source, judgment, lived evidence, and publishing decision with a person. Bulk or unsolicited automated replies cross X’s policy boundaries.