Methodology.
How What Works on LinkedIn measures the public posts of 346 prominent LinkedIn writers: what was collected, how engagement is indexed, the four lenses that separate craft from account size, how every band is defined, how the by-eye reads were done, and where the study stops. No author is named and no post is quoted anywhere in the study.
Contents
- Publisher
- NBK Social Research
- Research
- Matt Cunnelly, Josh Stoddard
- Study
- 002
- Profiles
- 346
- Posts collected
- 74,119
- Learning set
- 39,858 posts
- Authors in set
- 187
- Posts dated
- 2014-06-04 to 2026-09-10
- Collected
- 10 September 2026
- Published
- 9 October 2026
Contents
Abstract
What Works on LinkedIn is a study of the public engagement of 74,119 LinkedIn posts written by 346 prominent writers, collected on 10 September 2026 and dated from 2014 to 2026-09-10. Posts are grouped into bands by how they are written (length, line rhythm, the first line, the opener, the close, numbers, lists, voice), how they are formatted (format, image shape and count, links, hashtags, mentions, emoji), when they are posted (hour, day, the gap since the author’s last post, the author’s cadence) and what they are about (archetype and subject), and each band’s median engagement is indexed against the median of a learning set of 39,858 posts from the 187 authors whose own median is 100 or more. Each key band is then read on three further lenses, per 1,000 followers, on the comments share and within the author, to separate what lifts a post from what merely marks a large account. The 94 strongest image posts and 41 of the strongest authors’ banners were read by eye to a written protocol, and 26 profile signals were compared between the strongest 44 authors and all 344 profiles. The study publishes aggregate tables only.
Governance
No author is named and no post is quoted. The study is about the shapes of posts, never the people who wrote them. The tables are exported from NBK’s stored analysis by a script that refuses to write a file carrying an author, a handle, a post, a first line, a follower count or a cost, and the page renders only what that file holds. A reader cannot reconstruct any individual from this study, and that is by design.
Every index is correlational, and the page says so beside the figure. The lens that removes account size is shown wherever it was computed, and where the raw index and the within-author index disagree the study reports the within-author reading as the one to trust. Section 5 of the study is the worked case: a habit that indexed 1.19 across authors and 1.70 on comments reads 1.02 to 1.04 within the author, and the study changed its own working rules on that finding before it was published.
Nothing about the findings is, or ever will be, for sale. No commercial relationship with NBK Social, past, present or prospective, moved an author into or out of the population or a figure in any table.
Population and selection
The population is 346 public LinkedIn profiles selected by NBK Social in September 2026 as prominent writers on the platform, in two groups: a craft cohort chosen for how they write (authors widely read for the writing itself, whatever its subject) and a domain cohort chosen for working in social media and marketing. NBK’s own founders are in the collection as a benchmark and are excluded from every learning-set figure.
The learning set is the subset of authors whose median post earns 100 or more engagements: 115 craft authors with 24,649 posts (median 345) and 72 domain authors with 15,209 posts (median 288), 187 authors and 39,858 posts in all. Every band index is computed over this set; the two cohorts are also reported separately so a reader can see whether a finding holds in both (study section 17). The threshold is a selection and is stated as one: the study describes what distinguishes posts among accounts that already reach people.
There is no sampling within an author: every public post on the profile at collection is in the set, the oldest from 2014-06-04. Reposts and quote posts are kept and banded as such, because their engagement is part of the finding.
Data collected
Collection ran on 10 September 2026 from public profile and post pages. For each post: the text, the reaction count, the comment count, the post type (original, repost, quote), the format the platform labels it with, the media’s kind and pixel dimensions, the number of images, and the timestamp. For each profile: the headline, the About text, the experience entries, the follower and connection counts, the verified, Premium and Top Voice badges, creator mode as the profile recorded it, the creator hashtags, the website link, the education and certification entries, the top skills shown, and the banner image. In all, 74,119 posts carrying 21,171,919 reactions and 4,405,935 comments, 36,046 of the learning set’s posts with media, and 323 profiles with a banner. All data is public; no API permissions, no private analytics, no impressions.
The corpus is frozen. The analysis is stored once and never recomputed on a page; the study page renders the stored tables, so a figure on the site cannot drift from the figure the analysis produced.
Engagement and the index
Engagement
engagement = reactions + comments
The two public counts LinkedIn shows on a post, read as displayed at collection. Reposts of a post are not added to its engagement. Impressions are private to the author and are not measured.
The learning set
author median ≥ 100 engagements, author not NBK
Authors whose median post earns 100 or more engagements, excluding NBK's own founders: 187 authors and 39,858 posts. Every band index is computed over this set. The threshold keeps the study about writers whose posts reliably reach people, which is the question asked, and it is stated so the selection is visible.
The index
index = median(engagement, posts in the band) ÷ median(engagement, learning set)
The learning set's median post earns 320. A band is reported only when it holds at least twenty posts.
The distribution is the reason for the median. Across the learning set the median post earns 319, the 90th percentile 1,287, the 99th 6,173 and the largest post 96,042. A mean would be set by a handful of posts; the median describes the typical post in a band.
The four lenses
The raw index counts audience as much as craft: whatever the largest accounts habitually do reads well on it. So each key band is read three more ways. Where the lenses disagree, the disagreement is reported as the finding, and the within-author reading is the one the study trusts.
Per 1,000 followers
per_1k = median(engagement ÷ author followers × 1,000) · index = band per_1k ÷ set per_1k
Followers are the author's count at collection. The set reads 7.2 per thousand. This lens strips the size of the audience out of the comparison and is where several raw findings reverse.
Comments share
share = Σ comments ÷ Σ (reactions + comments) · index = band share ÷ set share
Summed over the band, not averaged per post. The set's share is 18.8%. A band that converses more than it is applauded reads above 1.00.
Within the author
lift = engagement ÷ author median · index = median(lift, posts in the band)
Each post against its own author’s median, then the median of those ratios across the band. An index of 1.07 means the band’s posts run 7% above what their authors usually get. This is the lens that removes account size; it was computed for the first-line dimensions, the archetypes, topics and formats, the close, and the timing, cadence and link comparisons, and it is shown wherever it exists.
Within the author, paired
ratio = median over authors of (median A ÷ median B) · wins = authors where A > B
For a direct comparison (the same author’s text-only posts against their image posts; their posts under twelve hours after the last against twenty or more hours after; their afternoon posts against every other hour; their busier weeks against their quieter ones), each author’s two medians are compared and the median ratio across authors is reported with the count of authors where the first side wins. Only authors with enough posts on both sides are included; the count is published with each ratio.
How posts are banded
Every band is assigned by the same rule to every post, from the text and the media fields, with no judgement call. The classifier is the one NBK’s own post checker runs; the bands are labelled in the tables exactly as the classifier names them.
| Dimension | Rule |
|---|---|
| Length | Characters in the post text, inclusive of line breaks: under 300, 300 to 700, 700 to 1,200, 1,200 to 2,000, over 2,000. Words are counted on whitespace. |
| Blank lines | Empty lines between text lines: none, 1 to 3, 4 to 8, over 8. |
| Line rhythm | The post’s typical line length: short lines under 60 characters each, medium 60 to 110, paragraphs over 110 characters a line. |
| Before the first break | Characters before the first line break: under 60, 60 to 140, over 140. |
| Sentence length | Mean words per sentence, sentences split on terminal punctuation. |
| First line | The text before the first line break: its characters (under 40, 40 to 80, 80 to 140, over 140), its words (1 to 3, 4 to 7, 8 to 12, 13 to 20, over 20), its shape (one full sentence; a fragment; a question; ends open on a colon or an ellipsis; two or more sentences on the line), an all-caps word, an emoji, a number, who it addresses (you; I or we; neither), whether it names a subject (a noun phrase before the verb), a negation or contrast word, and a dangling reference (a pronoun or "here is" with nothing before it). |
| Opener | How the post begins: a story anchored in time (a date, a year, an age, "last week"), a number, first person, a question, or another way. |
| Close | How the post ends: on an ask, on a P.S., on a statement, on a question, or on a list line. The P.S. and question-mark counts are also banded on their own. |
| Format | As the platform labels the post: video, image, document (a carousel), text only, or a link card (an article preview). Whether a URL sits in the text, and where (in the last fifth of the text or earlier), and whether the text says the link is in the comments. |
| Image shape | From the media’s pixel dimensions: portrait (taller than wide), square, landscape (wider than tall), or no image. The number of images is banded separately. |
| Hashtags, mentions, emoji | Counted in the text: hashtags (none, 1 to 3, over 3), tagged people or companies (none, one, 2 or 3, more), emoji (none, 1 or 2, 3 or more), exclamation marks, all-caps words, question marks. |
| Numbers and substance | Numerals in the text (none, 1 or 2, 3 to 5, 6 or more); a percentage or a currency figure; list lines (lines opening with a bullet, a dash or a numeral); a quoted passage; "you" and "I" per hundred words. |
| Post type | Original (regular), repost, or quote post (a repost with the author’s own text above it). |
| Timing | The post’s timestamp in UTC: the hour, the day of the week, and five slots (05 to 08, 08 to 12, 12 to 17, 17 to 22, 22 to 05). |
| Gap | Hours since the same author’s previous post in the set: under 6, 6 to 12, 12 to 20, 20 to 30, 30 to 54, 2 to 4 days, 4 to 8 days, over 8 days, or the author’s first post in the sample. |
| Cadence | The author’s typical posts per week over their active weeks, banded by author; the index for a cadence band is the median of the authors’ own median engagements, because cadence is a property of an author rather than a post. |
| Age at collection | Days between the post and 10 September 2026: under 7, 7 to 30, 30 to 90, over 90. Published so the under-count on fresh posts is visible. |
Classification by rule
Archetype. Each post is assigned one archetype by a fixed sequence of text rules, the first that matches: a how-to or numbered post (a numbered list or a "how to" or "steps" frame), a contrarian claim (an opening negation of a received idea), a personal story (first person with a time anchor and a narrative verb), an announcement (a launch, a release, a join or a milestone in the first lines), a data drop (a figure-led opening with a source or a chart), a question-led post (a question in the first line with the body answering it), and otherwise an observation or a take. The rules are applied to the stored text, so a post’s archetype is fixed and reproducible.
Subject. Nine subject families are tagged by word families (leadership and teams; hiring and careers; content craft; money; AI; failure and lessons; life outside work; LinkedIn itself; agency and client life), and a tenth marks a post that names a platform. A post can carry several subjects. The subject tables report each family against the posts that do not carry it.
Time-bound. A post is time-bound when it carries a phrase that fixes it to a date ("today", "this week", "just launched", a named season); the phrase list was pruned by hand so that a bare year or "next week" in evergreen advice does not mark it. 4,240 of 39,858 posts (10.6%) are time-bound.
The by-eye reads
Images. The corpus measured whether a post carried an image and its shape, not what it showed, so the 94 highest-lifting image posts (one per author, each post ranked by its lift over its own author’s median, the readable ones of the top 100) were read by one reader to a written protocol and classified as: a photograph of the author, a meme, a text card, a screenshot, a chart, or another kind; whether the image carries words; and whether people appear in it. No vision model was used. The set is reported with its size, and its lift figures are indexed against the median lift of the set itself.
Banners. 41 banners of the strongest authors were read by the same reader and classified by kind (a positioning line, decorative, an offer, a proof point) and by the elements present (any words, the author shown, a wordmark, a call to action, logos, a proof number), with the word count recorded where there were words. Each author has one banner, so no lift can be computed and none is claimed: the banner table says what the strongest accounts carry, not what works.
Profile signals
26 signals were read from the profile fields and compared between the 44 strongest authors by median engagement and all 344 profiles. Shares are the percentage of profiles with the signal; medians are the median value. Headline and About measures are computed on the stored text: length in characters, the first line of the About, whether the headline opens with "I help", carries a number, uses pipes or separators, or states a role, a company and a proof point; whether the About carries a list, a call to action or a number. A profile cannot be measured against itself, so every row is prevalence and the study says so.
Exclusions and limitations
- Every index is correlational. Authors differ in audience, subject, country and craft, and a band collects whichever authors happen to write that way. The lenses reduce the confounding and do not remove it.
- The population is prominent writers, not LinkedIn at large: the learning set's median post earns 320 engagements, which is far above a typical member's. The findings describe what distinguishes posts among accounts that already reach people.
- Posts under a week old at collection had not finished earning: the study measured them at 0.55 to 0.87 of a settled post. The age band is published rather than corrected for.
- Reaction and comment counts are read as the platform displayed them on the day. The platform rounds large counts, hides some comment threads and removes some posts; the study does not adjust for any of this.
- Follower counts are read at collection while posts date from up to twelve years earlier, so the per-follower lens slightly understates the oldest posts of the fastest-growing authors. The rule is identical for every author.
- Timing is in UTC and the authors post from many time zones. The hour bands describe when the feed engaged, not when any one author’s audience is awake.
- Profile signals are prevalence, not lift: a profile cannot be measured against itself, so the profile section says what the strongest accounts have, never what made them strong. Creator mode is reported as the collection recorded it on the profile; the platform has since folded the mode into the profile’s standard settings.
- The by-eye reads (94 image posts, 41 banners) are small sets classified by one reader to a written protocol, and are reported as such.
- Bands were classified by rule from the text and the media fields, and a rule misreads the occasional post (a line that opens with a numeral is counted as a list line whether or not it is one). The classifier is the same for every post, so a misread is noise rather than bias.
- Thirteen dimensions were read on all four lenses and the close was re-examined within the author in depth after a cross-author reading proved to be measuring account size. The remaining dimensions were not re-examined to the same depth on that question.
- Nothing here is reach. LinkedIn does not publish impressions, and no claim about reach is made.
Disclosures
- Funding: the study was funded and conducted by NBK Social Research, the research imprint of NBK Social.
- Competing interests: NBK Social provides social media services to brands, publishers and creators, and its founders publish on LinkedIn. No commercial relationship with any author in the population exists, and none influenced inclusion, measurement or publication. NBK Social has no relationship with LinkedIn.
- Review status: this study has not been academically peer reviewed.
Versioning, data availability and citation
The study published on its permanent address on 9 October 2026. The corpus is frozen and the stored analysis will not change; any post-publication amendment to the tables or this page is dated in a changelog here.
The aggregate tables behind every section are the study page itself. The underlying posts and profiles are not available, to anyone, for the reason in section 2.
NBK Social Research (2026). What Works on LinkedIn: 74,119 posts measured. nbksocial.com/research/what-works-on-linkedin/
Frequently asked questions
Why is the study indexed against a median rather than an average?
Engagement on LinkedIn is heavily skewed: the learning set’s median post earns a few hundred engagements, its 99th percentile a few thousand and its largest post tens of thousands. An average would be set by a handful of outliers; a median describes the typical post in a band.
What is the within-author lens and why does it matter?
Each post is divided by its own author’s median engagement before the band median is taken, so an index of 1.07 means the band’s posts run 7% above what their authors usually get. It removes account size, which the raw index cannot. Where the two disagree, the within-author reading is the one to trust.
Why are the authors not named?
Because the study is about the shapes of posts, not the people who wrote them, and because a list of names would turn a measurement into a ranking that nobody consented to. No author, handle, post, first line or follower count is published, and the export that builds the study page refuses to write a file carrying one.
Can a band be read as a cause?
No. Every index is a correlation across accounts of very different sizes, subjects and audiences. The lenses reduce the confounding; they do not remove it, and the study says so beside every headline figure.
Why are posts under a week old under-counted?
A post keeps earning reactions and comments for days after it goes up. The collection happened on one day, so a post published that week was read before it had finished; the study measured the shortfall at 0.55 to 0.87 of a settled post, and the age band is published so a reader can see it.
Who is behind the study?
What Works on LinkedIn is researched by Matt Cunnelly and Josh Stoddard and published by NBK Social Research, the research imprint of NBK Social, a social operations agency working between Dubai and Manchester.