How the Twitter Algorithm Works Right Now

The Most Counterintuitive Finding in the Data

Small accounts on X have a significant reach advantage over large ones.

In an analysis of 4,360 tweets across account sizes, nano accounts (under 1,000 followers) averaged 1.66 views per follower. Mega accounts (over 1 million followers) averaged 0.004 views per follower. That is a 400x difference in relative reach - in favor of small accounts.

Mid-tier accounts (10K-100K followers) averaged just 0.35 views per follower. That is the worst place to be. Big enough that the algorithm stops giving you the "new voice" boost, but not big enough to have the raw follower volume that makes low per-follower reach irrelevant.

The engagement rate data confirms this. Accounts under 10K followers averaged a 5.30% engagement rate. Accounts with 10K-100K followers dropped to 3.47%. Accounts over 100K fell to 2.12%. Small accounts outperform large ones on engagement rate by 2.5x.

If you have been sitting on the sidelines thinking "I only have 800 followers, so what's the point" - that is exactly backwards from how the algorithm is distributing reach right now.

What the Open-Source Code Says

X is the only major social platform that has open-sourced its algorithm - twice. The engagement weights are on GitHub. Every interaction type has a documented value.

Here is the confirmed scoring formula from the open-source code:

Read that table again. I watch creators optimize for likes every day. Likes are the lowest-weighted positive signal in the entire system.

A reply that gets a reply back from the original author carries a weight of 75 - that is 150x more powerful than a single like. When you post something and then reply to the people who reply to you, every one of those author-reply combinations is worth 150 likes in algorithmic value. This is why going back and engaging with your comment section is not optional - it is one of the highest-impact moves available.

The algorithm also predicts negative signals before they happen. It does not wait for actual blocks and mutes. The model predicts whether a user would block, mute, or report your content - and that prediction alone lowers your distribution score. The system has learned what "blockable" content looks like. Post things that generate that signal and your reach drops before a single real block happens.

The First 30 Minutes Determine Everything

When you post, the algorithm shows your tweet to roughly 10-20% of your followers in the first 15 minutes. What happens next is binary.

If that initial audience engages at above 3%, the algorithm interprets the content as valuable and expands distribution to 50-80% of your followers plus non-followers. That is a 5-10x impression boost triggered purely by early engagement velocity.

If engagement stays below 1% in that first window, distribution stops. The tweet effectively dies.

The recency decay schedule from the open-source code makes this concrete:

• 0-30 minutes: 100% distribution potential
• 30-60 minutes: 80%
• 1-2 hours: 60%
• 2-4 hours: 40%
• 4-8 hours: 20%
• 8+ hours: minimal organic reach

The practical implication here is significant. Posting when your audience is not active does not just mean fewer eyes on launch - it means the algorithm decides your content is low-quality and throttles it permanently. You cannot recover from a bad launch window by waiting. The tweet's fate is set in the first hour.

Post when your followers are online so you get engagement in the first ten minutes. Miss that window and the content is functionally dead regardless of quality.

The Link Penalty Is Worse Than You Think

Every analysis of the open-source code confirms it. External links in tweet bodies trigger a reach reduction of 30-50%. For non-Premium accounts specifically, the median engagement on posts with external links has dropped toward zero.

X wants users to stay on X. Links that take users off-platform contradict that goal, and the algorithm penalizes them directly.

The workaround is simple and used by high-performing accounts: post your content tweet with no link, let it gather early engagement, then drop the link in the first reply. The reply inherits the parent tweet's visibility. Your audience still gets the link. The parent tweet avoids the distribution penalty.

This single change - moving links from the tweet body to the first reply - is one of the fastest reach improvements available with zero additional content creation required.

The Reply Guy Strategy: The Numbers Behind It

Strategic replying to large accounts in your niche is the highest-organic-growth tactic in the current X environment.

Here is why it works mechanically. When you reply to a post from an account with 100,000 followers, your reply gets placed in front of that account's entire engaged audience. If your reply is good, those people click your profile. Profile clicks carry a weight of 12x in the algorithm - one of the highest signals available. Those profile visits then signal to X that your account is worth distributing.

The numbers practitioners are reporting: 50+ quality replies per day over two weeks has produced gains of 3,000-8,000 followers in 14 days with over 100 million impressions. One documented case tracked a path from zero to 2,000 followers plus 14 million impressions purely through 100+ replies per day.

The critical word is quality. After recent algorithm updates, volume without substance actively backfires. Five hundred shallow replies per day - generic agreements, single emoji reactions, "great point" responses - now trigger spam flags. The algorithm has gotten better at detecting reply farming versus genuine conversation.

What works: replies that add a specific data point, a contrarian angle, or a short story that relates to the original post. Replies that make someone click your name to see who said that. That is what converts profile visits into followers and what generates the algorithm signals that matter.

Hook Format Data From the Top 100 Highest-Performing Tweets

The top 100 highest-liked tweets in the dataset showed a clear pattern in hook formats - and it directly contradicts what most Twitter growth advice recommends.

Question hooks are one of the most recommended Twitter formats in growth advice content. They are also one of the least-represented formats in top-performing posts. Only 3% of the highest-performing tweets used a question as their opening hook.

Number and stat-based openings dominated at 34%. First-person story hooks came in second at 29%. Together, those two formats account for nearly two-thirds of the top-performing posts.

The pattern makes sense given the algorithm mechanics. Number-based hooks trigger pattern interruption - a specific figure in a feed of text stops the scroll. Story hooks generate dwell time, which is a positive engagement signal even without a click or like. Question hooks, by contrast, often feel like engagement bait, which the algorithm specifically penalizes.

Lead with a number or lead with "I" followed by something that happened. Those are the two formats the data supports.

Text vs. Video - The X-Specific Anomaly

X is the only major platform where text consistently outperforms video on engagement rate.

The engagement rate breakdown by content type:

• Long posts (500+ characters): 4.21% average engagement rate
• Threads (multi-part): 4.15%
• Short posts (under 100 characters): 4.00%
• Video posts: 2.77%

Video generates the lowest engagement rate despite being heavily promoted as the priority format. Long text posts and threads generate 50%+ higher engagement rates than video.

This does not mean video is useless on X. It means the platform culture skews toward text-first content more than any other major platform. When someone opens X, they are typically in a reading mindset. The implication for content strategy is to not port your video-first strategy from other platforms directly onto X and expect the same results.

The Posting Frequency Paradox

This is one of the findings that most contradicts standard growth advice.

In the dataset, accounts with only one post in the sample averaged 303 likes per tweet. Accounts with three or more posts in the sample averaged 94 likes per tweet. That is a 3x higher per-tweet performance for accounts posting less frequently.

The algorithm has what practitioners describe as an "Author Diversity Scorer" that penalizes rapid sequential posting from the same account. It prevents any single account from flooding a user's feed. The more you post in a short window, the lower the distribution weight on each individual post.

Think about your posting as individual drops, not a firehose. Each post you send competes with itself for distribution budget. If you post six times in a day, you are diluting the attention the algorithm gives each one.

One operator who has built a multi-million dollar business through organic X content started asking after every post: "why did that one do the numbers it did?" Then made the next one better. That feedback loop, not raw volume, is what compounds over time.

X Premium - What It Does and Does Not Do

Premium accounts receive a 2-4x initial reach boost and have their replies algorithmically prioritized in conversation threads. In the dataset, there were 25 positive mentions of Premium's impact versus only 8 negative ones - a 3:1 positive signal from practitioners who have tested it.

The details matter. One creator with direct testing experience put it this way: Premium does not give you reach directly. It improves your reply placement. That improved placement leads to more profile visits, which builds your follower base, and stronger initial distribution on future posts follows from there. The benefit compounds indirectly, not as a direct reach injection.

Bad content plus Premium still performs badly. The boost multiplies what is already working - it does not create performance where there is none.

Whether Premium is worth the cost depends on where you are in the growth curve. For an account that is already generating engagement and wants to accelerate the reply-to-profile-visit funnel, the math likely works. For an account that has not found content-market fit yet, paying for Premium before solving the content problem is the wrong order of operations.

The Suppression Reality - What Is Getting Throttled

The data captured 38 documented suppression and throttling complaints from creators in specific niches. Two patterns stand out.

First, crypto content. Accounts posting crypto charts, ticker symbols like $BTC and $ETH, and crypto price commentary reported reach reductions of up to 90% after recent algorithm updates. Multiple accounts tracked before-and-after metrics showing the same post format going from 50K impressions to under 5K.

Second, certain political content. One documented case showed an account going from an average of 1,000 likes plus 50K impressions to 1,000 likes plus only 5,000 impressions - a 90% reach reduction on the same engagement volume. The account attributed this to content category changes in how the algorithm classifies certain political topics.

The practical signal here is not to avoid these niches entirely. It is to be aware that some content categories carry higher suppression risk right now, and that if you are in those categories, your reach numbers may not reflect your content quality. Diversifying the types of content within a niche - mixing charts with analysis, mixing political commentary with adjacent topics - appears to reduce the suppression signal.

The Viral Threshold - What the Numbers Look Like

In the dataset, posts that crossed 10,000 views averaged 1,288 likes, 198 retweets, and 171 replies. Two ratios stand out from that data.

The retweet-to-like ratio in those top posts was 0.17 - meaning for every like, you need roughly one retweet per six likes to sustain viral distribution. I see this constantly - creators watching likes as the indicator. But a post with 1,000 likes and 50 retweets is going to cap out. A post with 1,000 likes and 200 retweets is going to keep expanding.

The reply-to-like ratio in viral posts was 0.45 - nearly one reply for every two likes. High reply counts are one of the strongest signals of genuine engagement versus passive appreciation. A high reply-to-like ratio signals that the content is generating real conversation and the algorithm distributes accordingly.

If you want to reverse-engineer what distribution-worthy content looks like, those ratios give you a target. The question to ask about any tweet you write is: will this make people want to share it and respond, or will it just make them tap the heart and scroll?

How Grok Changed the Ranking System

In January of the most recent update cycle, xAI released a Grok-powered version of the algorithm, replacing the legacy ranking system with a transformer model. The new system processes 500 million daily tweets and makes 5 billion ranking decisions per day.

The key difference in the new system: it understands content in terms of meaning. The previous system relied heavily on social graph signals - who engages with who, what accounts follow each other. The Grok-based system reads the content of every post and watches every video, matching users to content based on meaning rather than just social proximity.

This is why niche creators are getting discovery opportunities that were harder to access under the old system. A small account posting highly specific content about a narrow topic can get surfaced to users who engage with similar topics, even with zero social graph overlap. The SimClusters system groups users into approximately 145,000 topic communities - and your content can reach any of those communities if Grok identifies it as relevant.

The practical implication: specificity beats generality. A post that clearly signals what it is about, who it is for, and what category it belongs to performs better in a semantically-driven recommendation system than a post that is interesting-but-vague.

The Author Diversity Filter - Why Back-to-Back Posts Kill Each Other

There is a diversity mechanism in the algorithm that prevents any single account from dominating a user's feed. Even if you heavily engage with one account, the system limits how many consecutive posts from them appear.

This filter applies to your own posting cadence too. Rapid-fire posting - multiple tweets in a short window - activates the filter and reduces per-post distribution. The system treats bursts of posting from one account as spam behavior and moves that feed real estate to other accounts.

The accounts seeing the highest per-tweet performance in the data are posting selectively, not constantly. The "post 3-5 times per day" advice that circulates widely in X growth content is wrong. That frequency advice may work for building posting habits, but the per-post performance data does not support it as an optimization strategy.

What Is Working Right Now - The Practical Summary

Based on the engagement weight data, the reach findings, and the practitioner reports in the dataset, here is what is producing results:

Reply first, post second. Strategic replies to large accounts in your niche generate profile clicks (12x weight), which build the follower base that strengthens your future post distribution. Accounts documenting 100+ quality replies per day are consistently outpacing accounts that only post original content.

Optimize for retweets and replies, not likes. The scoring formula is not a secret. Retweets are worth 20x a like. Replies are worth 13.5x. Every piece of content you write should be evaluated by whether it makes someone want to share it or respond to it - not just tap the heart.

Put links in replies, not tweets. This one change removes a 30-50% reach penalty from every post where you would have included a link. Post the content, get engagement, then add the link in a reply.

Post when your audience is active and then engage hard in the first 30 minutes. The distribution decision is made in the first hour. If you can get early engagement above 3% in that window, the algorithm expands reach 5-10x. If not, the tweet caps out. The launch window is not a nice-to-have.

Reply to your own replies. Every author reply to a commenter carries a weight of 75 in the algorithm - 150x a like. Go back into your comment section and engage. It is one of the highest-return actions you can take after posting.

Lead with numbers or stories, not questions. The top-performing hook formats from the data are number/stat openings (34% of top posts) and first-person story openings (29%). Question hooks, despite being widely recommended, showed up in only 3% of the highest-performing posts.

To move faster on reply strategy and content optimization, Try SocialBoner free - it includes viral tweet search, AI tweet writing, and scheduling built specifically for X growth.

The One Metric Most Creators Ignore

Bookmarks carry a weight of 10x in the scoring formula. That is higher than link clicks, higher than profile clicks from the scoring hierarchy perspective.

Yet almost no creator tracks or optimizes for bookmark rate. I have looked at dozens of creator dashboards - almost none of them have even opened the bookmark metric.

A bookmark is someone saying "this is worth coming back to." It is one of the clearest signals of genuine value the algorithm can read - the user chose to save it, not just react to it in the moment. Posts that get heavy bookmark volume are explicitly signaling depth and usefulness to the distribution system.

Content formats that generate high bookmark rates: actionable frameworks, resource lists (placed in a reply to avoid the link penalty), step-by-step processes, and data-driven findings. These are the same formats that drive high retweet rates. The correlation is not coincidental - both actions signal that someone found the content worth preserving or sharing, which is exactly what the algorithm is trying to identify.