How Social Media Algorithms Work in 2026: The Shared Mechanics Behind Every Feed

Social media algorithms decide what each person sees in four steps: they gather a pool of candidate posts, predict how that person will respond to each one, rank the pool by predicted value, and learn from what actually happens. Every major network works this way. What differs is where candidates come from and which reactions count most.

Algorithm advice is full of folklore: magic posting times, secret hashtag counts, penalties for editing or for using a scheduler. Most of it falls apart once you read what the platforms themselves have published. Meta explains each of its feeds on its Transparency Center. X published its recommendation code on GitHub. LinkedIn described its new feed in an engineering post in March 2026. TikTok, YouTube, Pinterest and Bluesky have their own explainers. This guide pulls those sources together into one model of how feeds work, then breaks it down by platform, by signal, by surface and by account size, and ends with a method for working with any algorithm and diagnosing a drop in reach.

Key findings

• Feeds narrow thousands of posts to a few hundred before ranking. Facebook Feed runs a lightweight model to select "approximately 500 of the most relevant posts" and scores those (Meta). Instagram Feed narrows to about 700, then scores about 500 (Meta). Instagram Reels narrows to about 100 (Meta).
• Explore casts the widest net. Instagram Explore retrieves up to 1,500 items, then passes "the top 100 media" to its final ranking stage (Meta).
• The first three seconds are a named prediction. Instagram's Reels ranking predicts "how likely you are to watch less than three seconds of a reel" and whether you will watch more than 95% of it (Meta).
• Follower count is not a direct factor on TikTok. TikTok says "neither follower count nor whether the account has had previous high-performing videos are direct factors" in recommendations (TikTok).
• X scores each post as a weighted sum of predicted actions, including negative ones such as "not interested", mute, block and report, then applies a discount to out-of-network posts and a boost for new authors (X algorithm repository).
• LinkedIn's feed now uses LLM-based retrieval and a sequential ranking model that reads "more than a thousand" past interactions and serves more than 1.3 billion professionals (LinkedIn Engineering, March 2026).
• YouTube's system learns from more than 80 billion signals, including satisfaction surveys that help it "understand satisfaction, not just watch time" (YouTube Help).

What's changed in 2026

The basic machinery has not changed in years. Three things around it have.

Platforms published more detail than ever

X released its recommendation code, with updates logged in August and September 2026, showing exactly which actions it predicts and how scores combine (GitHub). LinkedIn explained how it replaced several separate candidate sources with "a unified retrieval system leveraging advances in LLMs" (LinkedIn Engineering). Meta keeps per-surface explainers for every Facebook and Instagram feed. Guesswork is less necessary than it was.

Originality became a ranking input

Meta updated its Facebook guidelines in March 2026 so that accounts which mostly post unoriginal content "may be deemed non-recommendable", and said views and time spent on original Reels "approximately doubled" in the second half of 2025 compared with a year earlier (Meta). Instagram extended similar rules to photos and carousels in April 2026 (Instagram). LinkedIn said in September 2026 that its classifiers "limit distribution of AI slop" and reported a 40% drop in views for content classified that way (LinkedIn).

Users got more direct control

Threads launched "Dear Algo" in February 2026, which lets people "tell Threads what topics you temporarily want to see more or less of" for three days by posting a request (Meta). Pinterest added a toggle that reduces generative AI Pins in the home feed (Pinterest Help). Bluesky has offered choosable custom feeds since 2023. The more control people have, the more reach depends on content they actively want rather than content that games a single ranking.

The four stages of almost every feed

Whatever the network, a feed is built in roughly the same four stages each time someone opens the app.

Stage 1: gathering candidates

The first question any feed asks is not "what is best?" but "what is possible?". For a following feed, candidates are recent posts from accounts you follow. For a recommendation feed, candidates come from accounts and content similar to what you engage with. Instagram Explore, for example, uses "author-based sources (media from authors that you've engaged with)" and "media-based sources (media similar to media you've engaged with)" (Meta). X uses an in-network source of recent posts from accounts you follow, and an out-of-network source that "embeds the viewer and each post as vectors, and returns the posts nearest the viewer" (X).

Stage 2: filtering and narrowing

Ranking every possible post with a heavy model would be too slow, so feeds first use a light model to narrow the pool. This is also where integrity processes remove or demote posts that break rules or are non-recommendable. If your account is non-recommendable, for example because most of its content is unoriginal, it can drop out here, before your post is ever scored.

Stage 3: prediction and ranking

This is the heart of every algorithm. The system predicts how likely this person is to take each of many actions on each candidate post, then combines those predictions into one score. Positive predictions (watch most of it, comment, share, follow) push a post up. Negative predictions (skip, hide, "not interested", mute, report) push it down. The weights differ per network and change often, but the shape is the same everywhere.

Stage 4: learning from feedback

After a post is shown, the actual responses feed back into the models. This is why early responses matter: a post that earns strong reactions from its first viewers becomes a better candidate for the next ones. It is also why one strong post can lift the next: the system learns who your content is for.

Breakdown by platform

Here is what each network has published about its own system, and the practical implication of each.

Facebook

Facebook Feed gathers posts from friends, Pages and groups, narrows them to about 500, and ranks those by relevance score while aiming for a mix of content types. The predictions Meta lists include whether you will scroll past or engage, whether you will click an outside link and how long you will spend there, and whether you will interact with friends or groups, using signals such as "past interactions with the post's author" (Meta). Implication: posts that start conversations among people who know each other travel furthest. See our guide to how the Facebook algorithm works.

Instagram

Instagram runs separate systems for Feed, Reels, Stories and Explore. Feed predictions include whether you will swipe through a whole carousel, spend time on the author's profile after seeing a post, and send a post to someone in a direct message (Meta). Reels predictions include watching under three seconds, watching more than 95%, resharing, following the author and using the audio (Meta). Implication: hooks, completion and shares matter most. Our guide to how the Instagram algorithm works covers each surface.

Threads

Threads gathers "a portion of the public content available on Threads" plus posts from accounts you follow, and predicts likes, replies, profile clicks, time spent viewing a post and scrolling past (Meta). Implication: Threads is a recommendation feed first; conversation and replies extend reach. See how the Threads algorithm works.

TikTok

TikTok weighs user interactions most, then video information such as captions, sounds and hashtags, with device and account settings given "lower weight". It says it does not recommend "duplicated content, content you've already seen before, or any content that's considered spam" (TikTok). Implication: a new account with a strong video can reach far beyond its followers. Read more in how the TikTok algorithm works.

X

X's For You feed blends in-network posts with out-of-network candidates, ranks them with a transformer model that reads the viewer's recent engagement history, and combines predicted actions as a weighted sum. It then applies an author diversity decay for repeat authors, a discount for out-of-network posts, and a boost for new authors (X). Implication: replies, reposts and dwell time all count, and negative feedback costs you. Our X algorithm guide goes deeper.

LinkedIn

LinkedIn retrieves candidates using LLM embeddings built from profile data such as industry, skills and headline, plus engagement history, and ranks them with a "Generative Recommender" that models a member's sequence of past interactions. Ranking actions include long dwells, likes, comments and shares, and the post says ranking relies on "professional signals and engagement patterns, not demographic attributes" (LinkedIn Engineering). Implication: posts that make the right professionals stop and read reach further. See how the LinkedIn algorithm works.

YouTube

YouTube compares your viewing habits with similar viewers' and learns from "over 80 billion signals", including watch history, search history, subscriptions, likes, dislikes, "Not interested", "Don't recommend channel" and satisfaction surveys (YouTube Help). Implication: titles and thumbnails must deliver what they promise, because satisfaction is measured. See how the YouTube algorithm works.

Pinterest

Pinterest recommends Pins "based on your saves, searches, and browsing activity", and lets people turn off activity-based recommendations or reduce AI content (Pinterest Help). Implication: Pinterest behaves like a search engine, and Pins can surface for months. See how the Pinterest algorithm works.

Bluesky

Bluesky offers "an open marketplace of feeds": anyone can build a custom feed, and users choose and pin the feeds they want instead of relying on one ranking algorithm (Bluesky). Implication: being included in the right topic feeds, and posting regularly for followers, matters more than gaming a single ranking. See how the Bluesky algorithm works.

Breakdown by signal

Platforms rarely publish exact weights, and the weights change often. The categories of signal, though, are consistent across every published source.

Relationship

How often a person interacts with your account. Facebook names "past interactions with the post's author"; Instagram names "authors you've followed, liked, commented on, or clicked on". People who regularly engage with you are more likely to see your next post, which is why replying to comments pays twice: once in the conversation, and again in the next ranking.

Interest

How well a post's topic matches what the person has engaged with. This is what lets recommendation feeds show your post to strangers. Clear, consistent topics help the system learn who your content is for.

Attention

Watch time, completion, dwell time on a text post, time on the author's profile. Instagram predicts watches under three seconds and over 95%; X predicts dwell time and video continuation; LinkedIn ranks on long dwells. The first second or first line decides most of this.

Meaningful actions

Comments, shares, saves, sends and follows usually count for more than a quick like, because they are rarer and show stronger intent. Instagram specifically predicts how likely you are to send a post in a direct message.

Negative feedback

Hides, "not interested", skips, mutes, blocks and reports reduce distribution. X lists "not interested, mute, block, report, not dwelled" as predicted negatives in its scoring. Engagement bait and clickbait that disappoints readers earn these.

Freshness

Most feeds favour recent posts. LinkedIn says its pipelines keep content current "within minutes". YouTube and Pinterest, which work more like search, surface older content for much longer.

Originality and quality

Increasingly, platforms filter for originality before ranking: Meta's non-recommendable rules, Instagram's limits on watermarked and recycled reels, LinkedIn's limits on "generic, repetitive, recycled" content.

Breakdown by content type

Because ranking follows predicted responses, the format that earns the most responses differs per network. Buffer's analysis of more than 45 million posts, published in March 2026, shows how much (Buffer).

Short video

Short video is built for recommendation feeds, where completion and rewatches are explicit predictions. It reaches furthest: Buffer found Instagram Reels reach 2.25 times as many people as single images. But reach is not engagement, and on several networks other formats earn a higher engagement rate.

Carousels

Carousels earn time spent and completion. Instagram predicts "how likely you are to swipe through all of a carousel's cards", which explains why carousels top its engagement rates. Our carousel best practices and LinkedIn carousel vs text post guides go deeper.

Images

Single images are quick to take in, which suits following feeds. They led on Facebook in Buffer's data. They earn less dwell time than carousels or video, so the image has to say something on its own.

Text

Text earns dwell time and replies, which is why it leads on X. On LinkedIn, long dwells are a named ranking action, so a text post that people stop to read can travel far. The first line carries the whole post.

Links

Algorithms do not ignore links; they predict what happens next. Facebook predicts whether you will click an outside link and how long you will spend there, using your "past click rate on external links" (Meta). X predicts link clicks among other actions. A link post works when the people who see it are likely to click and stay, so give the post enough context to earn that click, and tag links with a UTM builder to see what the click led to.

Breakdown by audience: B2B versus B2C

The same machinery produces different results for different audiences, because the predicted responses come from different people.

B2B audiences

Professional audiences are smaller and more specific. On LinkedIn, retrieval matches posts to members using profile data such as industry, skills and headline, so a post written for a defined professional reader reaches that reader more reliably than a broad one. Dwell time matters because professionals read before they react. Timing follows the working week; our guide to the best time to post on LinkedIn has the data.

B2C audiences

Consumer audiences are larger and browse in recommendation feeds. Short video, carousels and shareable images dominate, and sends to friends are a strong signal. Interest matching does the heavy lifting: a clear topic, such as recipes, skincare or home repair, helps the system find people who watch that kind of content.

Local businesses

Local audiences overlap with friends and community. On Facebook, interactions with friends and groups are a named prediction, so local posts that people tag friends in or discuss in groups do well. Location in captions and community events give the interest models something to match.

Creators

Creators are the main audience recommendation feeds are built around. Relationship signals grow fastest for creators who reply, run series and give viewers a reason to follow, which is itself a named prediction on Instagram Reels.

Breakdown by surface

Each app has several feeds, and each one ranks differently. Knowing which surface you are aiming for changes what to optimise.

Following feeds

Feeds of accounts you follow, such as Instagram Feed or X's Following tab, rank mainly on relationship and freshness. Your followers see you if they engage with you. Instagram's originality rules do not affect how content is shown to followers, per its April 2026 update. Consistency and conversation matter most here.

Recommendation feeds

For You, Reels, Explore and the Threads feed show posts from accounts people do not follow. These rank mainly on interest and attention, and they apply recommendability filters. This is where growth comes from, and where originality and format fit matter most.

Search

YouTube, Pinterest, TikTok and increasingly Instagram act as search engines. Here, words in titles, descriptions and captions matter, and content keeps finding viewers for months. Pinterest bases recommendations on searches as well as saves. YouTube lists search history as one of the signals that shape recommendations, so a video found through search can keep feeding recommendations later. For search surfaces, write titles and descriptions in the words your audience actually types, put the main keyword early, and describe what the viewer will get. A clear, specific title helps both the search ranking and the satisfaction signal, because the people who click are the people the video was made for.

User-chosen feeds

Bluesky custom feeds and Threads' Dear Algo requests let people steer what they see. On these surfaces, being clearly about a topic gets you into the feeds people pick.

Breakdown by account size

New and small accounts

Recommendation feeds give small accounts a real chance, because responses matter more than follower count. TikTok says so directly, and X's scoring includes a boost for new authors. The work is to make each post as strong as possible in its first seconds and to post consistently, so the system gets enough examples to learn who your content is for.

Growing accounts

Mid-size accounts have enough followers that relationship signals start to matter. Replying to comments, asking questions people want to answer and building a recognisable series turn occasional viewers into repeat engagers, which lifts every later post.

Large accounts

Large accounts get many candidates into feeds through following alone, but face diminishing returns: X applies a diversity decay to repeated posts from the same author, and Facebook aims for a mix of content types. Posting more is not always better; posting content that keeps earning meaningful actions is.

The small-audience test

Most recommendation feeds work like a series of tests. A new post is shown to a small group first. If they respond well, it becomes a better candidate for a larger group, and so on. That one idea explains most practical advice:

• Why hooks matter: the first second or line decides whether the test audience stays, and Instagram literally predicts under-three-second watches.
• Why a small account can go viral: distribution follows response, not only follower count.
• Why replying early helps: early conversation is a strong signal while the post is being tested.
• Why consistency matters: each post is another test, and more tests teach the system who your content is for. Buffer's 2026 analysis of more than 52 million posts found that accounts that skipped a week underperformed their own baseline growth (Buffer).

Common algorithm myths

How to work with any algorithm: an 8-step method

1. Pick your surface. Decide whether a post is for followers, for recommendations or for search, and shape it for that surface.
2. Earn the first seconds. Lead with the most interesting part. On video, the first frame; on text, the first line.
3. Use the native format. Vertical video for short-video feeds, carousels on Instagram and LinkedIn, strong text on X and Threads, 2:3 on Pinterest.
4. Aim for meaningful actions. Make content worth saving, sending, sharing or replying to.
5. Start conversations and reply. Replies build relationship signals for your next post.
6. Stay on topic. Clear content pillars help platforms match you with the right audience.
7. Stay consistent. Regular posting gives each system steady data. Our guide to staying consistent shows how.
8. Measure your own results. Your data beats generic advice. Track reach and engagement rate per post, per format, and see our metrics and KPIs guide.

What to do when reach drops

Before blaming the algorithm, work through these checks in order.

Worked example: one post through the algorithm

A fictional neighbourhood bakery, Rise and Crumb, posts a 25-second Reel of a croissant being laminated at 7:30 in the morning.

1. Candidates: the Reel becomes a candidate for followers' feeds and Reels tabs, and for non-followers who watch baking or local food content.
2. Narrowing: it is original, clean, vertical and unwatermarked, so it passes integrity checks and makes the shortlist for interested viewers.
3. Prediction: for a follower who often likes the bakery's posts, the predicted chance of watching most of it is high. For a stranger who watches pastry videos, the system predicts from their history.
4. Early feedback: the first viewers watch most of the video, several rewatch the butter fold, and a handful send it to friends.
5. Wider distribution: strong completion and sends make it a better candidate for more viewers like them.
6. Learning: the account gains a stronger association with pastry content, which helps the next similar video find its audience.

Now imagine the same bakery posts a landscape clip with a TikTok watermark and a slow ten-second intro. It may never pass the narrowing stage for non-followers, and the followers who do see it skip within three seconds. Same bakery, same croissant, very different result.

Nothing in that chain depends on a secret hashtag count or a magic minute. It depends on a strong opening, the right format, original content and the bakery posting again tomorrow.

How algorithms treat AI-generated content

AI writing and image tools have flooded feeds with polished but generic posts, and platforms have responded in two ways. First, by ranking: LinkedIn describes "AI slop" as content that "may sound polished on the surface but lacks a clear point of view", lets members flag it, and limits its distribution (LinkedIn). Second, by user choice: Pinterest's Show AI content toggle lowers the amount of generative AI Pins in a person's home feed (Pinterest Help). Neither platform penalises a post because a tool helped write it. Both penalise content that adds nothing. The practical rule: use AI for drafts if it helps, then add your own experience, data or opinion before anything is published.

A weekly routine that works with every algorithm

You do not need a different strategy for every algorithm. You need a routine that produces the signals they all reward.

A scheduler handles the Monday batch and the cadence, which frees the daily time for replies, the part no tool should do for you.

Where scheduling fits

Algorithms reward consistency, native formats and engagement. Scheduling helps with all three. Batching gives you time to work on hooks and formats instead of posting in a rush, scheduling keeps your cadence steady through busy weeks, and the time you save goes into replies. In Postprism, you plan every network on one content calendar, adapt each network's version in the same editor, and publish to all 6 networks on time. If you share one idea across networks, read does cross-posting hurt your reach first.

Common mistakes when trying to beat the algorithm

1. Chasing hacks instead of responses. Fix: focus on what makes your audience stop, watch and reply.
2. Weak openings. Fix: put the best moment first; Instagram predicts under-three-second watches.
3. Engagement bait. Fix: ask real questions; bait earns negative feedback.
4. Uploading watermarked or reposted files. Fix: upload your own clean originals.
5. Posting in the wrong format. Fix: match each surface's native format.
6. Going silent for weeks. Fix: schedule a steady cadence you can sustain.
7. Ignoring comments. Fix: reply in the first hour; it builds relationship signals.
8. Drifting across too many topics. Fix: keep three to five content pillars.
9. Judging by one post. Fix: compare medians over four to six weeks.
10. Stuffing hashtags. Fix: a few relevant tags where the network uses them.
11. Posting generic AI copy. Fix: edit for a point of view; LinkedIn limits "AI slop".
12. Misleading titles and thumbnails. Fix: deliver the promise; YouTube measures satisfaction.

Frequently Asked Questions

How do social media algorithms decide what I see?

They gather candidate posts from accounts you follow and content similar to what you engage with, predict how you will respond to each, rank them by predicted value, and learn from what you actually do.

What is the most important signal for social media algorithms?

There is no single one, but predicted engagement from the specific viewer is central everywhere: watching most of a video, commenting, sharing, saving, sending and following. Negative signals such as skips and hides count against you.

Do scheduled posts get less reach?

No published ranking explainer lists how a post was published as a signal. Posts published through official integrations are ranked on the same predicted responses as any other post.

Does follower count matter for reach?

Less than people think. TikTok says follower count is not a direct factor, and X boosts new authors. Followers help on following feeds, but recommendation feeds rank on response.

Why did my reach suddenly drop?

Usually a change in consistency, format, topic or originality, an unusually strong earlier post, or a platform-wide change. Compare median reach over several weeks before changing everything.

How often do social media algorithms change?

Constantly. Meta's explainers note its models and signals "change frequently". The categories of signal, such as relationship, interest, attention and meaningful actions, stay stable.

Are hashtags still important?

They help categorise and search on some networks, such as TikTok where they count as video information, but they carry less weight than how people respond. Use a few relevant tags.

How can I see why a post was recommended to me?

Meta publishes per-feed explainers on its Transparency Center, Threads lets you steer topics with Dear Algo, and Pinterest lets you refine recommendations in settings.