There’s a version of this post that would tell you to “post consistently and add value.” You’ve read that one. This isn’t it.

What follows is a breakdown of how the algorithms actually work right now: what they’re optimising for, what’s changed in the last year, and what that means practically for anyone trying to get seen in 2026. No recycled advice. No “just be authentic.” Just what the data and the platform engineers have actually said.

Let’s go platform by platform.

TikTok

The core mechanic

TikTok runs on an interest graph, not a social graph. It doesn’t care whether you follow the creator. It cares whether it can predict you’ll watch. Every video gets tested on a small cohort first, and in 2026, that cohort is now primarily your existing followers, not a random audience slice. This is the biggest shift in recent memory.

If your followers watch it through, share it, save it, the algorithm decides the content has earned a wider audience. If they don’t, it stops there. The implication: your existing following now functions as a quality filter before broad distribution is ever considered.

What it’s measuring

Watch time still matters but the bar has risen. TikTok now measures “Qualified Views,” which means views longer than five seconds. If people bounce before that, the video is effectively marked as low-quality. After that threshold, completion rate and rewatch rate are the signals that determine whether it goes wide. Shares and saves outweigh likes by a significant margin. A like takes one tap. A save or share signals real intent.

There’s also a search dimension that’s become impossible to ignore. According to Adobe’s January 2026 survey, 49% of US consumers now use TikTok as a search engine rising to 65% among Gen Z. The algorithm scans captions, on-screen text, and spoken words and importantly, on-screen text carries roughly the same weight as spoken keywords. If you’re creating content with a searchable angle, say the keyword out loud, write it on screen, and include it in your caption.

What changed

The follower-first testing model is the key 2026 update. Previously, new videos would be tested against a broader cold audience. Now your followers act as the first jury. This rewards accounts that have built a genuinely engaged following over accounts that chased volume. It’s a meaningful shift away from pure virality mechanics toward something closer to earned reach.

Also: TikTok’s originality detection has improved. The algorithm still actively suppresses recycled content and cross-posts from other platforms.

The short version

Rewatch rate is now the metric that separates content that goes wide from content that stays flat. Design for the rewatch. If you can create a natural loop point where the end of the video flows back into the beginning you’ll hold average view duration longer than a cleaner edit that just ends.

Instagram

The core mechanic

Instagram doesn’t have one algorithm. It has several each running independently across Feed, Reels, Explore, Stories, and Search. The mistake most people make is treating them as one system and wondering why the same strategy produces different results on different surfaces.

The most significant 2026 update: Instagram has moved to “Views” as the primary unified metric across all formats. Reels, photos, carousels all now ranked through the same lens. The follow-up significance: DM shares now carry three to five times the algorithmic weight of a like, confirmed by Mosseri in January 2025. Content that makes people send it to someone else gets disproportionate distribution.

What it’s measuring by surface

Reels still offer the most discovery potential they’re the only format Instagram actively pushes to non-followers. The Reels algorithm weights user behaviour first (what you’ve liked, shared, saved), then history with the posting account, then details about the Reel itself. Completion rate matters, and Instagram introduced an “Originality Score” in 2026 that actively detects and penalises recycled clips from other platforms, including TikTok watermarks.

Feed has shifted notably toward carousels. Carousels consistently outperform single images on both reach and saves the mechanism is straightforward: users spend longer on them, and swipe actions register as strong engagement signals. Carousels with 7–10 slides are now outperforming shorter ones.

Stories don’t drive discovery, but they do sustain relationship. Instagram’s own data shows accounts that post Stories regularly see fewer unfollows than those that don’t.

What changed

The Reels length picture is more nuanced than most guides suggest. Instagram extended the maximum Reel length to 3 minutes in January 2025, and longer Reels are now eligible for recommendations where they weren’t before. But Mosseri himself has said shorter is generally better for Reels, and both Hootsuite’s research and Sprout Social’s benchmarks point to 7–30 seconds as the sweet spot for discovery reach. The practical read: longer Reels can work when retention is genuinely exceptional, but length alone isn’t the lever completion rate relative to length is what the algorithm is actually measuring.

Also new: the “Your Algorithm for Reels” feature, launched December 2025, lets users explicitly tell Instagram what topics they want to see more or less of. This isn’t cosmetic. It means the algorithm now has explicit preference data to work with, not just inferred signals. For creators, this means content with a clear, consistent topic focus is easier for the algorithm to route correctly.

The short version

If you want discovery, build for Reels. If you want retention and relationship, build for carousels and Stories. DM shares are the signal Instagram is prioritising most aggressively right now. Content that makes people think “I need to send this to someone” will outperform content that makes people think “that was good.”

LinkedIn

The core mechanic

LinkedIn in 2026 has two things working against most people posting on it: organic reach for company pages has dropped to roughly 1.6% of followers, and the platform has shifted from rewarding broad engagement to what it calls “depth and authority.” The feed is less social graph, more interest graph optimising for professional relevance and demonstrated expertise over who you know.

The AI model behind the current LinkedIn feed (sometimes referenced as 360Brew) has introduced something called “Topic DNA.” It analyses your headline, about section, and posting history to assign your profile a topic cluster. Posts that fall outside that cluster get suppressed not because they’re bad content, but because they don’t match the signal the algorithm has established for who your content is for.

What it’s measuring

Dwell time is now a primary signal. Not likes. Not even comments in isolation. The algorithm measures how long someone actually spends looking at a post. A post someone reads for 30 seconds outperforms a post with 50 fast likes. The system can even detect “click bounces” when someone clicks and immediately leaves and deprioritises content that produces them.

Comments carry 15x more algorithmic weight than likes. And not any comment, a substantive comment, a few sentences, something that adds to the conversation. “Great post!” is almost worthless. A genuine reply that spawns a thread is algorithmically meaningful.

Posts with external links consistently see around 60% less reach than identical posts without them though LinkedIn’s official position is that this isn’t an intentional penalty, but a consequence of posts with links generating less engagement (because people click away, and the algorithm reads that as low-quality content). The outcome is the same regardless. The “link in first comment” workaround still reduces the hit somewhat, but also still carries a measurable reach reduction compared to a post with no link at all.

What changed

The early window after posting is decisive. Richard van der Blom’s Algorithm Insights Report, the most comprehensive independent study of LinkedIn’s algorithm analysing 1.8 million posts, identifies the first 30–60 minutes as the critical window. LinkedIn tests new posts with 2–5% of your network initially, and how that group engages determines whether the post gets broader distribution. Responding to early comments quickly matters significantly, and posting when your network is most active gives that initial test the best possible conditions.

Document posts (carousels uploaded as PDFs) are the highest-performing format in 2026. Single-image posts now underperform text-only content by 30%, reversing patterns from 2024–2025.

The short version

LinkedIn has become the platform where being a known expert on a specific topic matters more than audience size. The algorithm is explicitly trying to route your content to people interested in your specific niche. Consistency within a clear subject area, combined with content that earns genuine comments and reads for more than a few seconds, is the compound growth play here.

X (Twitter)

The core mechanic

X is the only major social platform to have open-sourced significant portions of its algorithm, meaning more of what follows is confirmed rather than inferred. The engagement weights are documented. A retweet is worth 20x a like. A bookmark is worth 10x. These aren’t estimates. They’re from the published code.

In January 2026, X replaced its legacy recommendation system with a Grok-powered transformer model that processes around 500 million posts per day and makes approximately 5 billion ranking decisions daily. The model now reads semantic meaning. If you post about “Java,” it knows from context whether you mean coffee or coding, and routes accordingly.

What it’s measuring

Conversation quality. Posts that generate genuine replies, particularly reply chains where the original author responds, are what the algorithm rewards most aggressively. The architecture isn’t built to surface content people passively like. It’s built to surface content people feel compelled to respond to.

On the engagement weights: the open-sourced code shows a like scores +0.5, while a reply that gets a reply back from the original author scores +75, making that specific interaction around 150x more valuable than a like. Sprout Social’s own analysis of the same code puts a standard reply at ~13.5x a like. The practical takeaway is the same either way: conversation depth is the primary signal, and responding quickly to your own replies is the highest-leverage thing you can do after posting.

X Premium (the paid tier) now functions as a meaningful algorithmic boost: a 4x boost for in-network content and 2x for out-of-network content. At the scale that matters for professional use, many creators are treating the subscription as effectively mandatory. There’s also an “author diversity penalty” if you post 10 times a day, the algorithm won’t show all 10 to your followers. It will surface the 2–3 strongest performers. More posting doesn’t compound; it cannibalises.

What changed

The Grok integration represents the most substantial architectural shift on the platform since Musk’s acquisition.

Smaller accounts were explicitly favoured in the 2026 update. The platform made a deliberate choice to surface content from newer and smaller creators more aggressively, provided it earns early engagement. This is the counter-weight to the Premium boost. The system isn’t purely pay-to-play, but Premium does give content a higher baseline from which to be tested.

The short version

X rewards hot takes, questions, and positions that make people want to respond. A post with 20 genuine replies will consistently outperform a post with 500 likes and no conversation. If you’re not investing time in your replies section, responding quickly and extending the thread, you’re leaving the most valuable engagement signal on the table.

YouTube

The core mechanic

YouTube processes over 80 billion signals daily and derives more than 70% of all watch time from algorithmic recommendations rather than search or subscriptions. But its goal is distinct from every other platform on this list: it’s not optimising for time on platform through novelty. It’s optimising for satisfaction. Whether viewers felt their time was well spent.

In 2025 and into 2026, YouTube shifted its entire recommendation model to what it calls “satisfaction-weighted discovery.” This means viewer surveys, repeat viewing behaviour, and “not interested” signals all feed directly into how content is ranked, not just click-through rate and watch time.

What it’s measuring

The five most consequential signals, in rough order of importance:

Click-through rate matters, but not as a standalone gate. Todd Beaupré, YouTube’s Senior Director of Growth and Discovery, made this explicit: a 5% CTR video reaching 100,000 people is far better than a 20% CTR video reaching 10,000. High CTR paired with poor watch time (the clickbait trap) is actively penalised. The algorithm is looking at CTR and retention together, not CTR alone.

Average view duration is the second key signal. High CTR with low watch time signals clickbait. The algorithm is explicitly built to penalise this: a misleading thumbnail might boost initial CTR but the retention crash that follows deprioritises the video going forward.

Satisfaction signals are now the third factor the algorithm weighs heavily. These include likes and shares, but also repeat viewing within a topic and “not interested” signals. If viewers consistently skip or click away, even videos with decent watch time get deprioritised.

YouTube doesn’t have one algorithm. It has five distinct systems for Home, Suggested, Search, Subscriptions, and Shorts, each with different ranking logic. A video built for Search (optimised for a specific keyword query) may not perform on Home at all, and vice versa.

What changed

The Shorts algorithm now runs for an extended virality window. Virality is no longer capped at 48 hours, and Shorts have been extended to a maximum of three minutes. The more significant change: YouTube now uses Shorts to identify audience fit, then applies that data to long-form recommendations. Channels combining both formats are seeing 41% faster growth than those using only one.

YouTube CEO Neal Mohan identified managing “AI slop” as a 2026 priority. The algorithm is increasingly analysing video content at a deep level: visuals, audio, speech, to assess quality and originality. Low-effort AI content faces the same retention and satisfaction headwinds as any other content that doesn’t satisfy viewers.

The short version

YouTube is fundamentally a search and recommendation engine, not a social feed. The question you should be asking about every video isn’t “will people like this?” It’s “will people feel satisfied after watching this?” Those are different questions. Satisfaction compounds. A video that consistently earns repeat viewing and low “not interested” rates will keep getting surfaced months after it’s published.

Substack

The core mechanic

Substack is the outlier in this list because its algorithm isn’t optimising for the same thing as every other platform. Mike Cohen, Substack’s head of machine learning and the architect of the Notes algorithm, has been unusually transparent about this: the platform is designed to drive subscriptions, not time spent. It doesn’t make money from ads. It makes money when readers become paid subscribers. Every algorithmic decision flows from that.

This produces a meaningfully different system. Notes that lead to subscriptions, either for the posting writer or for publications the reader gets introduced to, are what the algorithm rewards. The vanity metrics of other platforms are essentially irrelevant here.

What it’s measuring

The Notes algorithm weights audience overlap, not just engagement. If you restack someone’s Note and a significant portion of your audience isn’t already subscribed to them, that restack carries real discovery value. Boosting the same group of friends over and over triggers what Substack calls “saturation.” The algorithm actively recognises circular amplification and weights it down.

High-signal interactions, replies that add to the conversation and substantive responses, now outrank hearts and low-effort “love this” comments. The system uses sequential modelling, meaning it’s not just asking what a reader usually likes, but what would be the natural next thing for this reader right now. Context and reading journey matter.

The Notes feed has also shifted significantly. The majority of what appears in a typical Notes feed now comes from publications the reader has never followed, surfaced algorithmically based on overlap with what they do follow. For smaller creators, this is meaningful: your Notes can reach entirely new audiences without any additional action on your part.

What changed

The in-app discovery engine became genuinely functional. In 2025, Substack reported 32 million new free subscriptions driven from within the app over just three months. The platform is now the top source of subscriber growth for many publishers, above external social media. This is a significant shift from the prior advice to drive growth entirely from external platforms.

Notes scheduling was finally introduced natively in late March 2026, which matters practically: you can now batch-produce and schedule Notes without third-party tools.

The short version

Substack’s algorithm is designed to help you find subscribers, not an audience. The distinction matters. Content that demonstrates your actual perspective, not polished, not generic, and that makes sense to pass along to readers who aren’t yet in your world is what the system rewards. The most useful thing you can do on Notes isn’t to broadcast to your existing subscribers. It’s to create something worth encountering for the first time.

What this means across all of them

A few patterns are consistent across every platform right now.

Saves and shares have overtaken likes as the primary signal. Across almost every platform, a passive like is now close to algorithmically worthless. What the systems are trying to detect is high-intent engagement, the kind that suggests someone found real value. Saves say “I’ll return to this.” Shares say “someone else should see this.” Both signal value in a way a like never did.

The first window is decisive. On LinkedIn it’s 30–60 minutes per van der Blom’s research. On TikTok it’s the first few hours before the follower-testing phase concludes. On X it’s the first 30–60 minutes of reply velocity. Every platform runs an early audience test and then decides whether to expand based on what it sees. Posting when your core audience is active and responding quickly in that initial window isn’t optional. It’s structural.

AI-generated content is being actively identified and downweighted. TikTok, Instagram, and YouTube all have originality scoring now. It’s not a philosophical objection to AI tools. It’s a satisfaction problem. Content produced without genuine perspective tends to underperform on retention and satisfaction signals, and the algorithms are getting better at predicting this before the audience even has to tell them.

The platforms aren’t trying to suppress you. They’re trying to match their users with content those users will value. The question is whether what you’re making is actually that.

The irony of asking you to share this, right after explaining that shares are the most valuable signal on every algorithm, isn't lost on me. But if it was useful, send it to someone who needs it.

Share