LinkedIn Algorithm: Relevance and Expertise over Virality

What the LinkedIn algorithm evaluates

Between 2024 and 2026 LinkedIn rebuilt its feed, moving away from virality as the target metric. Instead of distributing a post to as many people as possible, the system tries to distribute it to the right ones. Daniel Roth, VP and Editor in Chief at LinkedIn, summed it up in these terms: the platform's goal is not to show content to the most people, but to the right ones.

Out of that come three dimensions that specialist analyses such as Hootsuite's LinkedIn algorithm guide describe individually:

Relevance: The system matches your post against the interests of individual members, not against your follower list. A post about Payload migrations reaches developers with an interest in headless CMS even if they do not follow you, and may not reach your own contacts in sales at all.

Expertise: LinkedIn infers from profile and posting history what somebody stands for. Anyone who publishes consistently on one topic is more likely to be treated as an authority. This is an accumulated signal, built over months rather than over a single post.

Engagement quality: What counts is not the volume of reactions but their kind. A considered comment says more about relevance than a like handed out while scrolling past, and according to Hootsuite a save has roughly five times the effect on reach that a like has.

This three-way split is not a weighting model published by LinkedIn. It is the most usable summary of what can be derived from platform statements, engineering papers and third-party analyses. Anyone who treats it as a formula overstates the evidence.

Why reach has fallen, and how solid the numbers are

The most quoted source on the decline in reach is Richard van der Blom's Algorithm InSights Report 2025. His analysis reports views down 50 percent, engagement down 25 percent and follower growth down 59 percent year on year, based on 1.8 million posts, 58,000 profiles and 31,000 company pages over twelve months to February 2025. His update from October 2025 covering around 400,000 profiles measures visibility down 47 percent, engagement down 39 percent and follower growth down 42 percent. A separate Q4 2025 analysis by the SayWhat team puts median impressions at 65 percent down since 2023.

Three caveats that are missing from most LinkedIn posts on the subject. First, none of these numbers come from LinkedIn, they are third-party analyses with their own samples. Second, base year and method differ, which is why minus 47, minus 50 and minus 65 percent are not measuring the same thing. Third, the full reports are paid products, and only the summary is verifiable in each case. Take the direction seriously, not the decimal place.

More revealing than the average is the distribution: according to the same analysis, the visibility share of top creators rose from 15 to 31 percent, while the share of the remaining creators fell from 57 to 28 percent. So reach has become smaller and at the same time more unequally distributed. For companies with a few active profiles that are clearly positioned around a topic, that is an opportunity rather than a problem.

How the LinkedIn algorithm ranks technically

Since 2025 the mechanics have been partly documented in public. The paper "Large Scale Retrieval for the LinkedIn Feed using Causal Language Models" describes a dual encoder fine-tuned on LLaMA 3 that selects around 2,000 posts per request out of hundreds of millions of candidates, within a latency budget of a few milliseconds. Only those 2,000 candidates go into the actual ranking.

Upstream of that sits 360Brew, a decoder model with 150 billion parameters that solves more than 30 prediction tasks previously handled by dedicated models. As of publication in January 2025, the authors describe it as a pre-production model. The paper was later withdrawn from arXiv over a licensing issue, and only the abstract is still retrievable. Third-party claims that the model has been rolled out to 40 to 100 percent of surfaces are explicitly flagged as an estimate, not as confirmation from LinkedIn.

The structural consequence is the same as on the other large platforms: distribution runs through interests, not through the relationship network. How this pipeline of candidate generation, pre-selection and ranking is built in general terms is covered in the overview of recommendation algorithms.

Signals that are actually documented

Dwell time, meaning the time until somebody scrolls on, was described as a feed signal by LinkedIn Engineering back in 2020, including a model that predicts the probability of a skip. The signal is old and stable. What is not documented: the percentage figures circulating for each second bucket, such as 15.6 percent engagement above 61 seconds. Those numbers have no traceable primary source, and LinkedIn does not publish breakdowns of that kind.

On the types of reaction, AuthoredUp's analysis supplies the most concrete values: comments count roughly twice as much as likes across 621,833 analysed posts, and saves and instant reposts are the strongest drivers of reach. On timing it still holds that the first 30 to 60 minutes determine the reach trajectory, while posts with genuine conversation stay visible in the feed for two to three weeks. The second part is the actual news: the long tail is back, which favours well researched content over topicality.

A practical detail from the same source: more than 91 percent of browsing happens on mobile, while 80 percent of profile editing happens on desktop. Anyone who formats their text on a 27-inch monitor rarely checks how the first three lines look on a phone. Those are exactly the lines that decide dwell time.

Formats and form factors

On formats the evidence is thinner than the advice literature suggests, and it shifts with every update. The same AuthoredUp analysis currently reports polls with a reach multiplier of 1.64 and document posts at 1.45 for personal profiles, while on company pages documents lead at 1.40. Multipliers of that order decide nothing if relevance and expertise are not right. They amplify a post that already fits, and they rescue none that misses its audience.

Two form factors are more concrete. For text length the analysis names 800 to 1,000 characters as the sweet spot, which corresponds roughly to an argument with one example, not to a statement and not to an essay. And hashtags help less than their reputation suggests: three to five hashtags tend to lower visibility slightly according to the same analysis rather than raise it. The much quoted 12.6 percent hashtag boost comes from an old, cross-platform study and does not hold up for LinkedIn.

Confirmed, study-based, folklore

The table is the actual task: for every LinkedIn tactic, check which row it belongs in before you point editorial resources at it.

The evergreen test and what it means

In July 2025 many users noticed that posts weeks old were turning up at the top of the feed again. Gyanda Sachdeva, Vice President of Product Management at LinkedIn, confirmed the test and explained that the point was to make valuable insights and career milestones visible again. After user criticism the test was partly reversed according to industry reports.

Treat that as an indication of direction, not as a permanent state. If relevance is the sorting criterion, recency loses weight. Content that still holds up in six months gains ground on commentary about yesterday's news.

What this means for DACH B2B

LinkedIn counts around 24 million registered members in the DACH region (LinkedIn figures via socialmediastatistik.de, as of early 2025). The number should be treated with caution, because the available country breakdowns mix registered members, monthly active users and advertising reach, and do not add up arithmetically. Globally LinkedIn itself cites more than one billion members, 63 million decision makers and 10 million C-level executives; the figure quoted alongside it, that four in five members make business decisions, rests on a study from 2016 and is correspondingly old.

Three things follow operationally.

Narrow positioning: Point profiles at one subject area instead of the full service portfolio. Expertise is an accumulated signal, and an account that serves three topics in parallel gives the system no clear pattern for any of them.

Audience as the reference value: Measure reach per post against the addressable audience, not against the follower count. As a rule of thumb, below roughly two percent organic reach per post despite a consistent expertise focus, a shift towards employee advocacy and paid becomes sensible.

Clarify the roles: Decide early what the company page is supposed to deliver and what the personal profiles deliver. The direct comparison of the two reach profiles is in the article on company page and profile. On the question of whether XING still plays a role alongside it in the DACH region, the comparison of the two B2B platforms is worth reading.

Typical mistakes

Jumping between topics: Anyone writing about recruiting this week, cryptocurrencies next week and leadership after that gives the expertise signal nothing to work with. The system can only amplify what it recognises again.

Leaving the link preview card in place: If you put a link in the post, remove the automatically generated preview card. This is one of the few formatting recommendations that shows up consistently across large data sets.

Engagement pods: Coordinated comments produce reactions without any topical fit. If relevance is the ranking criterion, that produces signals pointing to the wrong audience and degrades distribution over the medium term.

Measuring reach against followers: With interest-based distribution, followers are a stock figure without any guarantee of distribution. Which reference values work instead is covered in the article on engagement rates and benchmarks.

What you should measure

Three metrics are enough to start with. Impressions per post relative to the addressable audience show whether distribution is happening at all. The share of comments, saves and reposts across all reactions shows whether that distribution is going to the right people. And the trend over twelve weeks shows whether the expertise signal is taking hold, because individual outliers say little with this system architecture.

Dwell time itself cannot be read from the outside. As an approximation, compare short and detailed posts from the same account: if the longer ones consistently collect more reactions on the same topic, that argues for time spent as the driver. It is not proof, it stays a correlation inside your own account.

Two things regularly get confused during evaluation. The first is the outlier: a post that slipped into a large interest bubble says nothing about the distribution quality of the rest. The second is the mixing of post level and account level. If reach falls across all profiles of a company at the same time, the cause lies with the feed system rather than with individual posts; if it falls for one profile only, the topical focus is the more obvious explanation.

As of August 2026. The principles, meaning interest-based distribution, pre-selection through retrieval and scoring by reaction quality, are more stable than the individual numbers. The numbers change several times a year, the principle has not changed since 2024.