360Brew LinkedIn Ranking Explained for Practitioners

When SocialNexis tested fixed-interval against randomized-interval action scheduling across logged account sessions in 2025, restrictions arrived measurably faster on the clockwork accounts. That result predates most editorial coverage of 360Brew, and it points at what the architecture paper confirms: LinkedIn stopped counting engagement. It scores behavior now, at 150 billion parameters.

How Does 360Brew's LinkedIn Ranking Algorithm Work?

The architecture explains why old tactics behave strangely now. 360Brew is a decoder-only transformer with 150 billion parameters, built on LLaMA 3 and fine-tuned on approximately 1 trillion tokens of LinkedIn data. It replaced thousands of separate, task-specific recommendation systems with a single unified model capable of solving 30-plus personalization tasks without task-specific fine-tuning. Feed ranking, connection suggestions, job matching, notification selection: one set of weights, one opinion of your account.

That consolidation is the part most practitioner advice skips past, and it is the part that changes how you work. Under the old architecture, a tactic could produce results on one surface and nothing on another, because those surfaces were genuinely different models reading different features. Now the same model reads your profile when it ranks your post, and reads your posting history when it decides whether your profile is worth surfacing to someone else. There is one assessment of you, and it is applied everywhere at once.

LinkedIn's official feed ranking documentation still names three signal categories, and they reward a literal reading: Identity Signals (profile, location, skills), Content Signals (engagement metrics, recency, language, tagged entities), and Activity Signals (reactions, dwell time, follow patterns). The same documentation states that feed distribution is not influenced by third-party payments, with the single exception of labeled promoted content. Anyone selling you guaranteed reach is selling a tactic, not an arrangement.

The category names did not change when 360Brew shipped. The thing reading them did. Content Signals used to resolve to counters: how many reactions, how recent, which entities tagged. Under a 150-billion-parameter language model, Content Signals means the model has read the post, formed a representation of what it is about, and compared that representation against what it believes you post about and what it believes each reader wants to read. Same label, completely different resolution.

The system is working for LinkedIn, which is why you should plan around it rather than wait it out. Time spent per user grew 25% year to date through May 2026, the fastest of any major US social platform. Instagram came second at 6%. Everything else was flat or declining. No platform team rolls back the system producing that number. 360Brew is the permanent condition, and the optimization habits built for rule-based ranking are the things that need to change.

The Two-Stage Pipeline: Retrieval, Then Generative Scoring

Your post is scored twice, by two different mechanisms, and the first one is where most posts quietly die. Stage 1 uses an LLM dual-encoder retrieval system to narrow millions of candidate posts down to roughly 2,000 using semantic embeddings. Stage 2 applies a Generative Recommender that processes 1,000-plus historical user interactions as chronological sequences and produces a personalized relevance score for each surviving candidate. LinkedIn's engineering team described this architecture publicly in March 2026.

Stage 1 is not keyword matching, and treating it as keyword matching is how people waste an afternoon on phrasing. A dual encoder places your post and each reader's interest profile into the same vector space and retrieves by proximity. The words you chose matter only through the meaning they produce. Writing the phrase your audience searches for does nothing if the surrounding content sits somewhere else semantically. A post about hiring written by an account the model has embedded as a supply chain account is a hiring post in the wrong neighborhood.

Stage 2 is where the personalization happens, and it is why there is no such thing as your post's score. The Generative Recommender reads a chronological sequence of over a thousand prior interactions for the specific person whose feed is being assembled, then scores your post against that person's demonstrated behavior. Your post has as many scores as it has candidate viewers. Impressions are the sum of thousands of independent verdicts, not one number assigned at publish time.

The reported gains from this architecture are 15% better recall@10 and sub-50ms retrieval latency at production scale. Recall@10 measures how often the genuinely relevant items land in the top ten retrieved. A 15% improvement there means the candidate pool handed to Stage 2 is materially better targeted than it used to be, which is good for readers and unforgiving for anyone whose content sits outside a clear topical identity.

The failure mode worth naming here is retrieval-stage invisibility. A post that is semantically distant from its author's established topic profile enters Stage 1 at a disadvantage no amount of content quality fixes, because quality is evaluated in Stage 2 and the post never arrives there. In analytics this looks like nothing at all. No suppression notice, no reduced engagement rate, just a post with a few hundred impressions and no obvious defect. People respond by rewriting hooks. The hook was not the problem.

The practical consequence is that retrieval is upstream of everything else you tune. Format, timing, hook structure, and comment strategy all operate on candidates that made it into that 2,000. If you are not being retrieved, you are optimizing a stage you never reach.

Saves Outweigh Likes by 5x to 10x Under 360Brew

Saves are the highest-value engagement signal under 360Brew. Practitioner estimates put one save at 5x to 10x the weight of a like and roughly 2x the weight of a comment, which LinkedIn has not officially confirmed and you should hold loosely as a ratio while treating the ordering as solid. The logic behind the ordering is sound: a like costs nothing and predicts nothing, while a save is a reader telling the model that the content has value beyond the moment of the scroll.

The timing finding is the one that overturns common advice. Posts that collect saves and substantive comments 24 to 72 hours after publishing perform 4x to 6x better in Suggested feeds than posts that spike early and decay within the first hour. The golden-hour framing that dominated LinkedIn advice for years optimizes for exactly the shape 360Brew discounts: a burst of cheap reactions from people who already follow you, followed by silence. A post still collecting saves on day three is making a much stronger claim about its own value.

360Brew also reads the comment section semantically, and this is where pod habits show up. Generic low-entropy phrases like "Great post!" are explicitly downranked. Lexical diversity across a comment section is prioritized, because a thread where twenty people wrote twenty different things is behaviorally distinct from a thread where twenty people wrote the same thing in different orders. The model is scoring the discussion, not counting it.

Coordinated engagement from tight follower clusters is penalized as manufactured relevance. The system tracks comment velocity, account relationships, engagement history, and semantic comment content across network graphs. The failure pattern to name is the uniform comment section: high comment count, near-zero lexical diversity, all from accounts that engage with each other constantly and with nobody else. It reads as a closed loop from the graph side and as templated text from the language side, and those two signals confirm each other.

Replying to comments within 30 minutes of publishing correlates with 64% more total comments and 2.3x views, per AuthoredUp benchmark data. This is not in tension with the 24 to 72 hour finding. Fast author replies are what keep a thread alive long enough to reach that window. The goal is not a first-hour spike, it is a conversation that is still producing new, substantively different comments the next day.

If you want one behavioral change from this section: stop asking for likes and stop engineering the first hour. Write things specific enough that someone saves them, and answer every early comment with a real reply rather than an acknowledgment.

What Most Content Advice Gets Wrong About 360Brew LinkedIn Ranking

Most 360Brew guides describe an engagement counter with new weights attached. That framing produces a checklist: more saves, fewer links, no hashtags. The checklist is mostly correct and almost entirely beside the point. 360Brew is a semantic relevance model, and the question it answers is whether a post's content, its author's profile, and its readers' consumption history are mutually consistent. Consistency is the variable. Everything else is a modifier applied afterward.

Profile-to-content alignment is a direct ranking input, not a soft best practice. A post on a topic clearly supported by the author's headline, About section, and 90-day posting history receives a relevance boost. Inconsistent cross-topic posting dilutes the topic authority model and produces what practitioners have started calling distribution invisibility: content that is fine on its own terms and reaches almost nobody, because the model has no confident answer to the question of what this account is for.

External links carry an approximately 60% reach penalty, and the link-in-first-comment workaround is now penalized as well. That second part is where most advice is out of date, because the workaround circulated for years as settled practice and plenty of guides still recommend it.

We had ground truth on this earlier than most, because our automation was placing those comments. From mid-2025 onward, across SocialNexis-managed accounts, reach on posts where automation placed a link comment within the first 5 minutes degraded at the same rate as posts with the link in the body. Same curve, no meaningful gap. Editorial sources documented the change later. Running the tactic at volume across many accounts means you see detection catch up as a change in your own numbers rather than as an announcement.

The second thing no content-quality analysis can see is delivery method. Posting through LinkedIn's API produces no session-level behavioral fingerprint: no mouse movement, no dwell on the compose screen, no typing cadence. A real browser session produces all of it. SocialNexis hybrid workflows, where a draft is written ahead of time and a real browser session is used to post it, show meaningfully different early distribution patterns than API-posted content with identical text. Same words, different opening trajectory.

We are careful about the causal claim there, because we cannot see inside the ranking model. What we can say is that the text is not the whole input. Anyone evaluating 360Brew purely by reading posts and correlating features is measuring one channel of a system that is also reading how the post arrived.

Build the Topic Authority Profile That 360Brew Scores For

360Brew needs roughly 90 days of consistent posting within 2 to 3 core topic pillars before it classifies an account's topical domain and delivers the full distribution benefit. That is the single most useful number in this guide for anyone starting from a mixed or dormant profile, because it sets the horizon. Three months of deliberate consistency, not three good posts.

The cross-referencing is the mechanism. Your headline, your About section, and your recent posting history are three independent pieces of evidence about the same question, and the model gains confidence when they agree. A headline claiming supply chain expertise, an About section written for a job search four years ago, and a posting history split between hiring takes and travel photos gives the model three different answers. It resolves that by giving your content narrow, low-confidence distribution.

Inconsistent cross-topic posting dilutes the topic authority model regardless of how good the individual posts are. This is the part that feels unfair and is worth sitting with: posting across five unrelated subjects puts every one of those posts at a reach disadvantage, including the ones squarely inside your actual expertise. The dilution is a property of the account, not of the post. You do not get to have one strong pillar and four hobby topics without paying for it on the strong pillar.

Two or three pillars is the working range, and the constraint is tighter than most people want. Pillars should be adjacent enough that a reader interested in one is plausibly interested in another, because adjacency is what lets the embedding for your account stay compact. Three topics that share an audience read as depth. Three topics that share nothing read as noise with a professional headshot.

The advice to vary your content so the feed does not get bored was built for a ranking system that no longer exists. Consistent topic reinforcement is what 360Brew rewards over time, and the reward compounds, because a confident topic classification is also what gets you retrieved in Stage 1 for readers outside your follower graph. Variety within a pillar (format, angle, depth, contrarian versus explanatory) is free. Variety across pillars is expensive.

Format Reach Multipliers Under the Current 360Brew Algorithm

Document and carousel posts in PDF format lead every other format, with a 7.00% average engagement rate measured by Socialinsider across 1.3 million posts and a 1.39x reach multiplier against the platform average. If you want one format decision made for you, that is it. The gap is not marginal and it has held across the largest public dataset available.

The full engagement rate ordering by format: documents at 7.00%, multi-image at 6.45%, video at 6.00%, text at 4.50%, and link posts last at 3.25%. Two things stand out. Text posts, the format most LinkedIn advice treats as the default, sit in the bottom third. And link posts are not merely last, they are the format that carries a separate reach penalty on top of a low engagement rate, which is covered in the next section.

Native video carries a 0.86x reach multiplier, which puts it below the platform average, after declining over 70% in some analyses through mid-2025. That is a full reversal from video's earlier algorithmic advantage, and it is the format where old advice does the most damage, because video is also the most expensive thing to produce. Teams are spending the most effort per post on the format with the worst multiplier.

The wider context matters more than the ranking. Every format declined year over year through mid-2025: video down 72%, images down 45%, documents down 43%, and text down 34%, measured by AuthoredUp across more than 3 million posts. Median post impressions fell 47%, from 1,211 in June 2024 to 636 in May 2025. Documents are the best format in a market where the best format also lost 43%.

That reframes what format selection can do for you. A 1.39x multiplier against a 47% median decline does not restore your 2024 numbers, and treating format as the fix is how people end up rebuilding their entire content operation around carousels while the underlying problem stays untouched. Format is a modifier applied after retrieval. Topic authority determines whether there is anything to modify. Pick documents, then stop thinking about format and go fix the thing upstream of it.

Link Penalties, Hashtag Myths, and What Hurts Reach Most

External links carry an approximately 60% reach penalty, and the link-in-first-comment workaround is penalized at the same rate as a link in the post body. This is the most commonly repeated piece of outdated LinkedIn advice still in circulation. The workaround was real, it worked for a period, and it does not work now.

Our evidence for that came from operations rather than analysis. SocialNexis automation was placing link comments within the first 5 minutes of publishing across managed accounts, which produced a clean natural experiment at volume. From mid-2025, those posts showed the same reach curve as posts with the link in the body. Not a partial recovery, not a smaller penalty. The same curve. We were seeing it in account-level data before most editorial sources documented the change, which is the practical argument for watching your own numbers rather than waiting for someone to publish the shift.

Hashtags are no longer a positive signal. Posts without hashtags outperform posts with them by approximately 8%, and using 10 or more triggers spam suppression. The small negative for a few hashtags makes sense once you accept that a language model reads the post: hashtags are text that carries little meaning and pattern-matches to promotional content. The spam threshold at 10-plus is a separate and much harsher mechanism. If you have a hashtag block at the bottom of every post because it was a good idea in 2019, delete it.

Posting daily incurs a 45% reach penalty, with a minimum 18 to 24 hour gap between posts required to avoid suppression. This one contradicts almost every content calendar template in existence, and it is the constraint most teams break without noticing, usually by scheduling a company post and a founder post on the same account-adjacent rhythm. Five posts a week at proper spacing outperforms seven crammed together, and it costs less to produce.

In rough order of damage: an external link in the body or the first comment, then daily posting cadence, then a stack of 10-plus hashtags, then the weaker per-hashtag drag. Each of these is a self-inflicted multiplier applied to content that already had to survive retrieval. None of them are worth what they cost. The click you were protecting with a link in the body is worth less than the 60% of the audience who never saw the post.

When AI Content Gets Suppressed and When It Does Not

360Brew actively detects AI-generated content and suppresses reach for posts identified as raw, unedited AI output. The scoring inputs are specific and, usefully, they are things you can check yourself: specificity (named examples, exact figures, specific dates), sentence-length variance, and vocabulary repetition rate. Low repetition and high variance read as human. Uniform medium-length sentences with generic nouns read as a model.

The distinction that determines your outcome is what your editing pass adds, not how much of it you do. Rewording, reordering, and synonym substitution leave every measured property intact: the sentences stay the same length, the vocabulary stays as repetitive, the specificity stays at zero. Adding a named client, an exact figure from your own data, the date something broke, or a framework that did not exist in the draft changes the properties the model is scoring. Volume of editing is not the variable. Introduced specificity is.

This is where our own workflow produced something we have not seen anyone else isolate. SocialNexis hybrid workflows, where AI drafts the content and a real browser session delivers it after human editing, show different early distribution than API-posted content with identical text. The compose session generates a behavioral fingerprint (variable typing speed, cursor dwell on the preview) that API posting does not, and it appears to influence the size of the initial seed audience.

No competitor editorial source has isolated this interaction, and the reason is structural rather than a matter of effort. Analyzing AI content detection by reading posts and correlating textual features can only ever measure the content channel. Separating text from delivery method requires posting the same text two ways across enough accounts to see the difference, which requires running the automation. That is what we have and content analysts do not.

The working recommendation, stated plainly: drafting with AI is fine, shipping an unedited draft is not, and the edit that saves you is the one that adds facts only you have. If your editing pass did not introduce a number, a name, or a date that was absent from the draft, you have not changed anything the model measures. And if you are posting through an API integration because it is convenient, understand that convenience has a cost on the behavioral side that no amount of content quality compensates for.

Automation, Engagement Pods, and Behavioral Fingerprinting

Engagement pods and coordinated like or comment groups have been actively detected and penalized since late 2025, and Lempod's Chrome extension was banned in February 2026. The detection method is what makes pods unrecoverable rather than merely risky: the algorithm tracks comment velocity, account relationships, engagement history, and semantic comment content across network graphs. Every one of those is a property of the group, so leaving the pod does not remove the history that flagged you, and every member's activity is evidence about every other member.

On the automation side, LinkedIn's enforcement system uses biometric behavior analysis, including mouse movement patterns and typing cadence, to build a behavioral profile and flag non-human activity. Actions at exact fixed intervals are a clear detection signal. This is also why API-based automation is easily detected while headless-browser tools that simulate human behavior are significantly safer: one produces no behavioral evidence at all, the other produces evidence that can pass.

We can put numbers behind the interval finding because we have logged session data across accounts running real-browser automation on home IPs. Fixed-interval action scheduling, connecting every 60 seconds on a precise clock, triggers behavioral flagging measurably faster than randomized-interval automation running comparable volume. Restriction onset arrives earlier on the clockwork accounts. Editorial sources theorize about this; we have the session logs, and the practical implication is that the timing distribution of your actions is a stronger safety variable than the count.

For outreach, 360Brew evaluates message pattern entropy across a sender's 30-day message history, and it detects identical structural templates even when the surface words vary. This is the finding that invalidates a decade of spintax. Our own sequence testing confirms it from the other direction: accounts using genuinely varied message structures, meaning different opening question types and different value propositions per segment, maintain acceptance rates 2x to 2.5x higher than accounts running spintax-style variation on a fixed template. Entropy is measured at the intent and structure level, not at the word level.

The market-level numbers show what that detection did. Cold connection acceptance fell from 35 to 40% in 2023 down to 18 to 22% in early 2026, while warmed outreach with prior graph-level interactions (profile views, comments, likes before the request) achieves 38 to 45% in 2026. Templated personalization, meaning personalization that varies only at the job-title level, saw reply rates drop 40 to 60% and acceptance rates drop 30 to 50% between Q3 2025 and Q1 2026. Warming is now the difference between the 2023 baseline and half of it.

Volume limits still exist, and they are less important than the quality ratio behind them. The safe daily ceiling sits around 20 connection requests per day and 100 per week, but 23% of users face restrictions even while staying within official volume limits. Acceptance rate is the stronger restriction trigger: 40% acceptance on 20 requests a day is safer than 10% acceptance on 60. Poor acceptance tells LinkedIn's graph model that your targeting does not match your audience, which is a worse signal than sending a lot.

One last finding for anyone already restricted, because it is the question we get most and the least documented anywhere. Operational data from SocialNexis accounts across multiple industries shows that accounts coming off a restriction period need roughly 3 to 4 weeks of below-maximum-cadence posting before out-of-network Suggested distribution returns to pre-restriction levels. That is not a binary switch flipping back on. It suggests 360Brew's topic authority and behavioral trust scores degrade during the restriction and require a graduated re-warm. Coming off a restriction and immediately resuming maximum cadence is the single most common way people extend the damage they just finished serving.

Frequently asked questions

What is LinkedIn's 360Brew algorithm and how is it different from the previous ranking system?

360Brew is a 150-billion-parameter, decoder-only foundation model built on LLaMA 3 and trained on approximately 1 trillion tokens of LinkedIn data. It replaced thousands of separate task-specific recommendation systems with one model solving 30-plus personalization tasks simultaneously. The previous system matched posts to users through isolated rule-based signals; 360Brew evaluates semantic meaning, behavioral history, topic authority, and engagement quality as an integrated input.

How does 360Brew's algorithm rank LinkedIn content and decide feed distribution?

360Brew uses a two-stage pipeline. Stage 1 applies an LLM dual-encoder to narrow millions of posts to roughly 2,000 candidates using semantic embeddings. Stage 2 applies a Generative Recommender that processes 1,000-plus chronological interactions from a viewer's behavioral history to score each candidate. Posts that align with an author's established topic profile and a reader's consumption patterns score highest.

Why has LinkedIn organic reach declined so sharply since 2024?

Median post impressions fell 47%, from 1,211 (June 2024) to 636 (May 2025), based on AuthoredUp analysis of 3 million-plus posts. 360Brew shifted distribution toward personalized relevance rather than broad reach. Posts without topic-authority alignment, with low-quality engagement signals, or in reach-penalized formats (external links, daily posting cadence) now compete much harder for distribution than under the previous system.

What content formats get the most reach on LinkedIn in 2026 under 360Brew?

Document/carousel posts (PDF format) lead with a 7.00% average engagement rate and a 1.39x reach multiplier vs the platform average, per Socialinsider data across 1.3 million posts (2026). Multi-image posts score 6.45%. Native video carries a 0.86x reach multiplier after declining over 70% through mid-2025. Link posts score lowest at a 3.25% engagement rate.

How do saves affect LinkedIn reach compared to likes and comments under 360Brew?

Saves are the highest-value engagement signal under 360Brew. Practitioner estimates put 1 save at 5x-10x a like and 2x a comment in distribution weight (not officially confirmed by LinkedIn). Posts that receive saves and substantive comments 24-72 hours after publishing perform 4-6x better in Suggested feeds than posts that spike and decay within the first hour of publishing.

How long does it take to build topic authority that 360Brew recognizes on a LinkedIn profile?

Approximately 90 days of consistent posting within 2-3 core topic pillars is the established threshold for 360Brew to classify an account's topical domain and deliver full distribution benefit. The author's headline, About section, and recent posting history must all point to the same professional domain. Inconsistent cross-topic posting during this period disrupts the model's ability to place content accurately in relevant feeds.

Does 360Brew detect and penalize AI-generated LinkedIn content?

Yes. 360Brew scores posts for specificity (named examples, exact figures, specific dates), sentence-length variance, and low vocabulary repetition rate as originality signals. Raw, unedited AI output that fails these checks receives reach suppression. Human editing that adds genuine specificity can raise originality scores sufficiently to avoid the penalty. Surface-level edits such as word swaps or sentence reordering without added specifics do not register as meaningful changes.

Does LinkedIn's 360Brew algorithm penalize automation tools and engagement pods?

Both are actively penalized. Engagement pods have been detected and suppressed since late 2025; Lempod's Chrome extension was banned in February 2026. LinkedIn's enforcement system uses biometric behavior analysis including mouse movement patterns and typing cadence. Fixed-interval action scheduling is a clear detection signal. SocialNexis data confirms that restrictions arrive measurably faster on fixed-interval accounts than on randomized-interval accounts.

Do LinkedIn polls and comments count as engagement signals for AI visibility?

Comments count as engagement signals, but quality matters more than volume under 360Brew. Generic phrases like 'Great post!' are explicitly downranked; lexical diversity across a post's comment section is prioritized as a genuine audience signal. Polls generate engagement volume, but the semantic quality of the resulting discussion carries more weight than poll reaction counts alone. Substantive comments from non-clustered accounts are the highest-value signal.

What is dwell time and why does it matter more than likes for LinkedIn reach under 360Brew?

Dwell time is how long a user's scroll pauses on or returns to your post without an explicit engagement action. It is an Activity Signal in LinkedIn's official ranking framework and one of the behavioral inputs the Generative Recommender processes from a viewer's history. A post that holds attention without a like signals genuine interest to the model; a post scrolled past immediately signals low relevance, even if it later collects high reaction counts.

Sources and further reading

• 360Brew arXiv research paper (January 2025)
• LinkedIn Engineering Blog: the LLM dual-encoder and generative recommender architecture
• LinkedIn's official feed ranking signal documentation