Our three-level method
- A, established: official LinkedIn announcement, engineering paper, standard (accessibility) or peer-reviewed study.
- B, observed: correlation published on a large sample, often by a tool vendor. Useful, but not a cause.
- C, heuristic: good practice without published measurement.
The strongest studies compare each author with themselves, which neutralises audience size. No peer-reviewed controlled experiment on LinkedIn format tactics exists to date.
What is established (level A)
- Dwell time is a ranking signal documented by LinkedIn engineering. No threshold is published.
- On 20 May 2026, LinkedIn announced it would reduce distribution of generic AI-generated content beyond the immediate network, claiming 94% precision with no published false-positive data. A reporting button followed in late July.
- A post is limited to 3,000 characters.
- Unicode bold is read as a string of mathematical symbols by screen readers: an accessibility issue, not a performance one.
What is observed (level B)
- Concrete and lived: in the MagicPost study (40,937 posts), posts with neither concrete detail nor lived experience lose about 23% reach since late July 2026, posts with both gain about 15%. The clearest lever measured this year.
- External link in the body: less reach in every dataset, from −10% to −48% depending on method. Details.
- Question hook: less engagement in 5 of 6 studies (−8% to −34%). Details.
- Hashtags: no benefit beyond 3, null to negative effect depending on source. Details.
- First-person story: about +33% in two datasets, one within-author.
- Carousels (PDF documents): more median impressions than other formats; the best page count varies by study (often 9 to 12).
What is contested
- Post length: engagement studies favour long posts, reach-per-impression studies find no optimum.
- Number in line 1: +35% in one study, no effect in another.
- Share of AI content on LinkedIn: 41% to 81% depending on the detector. Detectors have high false-positive rates on formal writing and non-native authors.
What nobody knows
How the generic-content filter works exactly, how posts drafted with AI then rewritten are treated, the ideal first-line length outside English: no published data. Be wary of anyone quoting a precise figure on these.
What we do with it
Our post analyzer applies these findings without generative AI: it flags, explains with the source and offers mechanical fixes. It never rewrites your text or invents facts. Nor does it predict reach, since no public model does so reliably.
Paste your text in the editor, then tap “Check before copying”: flagged passages, one-click fixes, everything stays in your browser.
These figures come from observational studies published by tool vendors, not peer-reviewed. They describe correlations. LinkedIn is a trademark of LinkedIn Corporation; LinkedIn Format Pro is not affiliated with, endorsed or sponsored by LinkedIn.