TL;DR
- What shipped: Profile Relevance is an unsupervised LLM retrieval system that embeds your profile and matches it to posts from people you are not connected to.
- Your profile is the query: headline, summary, skills, work history with role descriptions, and education all get vectorized. Name and gender are excluded by design.
- Engagement is a gate, not a score: a post needs one like, comment, or share from an unconnected member before it is eligible for this distribution.
- Summarizing costs specificity: their rewrite step quietly dropped geographic context, so they patched it by appending country names to the prompt.
- The cold-start opening: members with little history saw a 17.7% lift in viral engagement. A precise profile now competes with network size.
What LinkedIn Actually Shipped
LinkedIn’s engineering team published a post called Profile Relevance. Most people will read it as a product announcement, nod at the words “LLM” and “embeddings,” and move on. Read it as a specification instead. It is the most detailed public description LinkedIn has given of what the feed reads, what it throws away, and what it refuses to distribute.
The system in one sentence: Profile Relevance is an unsupervised retrieval system that puts member profiles and posts into the same embedding space, then uses that shared space to surface Unconnected Content, which is LinkedIn’s term for posts from people you are not connected to.
Network proximity and engagement signals already decide most of the feed. They are good at relevance for people you follow and bad at discovery for people you do not. A creator building an audience from zero had no path in unless someone with reach shared them. Profile Relevance opens that path with professional context instead of the social graph.
Your Profile Is the Query
The core artifact of the system is the member prompt. It is a structured representation of the professional information on your public profile, assembled to give a language model enough context to answer one question: what does this person actually do? It is built from three blocks.
- Profile highlights: headline, summary, and key skills
- Work history: companies, titles, durations, and descriptions of past and current roles
- Education: schools, fields of study, and degrees
LinkedIn states plainly that the prompt excludes personal attributes like name and gender. What it does include is the part most people treat as decoration: the description field under each role. Durations too. If you have listed four titles with no descriptions, the model has four nouns and no context about what you did with them. The embedding is thin because the input is thin. That is not a metaphor. It is arithmetic on the input length.
Posts get the same treatment from the other side. Their embeddings are built from post text, titles, and transcripts, plus author metadata such as location when it is available. Which means a video with a transcript is not the same object as a video without one. LinkedIn can read the audio. If you publish a talking-head clip with no captions and no transcript, you are shipping a mostly empty document into a semantic index and hoping the thumbnail carries it.
Your profile is not a storefront anymore. It is the retrieval record. The feed is not ranking you by what you posted last week. It is retrieving you by what your profile says you are.
The Compression Tradeoff Nobody Budgets For
The most instructive part of the engineering post is not the result. It is the architecture decision they had to make twice. They evaluated two ways to turn raw text into something a retrieval model can compare.
| Approach | How it works | What it costs |
|---|---|---|
| Rewrite | A large generative model extracts the top topics from a profile or post, then a smaller embedding model turns those topics into vectors | Cheap to serve, but loses nuance and location context |
| Non-Rewrite | A retrieval-tuned embedding model with extended context reads the raw text directly | Preserves detail, requires far more compute |
| Shipped hybrid | The fine-tuned generative model pulls topics, a fine-tuned small embedding model handles retrieval across profiles and posts | Balanced cost and fidelity, with the losses that come from compressing |
They shipped with rewrite, then their own evaluation caught the bill. The rewrite step was attenuating geographic cues. Subtle location signals in a profile or a post were being dropped on the way from raw text to a short topic list, and the system started recommending content that was topically correct and geographically wrong. The fix was not a new model. It was appending the country name from the member’s physical location directly onto the prompt.
Two models do the work in the shipped version. A fine-tuned large open-source generative model pulls the top 10 topics from a profile. A fine-tuned small open-source embedding model runs the Entity-Based Retrieval pipeline that puts profiles and posts in one shared space. Two models, two jobs, one bill.
Members edit profiles constantly, so LinkedIn does not re-embed everyone. It diffs each generated prompt against a stored copy and refreshes only the members whose text changed, which puts the job at 24% of the compute a full refresh would cost. They scaled out instead of up, because bigger pods left GPUs idle waiting on data loading. Inference runs against the ETL data SLA, with a daily selective backfill for records that arrive late.
That is the lesson, and it applies to every content system I have built or audited. The moment you summarize before you embed, you are making a decision about what to lose. You usually make it without knowing. The summary looks clean, the vector space looks healthy, and three months later you discover you dropped the one attribute your audience actually cared about. LinkedIn found out because they had an evaluation system running daily. Most teams find out when a client asks why the recommendations feel off.
Engagement Is a Gate, Not a Score
The second half of the production system is a filter that most coverage will skip past. Profile Relevance does not serve a post just because the embedding is close. The post has to have received at least one viral action, meaning a like, a comment, or a share, from a member who is not connected to the author.
Read that carefully. It is one interaction, and it has to come from outside the author’s network. When I read the threshold, the phrase I could not get past was from an unconnected member. That one condition is the whole design.
Two things follow. First, your network is the ignition, not the audience. A post still has to earn its first outside reaction before LinkedIn will consider handing it to strangers, so the first hour or two still decides whether the post is eligible at all. Second, and this is the part most people will get wrong: a small group of connected peers upvoting each other does not clear this gate. The reaction has to come from someone unconnected. Engagement pods were already weak. For this distribution path they are structurally useless, because every member of a pod is connected to you by definition.
The gate also explains the quality numbers. LinkedIn reported a 0.54% increase in Feed Updates Viewed, a 0.94% increase in viral rate, and an 18.8% reduction in hide rate after adding it. Those look like rounding error until you remember the denominator. These are small percentages on a feed serving hundreds of millions of people a day. At that scale, a percentage point of viral rate is a category of content that exists or does not.
How They Grade a System With No Labels
There is a third decision in the post that matters more to content teams than either model choice. Profile Relevance is unsupervised, which means there are no click or conversion labels to train against and no clean ground truth to evaluate against. Early on, evaluation meant hand-reading roughly 20 member-post pairs per model variant. That is fine for a prototype and useless at the pace they needed to iterate.
Their answer was AutoEval, an LLM-driven evaluation system that scores each post against a member’s professional context across named factors: topic, location, job title, company, industry, seniority, and education. It replaced 20 human-labeled records with 1,000 per model, run automatically every day across every prompt and model variant.
The benchmark was their previous retrieval source, MemberEBR. Against it, Profile Relevance lifted on-topic rate by 34%, cut irrelevance by 33%, cut generic content by 24%, and lifted the share of members whose top ten results held at least five on-topic posts by 39%. That last number is the one I would put on a slide.
They defined three metrics worth stealing:
- Irrelevance rate: the share of posts that are irrelevant on every attribute
- Generic content: the share of posts that miss on all attribute relevance, meaning nothing about them is specifically for anyone
- On-topic member share: the share of members with at least five topic-relevant posts in their top ten results
The third metric is the interesting one. It is not a per-post score. It is a share-of-members measure, which means the system can look acceptable on average while a large minority receives nothing relevant at all. That is the exact failure mode every recommendation system hides, and almost nobody reports it because the average always looks fine.
Then they closed the loop properly. They compared AutoEval’s judgments against human reviewers across multiple relevance dimensions, found where the two disagreed, and rewrote the evaluation instructions until they aligned. That is the step most teams skip. An LLM judge is only as good as the evidence that it agrees with a human. This is layer four of the content engineering stack, and LinkedIn just published a working implementation of it.
The Cold-Start Opening
The results with the most strategic weight are buried near the end of the post. For members with limited activity history, LinkedIn measured a 17.7% increase in viral engagement and a 1.8% lift in daily active use.
Network-based ranking favors incumbents because it runs on history. A new account has none. Retrieval built on profile text does not need history, it needs a well specified profile. That is the first time LinkedIn has published evidence a new account can compete on something other than tenure.
Our own data from the 2026 LinkedIn Algorithm Report that I published with SignalScout points the same direction. Engagement per 1,000 followers drops from 9.7 for accounts in the 2K to 5K range to 0.8 above 50K, a 12x collapse in follower efficiency. I used to read that as audience fatigue. If retrieval replaces graph proximity for unconnected distribution, it stops being a quirk and becomes structural. Audience size stops being the asset. Audience specificity does. You can see the same pattern in the signals that actually predict reach.
What I Actually Think
I spent three years inside LinkedIn. The lesson that stuck with me is that the profile was never a storefront. It was the record every other system keyed off. The Social Selling Index pointed at it. Sales Navigator pointed at it. Every ranking system I touched treated the profile as the entity and everything else as an event. That is why I have been telling people for years that the profile is the homepage, and why most profile advice is too shallow to matter.
This post confirms LinkedIn has stopped treating the profile as a setting and started treating it as a signal. That is a bigger change than the announcement makes it sound, and it inverts the advice most creators are getting right now.
LinkedIn did not build a better feed. It built a better index of you.
If your profile is thin, posting more does not fix it. You are just publishing into a retrieval system that cannot tell what you are for.
The default reaction to any algorithm news is to post more. That is the wrong lever here. Post volume does not change your embedding. Your headline, your skills, your role descriptions, and your summary do. The people who benefit from this system are the ones who write a profile like an engineer writes a schema: specific, complete, and free of adjectives that carry no information.
My prediction: within a year, profile work becomes an engineering discipline with its own audit checklist, not a resume exercise you do once a decade. And the people who do it now will hold an advantage that compounds, because every post they publish lands in a better retrieval neighborhood. That is a flywheel, and flywheels do not care how many followers you have.
The Playbook: Five Changes That Move Your Embedding
Not “led marketing.” Name the function, the system you built, and the outcome, using the nouns a buyer would use. Include the stack. Include the number. Descriptions are an explicit input to the prompt, so a blank description is a blank field in the record.
LinkedIn appends country automatically. City, region, and industry do not come free. If you sell into a specific vertical or a specific geography, say it in the headline or the summary. Their whole geo patch exists because this context was missing.
Two sentences on who you help and what you do about it. Skip the biography and the mission statement. The model treats the summary as context, and vague context produces vague neighbors in the retrieval space.
Skills are an explicit input, not decoration. Fifteen real skills beat three generic ones. This is the cheapest lever on the page and most people leave it set to whatever LinkedIn guessed in 2019.
You are not optimizing for impressions. You are optimizing for one reaction from someone outside your network, because that is the gate. That usually means publishing when your unconnected audience is awake, not when your peers are.
LinkedIn moved the lever from your posting history to your profile record. The system reads what you say you are, gates distribution on one outside reaction, and hands the advantage to whoever is most legible. Legibility is now a growth channel.
None of this makes posting irrelevant. It makes posting the second half of the job. The first half is being findable by a retrieval system that has never met you. Write the profile for the machine, post for the humans, and stop treating the description field as optional.
