What 360brew means for LinkedIn B2B content visibility

LinkedIn swapped the engine under the feed in March 2026. The new one reads your entire posting history before it decides whether a single post deserves distribution. We manage accounts through that transition daily, and the behavioral triggers nobody has published are the ones costing people reach.

The Two-Stage Engine Powering LinkedIn's 360Brew Content Ranking Algorithm

This algorithm change has a name, a size, and a publication date, which is more than LinkedIn usually gives anyone. 360Brew is a 150-billion-parameter decoder-only foundation model built on Meta's LLaMA 3 architecture and fine-tuned on LinkedIn's proprietary data. Hristo Danchev announced the deployment on March 12, 2026, on the LinkedIn Engineering Blog. The design was public long before that. The paper titled 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation went up on arXiv on January 27, 2025, describing the pre-production architecture in enough detail that anyone reading it could see what was coming to the feed a year later.

What 360Brew replaced explains more than what it is. The old feed ran five separate fragmented retrieval pipelines. Each handled its own slice of signals or its own surface, each with its own scoring logic, and each largely blind to the others. That fragmentation is why the pre-2026 feed rewarded tactics: if you knew which pipeline surfaced which content type, you could aim at it. Reach hacks worked because there were five doors and each one had a different lock.

360Brew collapsed all five into a single two-stage system. Stage 1 is a Causal LLM that narrows millions of candidate posts down to roughly 2,000 using cosine similarity, matching post embeddings against a viewer's interest profile, and it completes that narrowing in under 50ms. This is a retrieval step, not a judgment step. It is asking a blunt question: could this post plausibly be relevant to this person at all? Most content that never reaches an audience dies here, silently, without a single impression to diagnose from.

Stage 2 is where distribution is decided. A Generative Recommender takes those roughly 2,000 surviving candidates and reads them against a chronological sequence of the viewer's past interactions, more than 1,000 of them, treated as an ordered sequence rather than a bag of preferences. Order matters. The model is not scoring you against a static profile of what someone likes; it is scoring you against a narrative of what that person has been paying attention to, in what order, and how recently their attention shifted. A post that would have ranked well for a viewer in January can rank poorly for the same viewer in June without anything about the post changing.

The unification runs wider than the feed. 360Brew handles over 30 predictive tasks simultaneously: feed ranking, job recommendations, people-you-may-know, ad targeting, and more, replacing what used to be dozens of separate specialist models with one foundation model inference pipeline. Tim Jurka, VP of Engineering at LinkedIn, has spoken publicly about the architecture and the consolidation behind it. For anyone operating a LinkedIn presence deliberately, this is the most underrated line in the entire story.

Here is why. When the feed ranker and the connections ranker were separate systems, your outreach behavior and your content behavior lived in separate accounting. One model now sits across more than 30 tasks. That does not automatically mean every signal leaks into every decision, and we want to be precise rather than alarmist about this: what it means is that the architectural separation people used to rely on no longer exists by default. The burden of proof has flipped. You can no longer assume your aggressive connection campaign is invisible to whatever decides your post's fate, and you should stop planning as though it is.

The practical read for a B2B operator is that 360Brew is less like a scoring rubric and more like a reviewer with a long memory. It has read everything you have posted. It has read the profile you wrote. It knows what each viewer has engaged with across their last thousand-plus interactions. Tactics that treat a post as a standalone artifact, the hook formula, the line-break rhythm, the comment-bait question at the end, are optimizing an object the model does not evaluate in isolation. That is the whole shift, and every section below is a consequence of it.

Saves Drive 5x More Reach Than Likes Under 360Brew

Save rate is the single strongest engagement signal under 360Brew. A save generates roughly 5x more reach than a like, and 2x more reach than a comment, with a direct message share registering at approximately 4.5x a like. If you have spent three years building a LinkedIn strategy around reactions and comment volume, that hierarchy is close to an inversion of your priorities. The engagement you were most reliably able to manufacture is the engagement worth the least.

The logic is not mysterious once you look at it from the model's side. A like costs a reader nothing and takes a fraction of a second. A save is a reader stating that this content has value beyond the moment they encountered it, that they intend to come back. A DM share is a reader putting their own professional judgment on the line by sending your post to a specific colleague. These are expensive actions. Expensive actions are hard to fake at volume, which is precisely why a ranking model trained to find genuine value weights them heavily. Every signal you can trivially generate is a signal the model has learned to discount.

The second inversion is about timing, and it is the one that quietly kills the old playbook. 360Brew weighs sequential engagement patterns rather than isolated reaction counts. A post that earns steady engagement across the 24-72-hour window after publication performs 4-6x better in ongoing distribution than a post with an identical reaction count concentrated into a burst immediately after publishing. Same numbers. Very different outcome. The model is reading the shape of the engagement curve, not the area under it.

That single mechanic invalidates most of what people still call growth tactics. The launch-window blitz, the notification-bell squad, the coordinated first-fifteen-minutes push, all of it produces a spike-then-flatline curve. Under the old fragmented pipelines that curve read as velocity and earned amplification. Under 360Brew it reads as an anomaly, because organic interest from a real professional audience does not arrive in a single synchronized wave and then stop. A post that people are still finding and saving on day two is describing genuine demand. A post that got everything it will ever get in ninety minutes is describing a distribution list.

There is a third figure that ties the whole section together. Content drawn from direct first-hand professional experience performs 3.5x better than generic or recycled content. That number is usually quoted as if it were a reward for authenticity, which misreads the mechanism. First-hand content does not perform better because the model appreciates sincerity. It performs better because first-hand content is what people save. Nobody bookmarks the fifth restatement of a LinkedIn best-practices list. People bookmark the post that contains a specific number someone measured, a named failure mode, a threshold they will need to remember when they hit the same problem next quarter.

That collapses two things most content teams manage separately. There is no longer a content-quality goal and a save-rate goal that you optimize with different levers. Save rate is the measurement of content quality, expressed in the only currency 360Brew reads. If your post has nothing in it worth returning to, no format trick, no hook rewrite, and no posting-time optimization will produce a save, and the distribution ceiling is set from there.

The operational change is small to describe and hard to do. Before publishing, ask one question: what in this post is worth coming back to? Not what is interesting, not what is true, but what would someone want to have on hand later. A threshold. A sequence. A number you measured. A mistake you made with the specific conditions that produced it. If the answer is nothing, the post will collect likes from people who already follow you and go no further, and under 360Brew that is close to the same as not posting.

Does Posting Frequency Affect Your 360Brew Content Ranking Score?

Yes, and the penalty is sharper than most guidance admits. Posting more than once per 24-hour window consistently suppresses reach for both posts, not just the second one. The mechanism is specific: the newer post enters the algorithm's 60-90-minute visibility window before the earlier post has accumulated enough quality signals to demonstrate its value. Both posts draw from the same per-account distribution budget. Both lose. Whatever you posted this morning gets cut off mid-evaluation by whatever you post this afternoon.

This is the failure mode we see most often in accounts that arrive already struggling, and it usually comes from a good intention. Someone decides to get serious about LinkedIn, batches a week of content on Sunday, and pushes two posts on Tuesday because both were ready. Their reach drops. They read the drop as the algorithm punishing them for something in the content, rewrite their hooks, and post three times the following week. The hole gets deeper. The variable they are tuning is not the variable causing the problem.

The cadence confirmed across multiple independent analyses is 3-4 posts per week, spaced at least 18-24 hours apart. The spacing is the operative half of that. Three posts on Monday, Wednesday, and Friday behave completely differently from three posts on Monday and Tuesday, even though the weekly count is identical. Each post needs a complete evaluation window before the next one enters the ranking queue, and 18-24 hours is the floor for that window, not the target.

The other end of the range has a penalty too, and it is easy to miss because it arrives slowly. Posting less than once per week produces reach decline as well, because the topical authority signal weakens without consistent reinforcement. This is not the algorithm being petty about loyalty. 360Brew maintains a semantic read of what you are credible on, and that read decays when there is nothing recent to reinforce it. An account that posts brilliantly once a month is asking the model to rank content from an author it has mostly forgotten.

Now the part no published marketing guide documents, and the reason we wrote this section rather than pointing at someone else's. From monitoring managed accounts through the 360Brew transition, we found that posting at a perfectly regular interval, with millisecond-level consistency, is itself a suppression trigger independent of content quality. The behavioral layer flags inhuman scheduling precision. A post that lands at exactly 09:00:00.000 every single weekday is not describing a person who writes and publishes; it is describing a cron job, and the model treats it accordingly.

This one is genuinely counterintuitive because it punishes the discipline that every scheduling tool sells. Rule-based schedulers fire at the configured second because that is what a scheduler is for, and their marketing treats that precision as the product. In our data, accounts running human-variance timing injection, randomized offsets of plus or minus 15-30 minutes around the target time, maintain stable reach that purely rule-based schedulers do not. Same content. Same cadence. Different publication timestamps at the sub-minute level. That is the whole difference.

We build a scheduler, so the honest framing is that we shipped this because we watched it happen to accounts we were responsible for, not because randomized timing makes a better feature list. If you are using any tool that publishes on an exact recurring timestamp, that is the first thing to check, and it costs nothing to fix. Set a target window rather than a target time. The reader does not care whether your post lands at 9:00 or 9:22. The model does.

Topic Authority: The 90-Day Credibility Window the Algorithm Measures

360Brew builds a semantic fingerprint from an author's full posting history and profile, then uses it to judge topical credibility before distribution. The question the system asks, in effect, is whether this account has earned the right to talk about this topic. It answers by examining the consistency, depth, and engagement history of everything you have previously posted on the same or adjacent themes. This is not a reputation score attached to your name. It is a topic-by-topic assessment, and you can hold high credibility on one theme and effectively none on another.

That single mechanism explains the reach gap that has confused so many people since early 2026. Two accounts with similar follower counts and similar post quality can see wildly different distribution, and the difference is not the post. One has 200 posts of accumulated signal on a coherent theme. The other has 200 posts spread across leadership, hiring, AI, remote work, and whatever was trending that month. The generalist account is asking 360Brew to distribute content on a topic where it has no evidence the author is credible, every single time it posts.

Establishing that fingerprint takes roughly 90 days of consistent, on-topic posting, and the clock resets. A significant topic pivot resets it. A multi-week posting gap resets it. An algorithmic suppression event resets it. That last one is the cruel case, because the reset compounds the original penalty: you take the suppression, and then you take the rebuild. This is also why the standard advice to broaden your topics when reach drops is close to the worst available move. Broadening dilutes the exact signal that determines whether you get distributed.

The aggregate numbers from the transition are worth sitting with. Richard van der Blom's independent analysis of more than 3 million posts documented a 47% median year-over-year reach decline as 360Brew phased in through mid-2025. That headline figure is real, and it is also misleading on its own, because the underlying distribution was not a uniform haircut. Company pages saw reach fall to 1-2% of their follower count. Personal profiles with strong topical authority saw up to 561% more reach relative to the old system. The median moved down because the losses were broad and the gains were concentrated.

Read correctly, that is not a story about LinkedIn suppressing organic reach. It is a story about a redistribution from accounts with no topical credibility to accounts with a lot of it, and the median hides it entirely. If you have been treating the 47% figure as evidence that LinkedIn organic is finished, you are drawing the wrong conclusion from a number that describes the middle of a bimodal split.

Our own contribution here concerns the recovery side, which is where published guidance stops. Every article names the 60-90-day shadow ban window for accounts flagged for coordinated unnatural engagement, and none of them say what to do inside it. The suppression window is graduated, not flat. From flagged managed accounts we have monitored through the full window, accounts that keep posting topically consistent, save-earning content during the suppression period exit the penalty faster than accounts that go dark or pivot to unrelated topics.

The reason follows directly from how the fingerprint works. The semantic fingerprint keeps accumulating positive signals even while distribution is throttled. A post that reaches a fraction of its usual audience and gets saved by that fraction is still evidence of topical credibility being earned. Going dark stops that accumulation cold and adds a posting gap on top of the flag, which is the reset trigger you least want during a rebuild. The intuitive move when reach collapses is to stop, wait it out, and come back fresh. In our data that is the slowest path out, and pivoting topics during a suppression window is slower still.

What LinkedIn Automation Tools Get Wrong About Suppression Detection

LinkedIn's detection layer does not primarily count your actions. It watches how they are performed. The layer uses behavioral biometrics: mouse movements, click patterns, keystroke dynamics, and browser fingerprinting. Any action delay under 5-10 seconds between related actions is flagged as suspicious. Tools that inject detectable code into LinkedIn's DOM accumulate risk signals over time regardless of how long the tool has been in circulation or how good its community reputation is. Reputation is not a detection input.

That is the gap in essentially every automation safety guide we have read. They are written as volume guides. Stay under this many connection requests, keep comments under that many per day, warm the account for these weeks. Volume caps are the easy half of the problem and the half that tool vendors are comfortable publishing, because respecting a documented limit is a feature you can put on a pricing page. The behavioral half determines whether an account gets flagged at half the published limit, and almost nobody writes about it because writing about it means describing your own architecture honestly.

The pod detection layer is where this becomes concrete for content operators rather than outreach operators. 360Brew's pod detection estimates 97% accuracy in identifying coordinated unnatural engagement, analyzing comment velocity, lexical diversity, and relationship cluster overlap. LinkedIn's VP of Product has confirmed the stated goal is to make engagement pods entirely ineffective. Accounts flagged for coordinated interaction face a 60-90-day shadow ban, where distribution is throttled rather than cut off. Nothing announces itself. Your posts still publish, your notifications still arrive, and your reach quietly sits at a fraction of where it was.

Note what the detection layer analyzes and what it does not. Lexical diversity and relationship cluster overlap are structural properties of a group of interactions. No individual comment in a well-run pod looks wrong. Each one is substantive, unique, and topically relevant. The pattern is only visible across the set, which is exactly why pod operators are so consistently confident their setup is undetectable right up until the reach drops. You cannot inspect your way out of a signal that lives above the level of the individual action.

The architectural question is the one that matters and the one nobody asks. Every tool comparison sorts on features, price, and whether the vendor claims to be safe. The relevant sort is whether the tool produces the signals a genuine human browser session produces at the network and browser level. An API call is not a human session. A headless browser is not a human session. A browser extension injecting into LinkedIn's DOM is a human session with an obvious tell attached to it.

SocialNexis runs actions through a real browser session on a residential IP rather than through an API call or a headless environment, which is why our engagement patterns are indistinguishable from a human user at the network and browser level. That architecture also sidesteps the two specific triggers named above: the under-5-second inter-action delay flag, and the extension DOM-injection pattern. We are describing our own product here and you should weigh that accordingly, but the underlying point holds no matter whose tool you use.

So the question to ask a vendor is not whether they are officially approved by LinkedIn, because no automation tool is. Ask how actions are executed. Ask whether anything is injected into the page. Ask what the minimum delay between related actions is, and whether it is randomized or fixed. A tool that cannot answer those three questions precisely is a tool whose risk profile you do not know. The volume limits in their documentation tell you almost nothing about whether your account survives the next six months.

Personal Profiles, Not Company Pages: Where B2B Distribution Actually Lands

The company page is done as a primary distribution surface for B2B. Company pages saw reach collapse to 1-2% of their follower base under 360Brew, while personal profiles with established topical authority gained up to 561% more reach relative to the pre-360Brew system. This is not a temporary swing or a rollout artifact you can wait out. It follows from what the model is built to evaluate.

360Brew weighs trust signals rooted in individual relationships and connection depth. A personal profile carries those by construction: it has a posting history attributable to one person, a professional record, a connection graph with acceptance rates and interaction depth attached to it, and a semantic fingerprint built from years of first-person material. A company page has followers and posts. It has no individual relationship graph to read, no personal credibility history, and no way to accumulate the signals the model treats as evidence. The model has no mechanism to substitute for what a company page structurally cannot have.

The reallocation this implies is uncomfortable for most marketing organizations, so it usually happens too slowly. If your B2B distribution plan still routes the primary content through the brand page and treats employee posting as amplification, the plan is inverted relative to how the platform now works. The personal profiles are the distribution channel. The company page is a reference surface people visit after they have already found you, which is a real job, just not the one it used to have.

On the outreach side, the static ceiling everyone memorized is gone. LinkedIn replaced the fixed 100-connections-per-week cap with a behavioral Trust Score. High-trust accounts may reach 200 weekly connections. Flagged accounts may be capped at 50. Connection acceptance rates below 20% increase restriction probability, which means the composition of who you invite now affects your capacity to invite anyone at all. Spraying invitations at a poorly targeted list does not just waste the invitations. It lowers the ceiling.

Here is the finding that changes how you should structure the whole operation, and it is one generic LinkedIn safety guides do not capture. From our account monitoring, the Trust Score is surface-specific rather than account-wide. An account facing restrictions on connection request behavior can retain full feed distribution reach. The score governs the surface where the behavior occurred. It does not automatically follow you into the feed.

That has a direct architectural consequence for anyone running both motions. Operators who separate outreach automation from content automation avoid cross-surface contamination of their credibility. If your growth platform runs connection campaigns, comment automation, and content publishing through one integrated session and one behavioral profile, a flag earned by an over-aggressive outreach sequence is happening inside the same operational envelope as your content. Separating them is cheap insurance, and it is the main structural reason purpose-built content tools produce different outcomes than all-in-one growth platforms.

There is a tension worth naming honestly. The unified model across more than 30 predictive tasks argues for more signal sharing between surfaces, and the surface-specific Trust Score behavior we observe argues for less. Both are true as far as we can tell: the ranking model is unified, but the trust and restriction machinery still enforces per-surface. We would not bet a client's account on that separation holding indefinitely, which is a further argument for keeping the two motions structurally apart rather than relying on the platform to keep them apart for you.

Calibrate Your B2B Posting Behavior to the 360Brew LinkedIn Ranking Algorithm

LinkedIn moved against low-quality AI content officially on May 20, 2026, announcing algorithm action against low-quality AI-generated posts and bot-generated comments. Generic AI-generated content receives an estimated 45% less engagement than human-authored posts. 360Brew's semantic analysis identifies the tells: uniform sentence length, templated openers and closers, and an absence of first-hand specificity. Those three are worth reading closely, because they describe structure and substance, not authorship.

This distinction is where most coverage goes wrong, and it matters commercially for anyone running a content operation. The suppression targets content that lacks direct practitioner experience. It does not target content where AI assisted in editing or structuring material that originated from human observation. A post that begins with a number you measured, names a failure you watched happen, and states a threshold you tested is not detectable as slop, because it is not slop, regardless of what helped you tighten the sentences. A post generated from a prompt like write a thought leadership post about AI in sales has nothing in it that a semantic model can distinguish from the ten thousand identical posts it has already ranked.

Comment behavior is the least calibrated area in published guidance, and the most consequential for anyone doing engagement work. The usual definition of a substantive comment is 3 or more sentences or 25 or more words. That threshold is real and also nearly useless on its own, because it measures the wrong dimension. Length tells you whether a comment is worth reading. It tells you nothing about whether the comment helps or harms the account leaving it.

From monitoring managed accounts through the transition, we found the operative variable is relationship cluster diversity, not comment content. Comments arriving from the same small network cluster inside a compressed time window are discounted as a pod signal even when each comment is individually substantive and unique. The same comment volume, spread across diverse second-degree connections over a 6-12-hour window, passes the cluster overlap check. Identical comments, identical volume, opposite outcome. The variable is who is commenting and how tightly they are connected to each other, distributed over how long.

This is the specific reason well-intentioned engagement groups keep getting flagged while their members insist they are doing nothing wrong. They optimized the thing they could see, comment quality, and left untouched the thing the detection layer measures, cluster structure. If your engagement strategy consists of the same eight people supporting each other's posts within an hour of publication, thoughtful comments will not save it, and writing longer ones makes no difference to the signal being evaluated.

Pulling the operating rules together: post 3-4 times per week with at least 18-24 hours between posts, never twice in a 24-hour window; publish inside a target window with randomized offsets rather than at a fixed timestamp; write toward saves rather than likes, since a save carries roughly 5x the weight of a like; hold topical consistency for at least 90 days before expecting stable distribution, and hold it harder rather than broadening if reach drops; avoid tools that operate through API calls or DOM injection; and keep outreach automation structurally separate from content automation.

None of that is a growth hack, and it will not produce a spike next week. It is the boring version, and the boring version is what 360Brew is built to reward, because the model is explicitly designed to discount anything that can be manufactured quickly. The accounts we watch doing well through this transition are not the ones that found a new trick. They are the ones that picked one topic, said something specific about it a few times a week, and did not give the behavioral layer anything to flag. The 561% gainers were not lucky. They were consistent in the one dimension the new model measures.

Frequently asked questions

How does 360Brew decide which LinkedIn posts get distributed beyond an author's immediate network?

360Brew uses a two-stage process. Stage 1 narrows millions of candidate posts to roughly 2,000 using cosine similarity between post embeddings and each viewer's interest profile. Stage 2 analyzes over 1,000 of the viewer's past interactions as a chronological sequence to produce the final ranking. Distribution beyond the immediate network depends primarily on whether early engagement, particularly saves and direct message shares, signals that the content is worth pushing to a wider audience.

What is the exact two-stage ranking process 360Brew uses and how does each stage affect content visibility?

Stage 1 is a Causal LLM that runs cosine similarity matching between post embeddings and user interest profiles, narrowing millions of candidates to roughly 2,000 in under 50ms. Stage 2 is a Generative Recommender that reads those candidates against a sequential record of each viewer's 1,000-plus past interactions to produce a ranked output. A post that clears Stage 1 but earns low saves or comments in early distribution rarely recovers in Stage 2 scoring for subsequent viewers.

Does posting frequency affect 360Brew content ranking and what cadence triggers a reach penalty?

Posting more than once per 24 hours suppresses both posts. The newer post enters the algorithmic visibility window before the earlier post has accumulated sufficient quality signals, draining the distribution budget for both. Stable reach requires 3-4 posts per week with at least 18-24 hours between each. Posting less than once per week also reduces reach because the topical authority signal weakens without consistent reinforcement.

What engagement signals does 360Brew weight most heavily for B2B founder posts in 2026?

Saves are the top signal, generating roughly 5x more reach than a like. Direct message shares register at approximately 4.5x a like. Comments weigh in at around 2x a like. The model also weighs sequential engagement: a post earning steady engagement over 24-72 hours after publication performs 4-6x better in ongoing distribution than one with the same reaction count concentrated in the first two hours. First-hand content earns 3.5x better performance than generic content because it generates more saves.

How does 360Brew detect engagement pods and what happens to accounts that are flagged?

360Brew's detection layer analyzes comment velocity, lexical diversity, and relationship cluster overlap, and estimates 97% accuracy in identifying coordinated unnatural interaction. Accounts flagged for pod behavior face a 60-90-day shadow ban where distribution is throttled rather than cut entirely. LinkedIn's VP of Product stated the goal is to make pods entirely ineffective. Accounts that continue posting topically consistent, save-earning content during the suppression period exit the penalty window faster than those that go dark.

Will LinkedIn automation and scheduling tools trigger suppression under the 360Brew algorithm?

The suppression risk depends on the tool's architecture, not its approval status. LinkedIn's detection layer uses behavioral biometrics including mouse movements, keystroke dynamics, and browser fingerprinting. Action delays under 5-10 seconds between related actions are flagged as suspicious. Tools that operate through API calls or inject code into LinkedIn's DOM accumulate risk over time. Tools that run through genuine browser sessions on residential IP addresses produce engagement patterns indistinguishable from human behavior at the network level.

How long does it take to build topic authority under 360Brew and what resets the clock?

Building a stable topical credibility fingerprint takes roughly 90 days of consistent on-topic posting. The clock resets after a significant topic pivot, a multi-week posting gap, or an algorithmic suppression event. Active posting of topically consistent, save-earning content during a suppression window shortens the recovery period. Stopping posting or pivoting topics during a shadow ban does not accelerate recovery and may extend it.

Why did LinkedIn reach drop 47% in 2025 and is 360Brew the confirmed cause?

Independent researcher Richard van der Blom's analysis of over 3 million posts documented a 47% median year-over-year reach decline as 360Brew phased in through mid-2025. The model's replacement of five retrieval pipelines with a unified semantic system changed how content was matched to audiences and how engagement signals were interpreted. Accounts without topical authority or save-generating content saw the steepest declines. LinkedIn has not published platform-level reach statistics to confirm the figure, but the timing aligns precisely with the 360Brew rollout.

How does 360Brew evaluate AI-generated content differently from human-authored posts?

360Brew's semantic analysis identifies signals associated with low-quality AI generation: uniform sentence length, templated openers and closers, and absence of first-hand specificity. LinkedIn acted officially against AI-generated posts and bot-generated comments on May 20, 2026. Generic AI content receives an estimated 45% less engagement than human-authored posts. The penalty applies to content that lacks direct practitioner experience, not to content where AI assisted while incorporating original human insight.

What is the difference between how 360Brew ranks personal profiles versus company pages?

Personal profiles with established topical authority saw reach gains of up to 561% relative to the pre-360Brew system. Company pages saw reach fall to 1-2% of their follower base. The gap reflects how 360Brew weighs trust signals rooted in individual relationships and connection depth, which personal profiles carry and company pages do not. B2B distribution that relied on company page reach before 2026 needs to be rebuilt around consistently posting personal profiles with defined topic focus.

Sources and further reading

• 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation
• How LinkedIn replaced five feed retrieval systems with one LLM model
• Engineering the next generation of LinkedIn's Feed