Methodology

• Aggregates only. We never publish names, handles, post text, example posts or lists of top posts or creators.
• Minimum group sizes: every published number covers at least 30 authors and 200 posts, or 20 authors and 100 posts inside a follower band or region. Smaller cells are suppressed, not rounded.
• We never cross two small cuts (for example country by band by format).
• Working files with post text were processed locally and deleted after the analysis.
• Our privacy policy covers publishing aggregated, anonymised statistics.

Engagement is reactions + comments + reposts. We do not have views, saves or dwell time for most posts.

Within-author comparison. For each post we take log(1 + engagement) and subtract the average for the same author in the same year. A group's effect is the average of those differences, first within each author, then across authors, so every author counts once and a few big accounts cannot dominate. "+36% vs text" compares the same authors' image posts with their text-only posts. Comparing within the same year also removes follower growth over time.

Confidence intervals are 95% intervals from 400 bootstrap resamples of authors. Engagement snapshot: posts need at least 7 days between publishing and our engagement snapshot; posts can keep gaining engagement after that, which slightly understates totals for recent posts but affects all groups alike.

Local time uses the author's profile location, for single-time-zone countries plus US states, Canadian provinces and Australian states.

Follower growth uses monthly follower snapshots from June 2025 to August 2026 (months in which the person posted). Growth between consecutive months, and for longer histories the compound monthly growth from first to last snapshot.

Follower counts are current. We plotted all profiles on a log scale and looked for natural breaks: a jump in profile counts at 1,000, a dip near 3,000, a second cluster from about 4,000 to 6,000 that thins after 6,000, and a tail past 10,000. Jenks natural breaks on the same data give about 760, 2,900 and 11,000 (four groups) or 1,350, 3,550 and 10,400 (six groups). We used 1,000, 3,000, 6,000 and 10,000.

Post types, opening lines, topics, traits and "reads as AI" were labelled by Jev, a fast judgment model from TypeSafe that answers narrow questions with a probability, a choice or a score. It does not write. Each question was calibrated against posts we checked by hand before we trusted it (a single labeller; no post is quoted anywhere). Every eligible post was labelled for every question in one consistent pass with the same model version: all 130,976 posts of 150+ characters from January 2023 onward, plus the 1,554 pre-ChatGPT baseline posts. That covers 119,961 posts from 1,466 authors in the analysis set. On 1,800 posts checked against the earlier calibrated run, 98% to 99.9% of labels matched. Posts under 150 characters were not labelled.

The "reads as AI" judge also scored 1,554 posts written before ChatGPT's release (November 2022). It flagged 20% of them. We publish that false-positive floor next to every AI trend.

Posts published through LigoSocial itself are excluded from the AI-writing and content-label analyses, so our own product cannot influence those results.

• Observational data: associations, not proof of cause. Within-author comparisons remove many confounders, not all.
• Our authors chose an AI writing assistant, so they are more active and more AI-curious than the average LinkedIn member.
• No views, saves or dwell time; poll votes are not counted as engagement.
• Follower counts are current, and growth history covers about 15 months.
• 2026 covers January to September.

Everything here is free to use under CC BY 4.0. Please credit "LigoSocial Research" and link to the page you cite. The CSV files below hold every published number. If you spot an error, tell us at [email protected] and we will correct it and note the change.