LinkedIn’s 2026 feed rebuild replaces separate ranking systems with a unified LLM-based retrieval and a sequential ranking model that reads your history and every scroll. LinkedIn’s engineering lead described the change in the Engineering Blog post “Engineering the next generation of LinkedIn’s Feed” on 12 March 2026, naming signals like reads, likes, comments, returned visits, scrolled-past posts and long dwells. “Depth Score” is not a term LinkedIn uses; the phrase functions as sector shorthand for the depth-oriented reading of those signals. If your reach dropped and quick reactions no longer travel far, the fix is not a new hack; it is a posting practice built around the signals LinkedIn already documents in public.
Operating rule
Do not chase reach. Chase the read. Every post-shape decision should be grounded in a signal LinkedIn’s Engineering Blog names, not in a signal you invented on the way to a template.
What LinkedIn’s engineering team actually says changed
The LinkedIn Engineering Blog post from 12 March 2026 describes replacing “multiple retrieval sources — trending content, collaborative filtering, and embedding-based systems” with a “hybrid complementary relevance solution — a unified retrieval system leveraging advances in LLMs”. The post also names a sequential ranking model with a transformer architecture that “captures how professionals engage with content over time” and “processes more than a thousand of your historical interactions”. The engagement signals it names — read, liked, commented on, returned back to, scrolled past, long dwells, likes, comments, shares — are the material a post must earn to travel.
Two clarifications matter before we go further. First, LinkedIn’s Engineering post does not name the frequently-cited “360Brew” system as the production feed model; the LiRank paper on arXiv from LinkedIn’s engineering team is the closest peer-reviewed description of the ranking framework and covers Residual DCN, Dense Gating and Transformer components. Second, no LinkedIn document I could verify discloses per-signal weights. Anyone quoting an exact percentage on a signal — “dwell is 32% of the score” — is inventing a number. Treat that pattern as a red flag, not evidence.
Visual slot: Add an “old-signal vs new-signal” contrast diagram here — separate specialised ranking models on the left, unified LLM retrieval + sequential ranking on the right, engagement-signal legend beneath both. Replace this block with the final visual later.
| What LinkedIn documents | Where it appears | Practical implication | Do not extend this into |
|---|---|---|---|
| Unified LLM retrieval + sequential ranking | Engineering Blog, 2026-03-12 | One system now reads a member’s history end-to-end | A claim that “360Brew” runs your feed |
| Transformer over 1000+ historical interactions | Same Engineering post | Recency is not the only signal — sequence matters | A specific per-post weight |
| Long-dwell classifier | Engineering Blog, 2020-05-12 | Passive reading is a positive signal | Disclosed dwell-time threshold in seconds |
| Residual DCN + Dense Gating + Transformers | LiRank paper, arXiv 2402.06859 | The ranking layer is a stack, not a single feature | Signal weights the paper does not publish |
| Read, liked, commented, returned, scrolled past, shares | Engineering Blog, 2026-03-12 | Six behaviours drive distribution, only one is “like” | The five other signals are equal to a like |
Where “Depth Score” came from and what it does not mean
Naming rule
“Depth Score” is sector shorthand. Treat it like a nickname, not like a product name.
The phrase “Depth Score” is not present in any LinkedIn document I could verify. It circulates in industry write-ups as a name for the depth-oriented reading of the signals LinkedIn does describe. That is a useful teaching label — as long as you remember it is not LinkedIn’s own name. Because the term is a compression of “distribution follows depth, not reactions”, it is fine as a heuristic for your own team, but do not quote it inside a claim about how LinkedIn officially describes its system.
The practical translation is short. LinkedIn’s Engineering Blog names a set of signals. The industry compresses that set into a nickname. Your team can use the nickname; your team must not attribute it to LinkedIn. A footer that reads “as described by LinkedIn’s engineering team” belongs above a signal list drawn from the actual Engineering Blog post — not above a paragraph that names “the Depth Score model”.
The signals LinkedIn documents (and the ones it does not)
Dwell time and long-dwell — the underpaid signal
LinkedIn’s own Engineering Blog post on feed dwell time, authored by Siddharth Dangi, Johnson Jia, Manas Somaiya and Ying Xuan on 12 May 2020, defines two measures. On-feed dwell time “starts measuring when at least half of a feed update is visible”. Post-click dwell time is “the time spent on content after clicking on an update”. The article also explores a long-dwell binary classifier that predicts whether dwell time on a post exceeds a context-dependent threshold — but explicitly does not disclose the numerical value. Passive scrolling counts. A member who lingers on a post they never click, never react to and never comment on has still voted.
The takeaway for a post-writer is unusual because it is dull. If your only measurement window is impressions, you are missing the signal that actually tells LinkedIn something. A post that earns two seconds of scroll and zero reactions is a different object from a post that earns twelve seconds of scroll and zero reactions — even though both look identical in your dashboard.
Comments and shares — depth over volume
LinkedIn’s 2026 Engineering post lists comments and shares among the signals its sequential ranker reads. The LiRank paper on arXiv (Borisyuk et al., revised August 2024) treats comment- and share-style signals as multi-task prediction targets alongside reading behaviour. Neither source publishes a weight. Both make clear that the count of comments is not the object of interest on its own — the reading, the return visits and the follow-through are the object of interest. That is why an engagement-bait comment thread does not travel the way a substantive one does.
Visual slot: Add a signal-to-post-shape worksheet here — signal, what generates it, what post shape earns it, what to measure. Replace this block with the final visual later.
| Signal | What generates it | Post-shape that earns it | What LinkedIn does not publish |
|---|---|---|---|
| Long dwell | A reader stays past the fold without clicking | Story with a real setup, not a listicle stack | Exact threshold in seconds |
| Reads returned to | A reader comes back later in the session | Reference posts, tables, checklists worth revisiting | How re-entry weight compounds |
| Substantive comment | A reader writes more than a reaction | Open question with a bounded scope | Whether words per comment enter the score |
| Direct share / DM | A reader forwards the post | Post about a decision a peer would replay | Whether DM shares outrank public reshares |
| Scroll past | A reader skips within the first fold | Fixable only by rewriting hook + first line | The skip threshold value |
Formats that earn reading time
Format rule
Choose the format the signal you need is easiest to earn in. Every other reason for a format is decoration.
Video is a specific case. LinkedIn Marketing Solutions’ best-practices page for video ads states that “videos under 30 seconds reported a 200% lift in view completion rates (Source: LinkedIn internal study, 2018)”. That statement is bounded — it is about video ads and about completion rate — but it points to the same posture the Engineering Blog described in 2026: shorter, more attentive units earn the pieces of a member’s session that ranking rewards. Long-form video can still work when the first ten seconds justify staying; the completion window is not going to widen because you filmed twelve minutes of intro.
Text-only posts are not obsolete. They are, however, competing directly with the “read” signal — every extra sentence has to earn its way past a scroll-past. Documents, carousels and native articles can help long-form content survive because they invite return visits. If your goal for a piece is a re-entry rather than a first-glance react, the format decision is different from a short punchy hook.
| Format | Signal it best earns | Documented boundary | Not a promise |
|---|---|---|---|
| Short video (< 30 s) | Completion, direct engagement | LinkedIn video-ads best-practices page, 2018 internal study | Organic reach lift |
| Long-form text | Long dwell, thoughtful comments | Engineering Blog, 2026-03-12: reads and long dwells named | A specific completion rate |
| Document / carousel | Return visits, saves | Ad-format Help page lists document ads as a distinct format | Return count guarantees ranking |
| Native article / newsletter | DM shares, subscription | Ad-format Help page lists article/newsletter ads | Newsletter subscribes as feed weight |
| Text with an image | Scroll retention past the fold | Feed engineering post: reads and scroll-past both named | Reach uplift by adding an image |
Where paid distribution intersects with organic depth
Two ad-side surfaces are worth knowing when you are thinking about depth signals. LinkedIn Marketing Solutions Help’s page on thought leader ads defines them as ads that “promote content from LinkedIn members, which can increase your reach and help you meet your brand goals” — with the creator’s explicit permission and specific role permissions on the company Page. Celebrations, documents, polls, multi-images and reshared posts are ineligible. This is not a shortcut past ranking; it is a paid distribution surface for a member post, and the underlying organic depth signals still matter for what that post’s public reception looks like.
The Help page for ad formats and objectives lists 13 objective-based formats. Brand Awareness supports ten of them (single image, carousel, video, event, text, spotlight, follower, document, article/newsletter, and thought leader ads); Video Views supports video ads only. That means if you decide to reinforce an organic depth-post with paid amplification, the objective you choose narrows the format list quickly. Choose the pairing on purpose. A brand-awareness thought leader ad promoting a member’s video is a different piece of media from a video-views campaign on the same asset — and neither is a substitute for the piece earning long dwell in feed.
Visual slot: Add a “what LinkedIn Engineering documents / what sector shorthand adds” comparison board here — column A: signals from Engineering Blog; column B: sector nicknames; column C: safe usage rule per row. Replace this block with the final visual later.
A weekly posting system built around documented signals
System rule
One post-shape per week, one signal per shape, one measurement per shape. Anything more is a hobby, not a system.
The point of a weekly system is not to fill the calendar. It is to make the signal you are testing explicit before you write. If you already know the shape and the signal you are asking that shape to earn, you can decide the format, hook, length and follow-up without renegotiating your tone every day.
Use the following cadence as a starting point, not a template you must copy. The purpose is to keep the signals varied enough that you learn which shape earns which behaviour in your specific account.
Visual slot: Add a weekly cadence board here — day, post-shape, signal to earn, target measurement, notes column. Replace this block with the final visual later.
| Slot | Post-shape | Signal earned | Measurement | Do not track |
|---|---|---|---|---|
| Monday | Long-form text with an operating rule | Long dwell + comments | Average reading time + substantive comments | Emoji reactions |
| Wednesday | Short video (under 30 seconds) with one visible decision | Video completion | Watch-through percentage on the ad-level metric | Public views count in isolation |
| Friday | Document or carousel worth returning to | Saves + return visits | Save count + late-week impressions | Impressions in first hour |
| Weekend | Reflection post or short essay | DM shares + reader replies | DM shares + first-response quality | Follower delta over 24 hours |
Detailed operational example
Assume a solo B2B consultant with an account that used to average two hundred reactions per post and now averages sixty. Rather than looking for an algorithm hack, the consultant runs a four-week experiment. In week one, they publish a long-form text piece and a short video only. In week two, they add a carousel with a follow-up post that references it. In week three, they publish a reflection essay grounded in a specific client scenario, clearly labelled as a scenario rather than a case study. In week four, they compare the four post-shapes on three axes: average reading time when analytics discloses it, thoughtful-comment count, and direct-message shares.
The consultant does not conclude that the video works better than the essay for their audience because the video collected more reactions. Reactions are volume, not depth. They conclude which shapes their specific audience actually read to the end, which shapes their audience quoted in the comments, and which shapes they wrote to peers about. If the video earned the reactions and the essay earned the DMs, the essay was the more distribution-worthy post for that reader base — and the following week starts from that finding, not from a template.
Measurement beyond impressions
LinkedIn’s built-in analytics does not disclose every signal the Engineering Blog names. It does not, for example, tell you the numerical dwell-time bucket a post earned. That is not a reason to ignore the ones it does surface. Read time (when shown), impressions vs unique members reached, meaningful comments, saves, and — where you can observe them — direct-message shares are all useful proxies. The important discipline is not the tool; it is the habit of comparing shapes and weeks against each other, not against an absolute number.
Visual slot: Add a measurement dashboard sketch here — post-shape, impressions, unique reach, comments (thoughtful vs reaction-style), saves, DMs, editorial note column. Replace this block with the final visual later.
Risks: over-indexing on one signal, faking depth, ignoring the LLM shift
Risk rule
Any tactic that would embarrass you in a client meeting is a distribution risk. LinkedIn’s Engineering Blog already reads for meaningful conversation; the tactic reads as unmeaningful because it is.
The risks below are not novel to Business Agent for Leads-style formats. They apply to every posting strategy that reads a public feed engineering description as a to-do list.
- Over-indexing on one signal: Building every post to farm dwell time, or every post to farm comments, produces uniform content that eventually stops earning either. The Engineering Blog names several signals for a reason.
- Faking depth: Asking a question you already answered, or padding a short thought into a long post, is legible to readers and — via reader behaviour — legible to a model reading long dwells and returns.
- Ignoring the LLM shift: The 2026 change explicitly reads a member’s history over 1,000+ interactions. Posting a single strong piece into an account that otherwise posts generic filler will not reset the sequence view of your account.
- Generic AI-drafted posts: LinkedIn has not published a rule that penalises AI content per se. The reasonable reading is that generic phrasing — regardless of author — earns less dwell and fewer substantive comments and therefore travels less. Do not assume the origin is the failure; the phrasing is the failure.
- Chasing “Depth Score” as an official term: Naming it as a LinkedIn product in your posts undermines the piece the moment a reader searches the phrase inside LinkedIn’s own materials and finds no match.
Common mistakes with “algorithm hacks”
These are the recurring failure modes when a team responds to a feed change by looking for a new hack instead of a new discipline. None of them are unique to LinkedIn; they will apply to the next feed rebuild on the next platform.
- Quoting an exact percentage for a signal weight (“dwell is 32% of the score”) that no LinkedIn document publishes.
- Attributing “Depth Score” to LinkedIn as if it were a product name.
- Assuming “360Brew” runs the live feed because a widely-shared arXiv paper described it — even though the paper was withdrawn.
- Using engagement bait (“Comment YES if you agree”) and reading the resulting comment count as depth.
- Publishing generic AI-drafted posts and then blaming “the algorithm” when they underperform.
- Chasing follower count as if it were the ranking signal.
- Buying a Thought Leader Ads campaign to bypass depth-signals instead of to reinforce a post that already earns them.
Set one post-shape per week and measure the retention curve
The disciplined move for a team that has just watched its LinkedIn reach drop is to stop looking for a new “algorithm” article and pick one post-shape per week — long-form text, short video, document or reflection essay — and commit to writing it well enough to earn a documented signal. Long dwell, thoughtful comment, save or DM share. If you can hold that discipline for four weeks and compare shapes against each other rather than against a hypothetical target, you will have earned a working sense of what your specific audience is willing to read. That is a more durable output than any percentage-weight claim you could quote.
Write down the signal you were trying to earn before you write the post, and read the post one more time asking whether it actually earns that signal. If the post-shape you chose is not doing the job, replace the shape, not the whole system.
A pilot week is easier to defend to yourself when its target is a signal you already know LinkedIn reads. That is why grounding the plan in the Engineering Blog’s language, rather than in a trending nickname, is not pedantry — it is what keeps the week from degrading into content improvisation the moment a reader disagrees in the comments.
Visual slot: Add a post-shape decision worksheet here — signal target, chosen shape, hook plan, follow-up plan, review date. Replace this block with the final visual later.
Want a content system built around documented feed signals?
A content-system review reads your posting mix, the signal each shape earns and the way your account presents to LinkedIn’s sequential ranker — and returns a weekly plan grounded in the signals LinkedIn actually documents. No follower promises, no reach targets, no viral hooks.
Sources and evidence
- LinkedIn Engineering Blog — “Engineering the next generation of LinkedIn’s Feed” by Hristo Danchev (12 March 2026)
- LinkedIn Engineering Blog — “Understanding dwell time to improve LinkedIn feed ranking” by Dangi, Jia, Somaiya, Xuan (12 May 2020)
- Borisyuk et al., “LiRank: Industrial Large Scale Ranking Models at LinkedIn”, arXiv:2402.06859 (submitted 2024-02-10, revised 2024-08-07)
- LinkedIn Marketing Solutions Help — Thought leader ads
- LinkedIn Marketing Solutions Help — Ad formats for objective-based advertising
- LinkedIn Marketing Solutions — Best practices for video ads
FAQ
Is “Depth Score” really LinkedIn’s official term?
No. “Depth Score” is not present in LinkedIn’s Engineering Blog or in the Marketing Solutions Help pages I could verify. It functions as sector shorthand for the depth-oriented reading of the signals LinkedIn does describe — long dwells, meaningful comments, returns, saves and DM shares. The label is fine as a team heuristic; it is not something to attribute to LinkedIn in your own posts.
Do likes still matter on LinkedIn in 2026?
LinkedIn’s Engineering Blog names likes as one of the signals its 2026 ranking model reads, alongside reads, comments, shares, returned visits, scrolled-past posts and long dwells. Likes still count; they are one signal in a broader set, and quick reactions alone do not carry a post the way a substantive comment thread and long dwell can.
Does LinkedIn penalise AI-written posts?
LinkedIn has not published a public rule that penalises AI-written posts as such. The reasonable reading of the 2026 Engineering post is that generic phrasing earns less dwell and fewer substantive comments and therefore travels less — regardless of the drafting tool. The origin of the post is not the failure; the phrasing is.
Is 360Brew running LinkedIn’s feed?
The 360Brew paper on arXiv (2501.16450) was withdrawn by arXiv administrators on 2025-08-23. LinkedIn’s own 2026 Engineering Blog post about the feed rebuild does not name 360Brew. Third-party guides frequently claim that 360Brew now runs LinkedIn’s feed; that claim is not supported by any LinkedIn source I could verify.
What post length earns the most reading time?
LinkedIn does not publish a length-to-reading-time relationship. The Marketing Solutions video-ads page notes that “videos under 30 seconds reported a 200% lift in view completion rates” from a 2018 LinkedIn internal study, but that is a video-completion claim, not an organic long-form claim. For text posts, the honest guidance is to write the length the argument actually needs and to test which length earns long dwell in your specific account.
How often should I post now?
There is no LinkedIn-disclosed posting frequency. The 2026 Engineering post makes clear that the ranker reads a member’s history over 1,000+ interactions, so a shape you can hold for months matters more than a temporary daily sprint. A weekly system with one deliberate shape per slot is easier to hold than a daily posting habit that trends towards generic filler.
Rebuild your posting practice around documented signals.
A content-system consulting engagement reviews four weeks of your posts against the signals LinkedIn documents, replaces the “reach hacks” with a weekly shape-and-signal plan, and returns a written system — not a promise about a follower count.