Social Media Algorithms & Distribution

Your follower count no longer decides who sees your posts in 2026. A recommender system makes that decision, freshly every time. Every post competes against the platform's entire content pool at every single delivery. The decision falls in milliseconds, through a weighted sum of predicted user reactions. Followers shift that calculation, they do not replace it.

This overview sorts what is documented about distribution on Instagram, LinkedIn, TikTok, YouTube and X, what remains a platform statement and what is folklore. That dividing line matters more than any single figure: only X has published ranking code, and even that is partly outdated. Everything else is platform statements, engineering papers or third party analyses that measure correlations and not causes.

How a social media algorithm distributes reach

The large platforms build their recommender systems to the same pattern. First, candidate generation pulls a few thousand candidates out of billions of possible items, usually via embedding search. A cheap light ranking model then makes a first cut. After that, an expensive neural model scores the remaining candidates and predicts several action probabilities for each one at once: like, comment, forward, watch time, report. At the end, a re-ranking step applies diversity rules, business rules and safety filters.

The orders of magnitude are publicly documented. In an engineering paper from October 2025, LinkedIn describes how around 2,000 candidates are preselected from hundreds of millions of posts before the actual ranking starts. According to its 2023 engineering blog, X pulled roughly 1,500 candidates per session for the For You timeline. Meta Engineering describes the same stage for its Andromeda retrieval system: tens of millions of ads are reduced to a few thousand candidates before the ranking model takes over.

The decisive building block is the value model: a weighted function over the predicted action probabilities. The arXiv paper "Choosing the Right Weights" describes this aggregation as a linear combination of the individual signals and puts the weights disclosed by X into context. In practice that means there is nothing to outsmart. There is an objective function that pays more for some user reactions than for others. How the stages interact in detail, what role two-tower models play and why re-ranking often changes more than the ranking itself is covered in the foundational article on the recommendation algorithm.

Anyone coming from SEO will recognise the structure. Google also works with candidate retrieval, scoring and post-processing, just with different signals: comparing it with Google's ranking factors and with engagement signals such as NavBoost makes clear how similar the logic is and where the two part ways.

From the social graph to the content graph

The structural shift of recent years fits into one sentence: delivery follows the question of what interests users who behave like you. Who you follow has become one factor among several. Instagram now separates this explicitly into connected reach (followers) and unconnected reach (recommendations to non-followers). Both use the same signals but weight them differently.

That has two practical consequences. First, follower count is largely worthless as a steering metric. Second, every post is essentially tested from scratch, regardless of how well the last one performed. That explains the high variance that teams often misread as an algorithm change.

Adam Mosseri named the Instagram logic openly in January 2025: watch time, likes and sends are the three most important ranking signals, with likes working slightly harder for content shown to followers and sends slightly harder for content shown to non-followers. Instagram runs a separate ranking system per surface: feed, Reels, Stories, Explore and search. Anyone who wants to optimise "the Instagram algorithm" is therefore always optimising for one specific surface.

At the same time, regulation is giving users more control over distribution. The Digital Services Act obliges platforms to set out the main parameters of their recommender systems in plain language, and very large platforms additionally to offer at least one recommendation option without profiling. TikTok announced in August 2023 that European users would be able to switch personalisation off. In January 2026 Instagram rolled out a dashboard called "Your Algorithm", at first worldwide in English only, through which users can add and remove topics in the Reels feed. The more users make use of it, the more important content becomes that also works without personalisation: clearly named topics, search terms, evergreen substance.

Signals that count across platform boundaries

Four families of signals appear in every documented system.

Dwell time and watch time: for video, watch time is the dominant signal. Instagram measures playback as total time including replays, not as a count of short views. LinkedIn uses dwell time as a quality signal and describes a dedicated long-dwell classifier for passive consumption in the LiRank paper.

Private forwards: sends per reach, meaning forwards by direct message, are weighted slightly higher than likes on Instagram for reach among non-followers. In a 2025 conversation with Business Insider, Mosseri pointed out that considerably more photos and videos are shared in DMs than in Stories or the feed. This distribution through private channels never shows up in full in any analytics interface. The trend is measurable: across more than 70 million posts, Socialinsider counted 45% more shares per post on TikTok and 12% more on Instagram, while comments per post fell by 24% on TikTok and by 16% on Instagram.

Conversation instead of reaction: on X the effect is visible in the code, on LinkedIn in third party analyses. On the basis of more than 620,000 posts, AuthoredUp concludes that comments carry roughly twice the weight of likes. The figure circulating in creator circles that "comments count 15 times over" has no verifiable primary source.

Originality: reposts and aggregated third party content lose distribution. The one hard figure Instagram gives is that 75% of recommendations in the US come from original posts; accounts that mostly upload other people's content no longer qualify for recommendations and become eligible again only once the majority of their posts in the rolling 30 day window counts as original. In 2024 Instagram still named a fixed threshold for this: accounts with 10 or more reposts of other people's content in 30 days were removed from the recommendation surfaces. The current rule works with the majority share in the rolling window instead.

Platform by platform: what is actually documented

Instagram and Facebook. The three Mosseri signals are the most solid basis available for Meta. The frequently quoted claim that sends weigh three to five times as much as likes, by contrast, comes from tool vendor guides and is an interpretation, not a value published by Meta. On the advertising side Meta has rebuilt its infrastructure: according to Meta, the foundation model GEM is 4 times more efficient than the previous generation and raised ad conversions on Instagram by 5% in the second quarter of 2025. Because GEM learns from organic and paid interactions together, organic creative testing becomes a direct input for paid performance, which calls the split between the budgets into question and is examined more closely in the paid social fundamentals. Facebook sits at an average reach rate of 1.65% against 3.50% on Instagram and, in our assessment, stays relevant in the DACH region mainly for older audiences, groups and local visibility. Details on signals, surfaces and the originality policy are in the article on the Instagram algorithm 2026.

LinkedIn. For DACH B2B, LinkedIn is the central channel, and it is there that the ranking has shifted most clearly: away from virality, towards topical fit and demonstrated expertise. Daniel Roth, VP and Editor in Chief at LinkedIn, described the goal as reaching not the most people but the right ones. Operationally that means clarity in a niche beats breadth. The evidence on the decline in reach is directionally consistent but methodologically uneven. For 2025 Richard van der Blom reports, on the basis of 1.8 million posts, views minus 50%, engagement minus 25% and follower growth minus 59%, and in his October update covering around 400,000 profiles visibility minus 47%, engagement minus 39%, follower growth minus 42%. A separate analysis from Q4 2025 puts median impressions at minus 65% since 2023. None of these figures comes from LinkedIn itself, and base years and methods differ. Solid as a trend, useless as a target. In parallel, visibility is concentrating: the top creators' share of feed visibility rose from 15% to 31%, while the rest fell from 57% to 28%. Time windows and lifespan have changed as well: the first 30 to 60 minutes determine the reach trajectory, and good posts stay visible in the feed for 2 to 3 weeks, and according to product lead Gyanda Sachdeva, LinkedIn is explicitly testing bringing older posts that remain relevant back to the top. What that means for editorial planning and topic choice is in the article on the LinkedIn algorithm.

TikTok. In 2022 ByteDance published Monolith, the architecture of its recommender system: online training and a collision free embedding table. The paper mentions neither TikTok nor the For You page. It describes the construction that explains how quickly trends get picked up: the model keeps learning as it goes instead of in batches. In its 2020 explainer TikTok itself names completion and watch time as the strongest signals, followed by shares, likes and comments. The originality rule has been tightened: since 13 September 2025, unoriginal or reused content without a contribution of your own is no longer eligible for the For You feed. For search, TikTok provides its own keyword tool with Creator Search Insights. Architecture, documented signals and traffic source analysis are covered in the article on the TikTok For You page.

YouTube. The basic formula for long-form remains click-through rate times average view duration, plus the contribution to the session. YouTube writes in its own help pages that a high click-through rate with short view duration counts as a clickbait signal and leads to fewer recommendations. Todd Beaupré, Senior Director of Growth and Discovery at YouTube, puts not the behaviour of viewers at the centre of this but how they rate the time they invested. Shorts are judged differently. There, instead of click-through rate, the ratio of viewed to swiped away counts, plus completion and loops. The retention thresholds in circulation of around 65% under 30 seconds and around 50% at 30 to 60 seconds come from trade sources, not from YouTube. For B2B there is a secondary benefit: YouTube material gets cited by AI systems. In an analysis covering 75,000 brands, Ahrefs finds the strongest single correlation between mentions of a brand on YouTube and its AI visibility, at 0.737. What is measured is mentions of the brand on YouTube, not the performance of your own channel, and it remains a correlation. YouTube is also, at 23.5%, the third most cited source in LLM answers. Small channels get through in YouTube search too: a Semrush study from 2021 covering 15,000 keywords found that 18% of the top 10 search results come from channels with fewer than 1,000 subscribers. Surface logic, satisfaction signals and the decoupling of Shorts and long-form are covered in the YouTube article.

X. X remains the only platform with disclosed ranking code and is therefore the best reference for anyone who wants to understand what a value model looks like. The candidate sources from in-network search, graph traversal, SimClusters and TwHIN are documented in the repository. More revealing are the weights in the Heavy Ranker, as published in 2023: a reply weighs 13.5, a reply answered by the author 75.0, a like 0.5 and a repost 1.0, while a report counts for minus 369. Conversation carries the ranking on X, everything else is trimming. The subscription boost is documented as well: Premium accounts get a multiplier of 4.0 in network and 2.0 outside it in the 2023 code. In the updated main branch this multiplier no longer exists under the same name. Since a commit in September 2025, Grok based components for topic assignment and filtering are visible in the For You code. What the code shows and what it does not is in the article on the disclosed X algorithm.

Formats decide the distribution

Format choice is the lever with the greatest effect, because different formats run into different distribution systems. Across more than 4 million posts, Buffer measures that Reels achieve 36% more reach than carousels and 125% more than single images. For the median engagement rate the order flips: carousels lead at 6.9%, ahead of single images at 4.4% and Reels at 3.3%.

Metricool confirms the direction: in an analysis of 24.3 million Instagram posts from 375,000 accounts, average Reels watch time more than doubled year on year to 8.5 seconds, and carousels generate nine times more saves than single images. Single images, by contrast, are losing across the board: reach minus 21.96%, interactions minus 25.41%, engagement minus 45.98% comparing January and February 2025 with 2026. Metricool's LinkedIn analysis of 673,658 posts from 63,108 accounts shows a different picture: polls achieve the highest average impressions on company pages, even though they make up less than 1% of posts, with carousels and multi-image posts ahead of single images.

These figures are correlations, not causality. Post nothing but polls and you get a credibility problem rather than reach. The format decision for Instagram is covered in the article Reels vs carousel, production and platform native re-editing in the content creation overview.

Frequency, half-life and what consistency actually delivers

Posting frequency is the topic with the best data and the most wrong conclusions. Buffer analysed 2.1 million Instagram posts from 102,000 accounts: 3 to 5 posts per week more than double the follower growth rate compared with 1 to 2 posts, 6 to 9 reach around 3.7 times and 10 or more around 5.5 times, while reach per post rises by only 12% at 3 to 5 and by 24% at 10 posts per week. Marginal returns fall off sharply above 5 posts.

The more important finding is the counter-test: in an analysis of 4.8 million channel weeks, weeks without posts underperform the channel's own baseline. Missed weeks cost more than extra posts above five bring in. For a two to five person marketing team in the DACH region this is a capacity question: two channels with a reliable cadence beat five with gaps. How frequency and content half-life differ per platform, and what that means for editorial planning, is in the article on posting frequency and half-life.

Personal profile or company page

For DACH B2B this is the most expensive single decision, because it determines budget, roles and governance. The evidence is clear enough to make it. Van der Blom's analysis puts company pages at 1 to 2% organic reach relative to their followers. An analysis by Whitehat puts the share of organic company page posts in the feed at around 2%, while personal profiles achieve 2.75 times more impressions and 5 times more engagement.

The mechanism behind it is the weighting of peer interactions and early conversation. Van der Blom's analysis shows that posts with three or more commenters in the first 60 minutes get around 5.2 times more reach, measured across 1.8 million posts. This is exactly where the grey area starts: coordination within the team is legitimate, automation is not. In March 2026 LinkedIn explicitly announced that it would act against comment automation, engagement pods and unauthorised third party tools.

Be careful with advocacy figures: the much quoted 561% more reach and 7 times higher lead conversion come from vendor analyses with very small samples and have not been independently reproduced. Directionally plausible, useless as a business case. On the paid side, Dreamdata supplies one of the few solid comparative figures for B2B: LinkedIn reaches a ROAS of 121% against 67% for Google Search and 51% for Meta, and takes 41% of B2B advertising budgets. Benchmarks, peer weighting and the question of how an advocacy programme works without fake engagement are covered in the article company page or profile. On the DACH special case XING: New Work SE reports 22.5 million registered members in the DACH region and names the platform's repositioning as a jobs network as its strategic focus. As a content channel for B2B, XING barely matters any more in our assessment; the comparison LinkedIn vs XING sets out the trade-off.

Reading benchmarks without lying to yourself

Benchmark figures scatter dramatically because they use different denominators. For Instagram, Socialinsider measures an engagement rate relative to followers of 0.45% in Q1 and Q2 2026, Emplifi reports for the same platform a fall in the median engagement rate from 16.9% in Q1 2024 to 9.7% in Q4 2025, and Hootsuite names 3.5% for Instagram, 3.4% for LinkedIn, 1.8% for X, 1.5% for TikTok and 1.3% for Facebook. The three values measure different things: engagement per follower, median per post within one customer base, average across industries.

The more useful measure is engagement by reach. Socialinsider argues for it on the grounds that this formula shows how relevant a piece of content was for the users actually reached, instead of measuring it against a follower count that has less and less to do with delivery. Orders of magnitude per platform and format, including a methodology column, are supplied by the article what is a good engagement rate; how you build a KPI set from that which lets you steer instead of only report is covered in the analytics and KPI overview.

For the reality in the DACH region, the distribution of usage matters more than any engagement rate. According to GWI data for Germany (fieldwork Q2 2025), 85.1% of internet users aged 16 and over use WhatsApp monthly, 58.9% Facebook, 56.2% Instagram, 48.2% YouTube and 34.0% TikTok. On time spent the picture reverses: TikTok comes to around 75 minutes a day, YouTube to more than 30, Facebook to 23, Instagram to 21.5 and WhatsApp to 18.2 minutes. Instagram's potential advertising reach in Germany stands at 31.3 million, down 2.3% between October 2024 and October 2025, while YouTube reaches an ad reach of 64.7 million. Which platform carries which buying centre is sorted out in the platform comparison.

One example of careless handling of numbers is the perennial statistic that 75% of B2B buyers use social media to support purchasing decisions. It comes from an IDC study from 2014 with 760 respondents in eight countries and is not DACH specific. As an argument in a board paper it is worthless.

Shadowban, cold start and other folklore

New posts get tested on small audiences, and the quality of the early signals decides whether they scale. That is plausible, but no platform has documented it as a primary source. Fixed thresholds even less so. There is no documented "golden hour" with fixed target values, no evidenced tiered TikTok test pool with 500 views as stage one and no cross-platform reach penalty for external links. TikTok itself describes no such thresholds in its explainer on video recommendation.

On X, a downgrade in the search ranking is built in through a spam scoring function in the code. The For You code published in 2023, by contrast, holds no general link penalty; what it does hold explicitly is the Premium multiplier. On LinkedIn and Instagram a link penalty is widely reported, but the evidence is inconsistent. The pragmatic approach remains putting the link in the first comment or the bio, without overstating the effect.

The shadowban narrative itself is usually a misreading of cold start variance. Platforms reduce reach on a rule basis and name some of those rules, for example for unoriginal content or for undisclosed AI content. Which effects are documented and which are not is separated out in the article what is a shadowban.

Search is the second distribution path

Discovery is moving from the feed into search, both inside the platforms and upstream of them. For the first four months of 2026, SparkToro measures that 68.01% of US Google searches end without a click, against 60.45% in 2024. Anyone planning distribution can no longer rely on attention converting into website traffic. The value has to be created on the platform itself.

Inside the platforms it looks like this: TikTok states that 23% of users start a search within 30 seconds of opening the app and 57% use TikTok search, and it has introduced search ads campaigns with keyword suggestions there as well. In the US, according to an Adobe survey published in February 2026, 49% of respondents have already used TikTok as a search engine, up from 41% in 2024, while the share of Gen Z who prefer TikTok over Google fell from 8% to 4%. Social search is growing but is not replacing Google. Solid comparative figures for the DACH region are missing.

At the end of 2024 Instagram removed the option to follow hashtags and has been more keyword driven since. LinkedIn has described its retrieval as LLM based since 2025. In practice that means the same thing for all three platforms: terms belong in the caption, the on-screen text, the spoken text and the file names, not in a cloud of hashtags at the end. How the search surfaces work on each platform is covered in the article on social SEO.

Regulation has become part of the distribution logic

Two pieces of legislation reach directly into distribution. The DSA requires disclosure of the main parameters and, for very large platforms, a profiling free feed option. Article 50 of the EU AI Act imposes transparency and labelling duties for AI generated content from 2 August 2026. This is not purely a compliance question: unlabelled AI content additionally risks sanctions from the platform. YouTube makes clear in its own documentation that labelling restricts neither the reach nor the monetisation of a video, while repeatedly missing disclosures can result in a label being applied for you, the removal of the video or exclusion from the partner programme. What works in production with AI and what costs distribution is described in the existing knowledge base under AI assisted content creation.

What follows for how you steer

A short decision routine can be derived from the evidence. It does not replace a strategy, but it prevents the most common misinvestments.

The practical core is unspectacular: produce the signals the value model pays well for, and measure them against reach, not against follower count. Everything else is interpretation of numbers that nobody outside the platforms really knows. The principles last longer than the individual figures, and the individual figures in this text are as of August 2026.