How we measure LinkedIn engagement

1. The study behind the model

We scraped every original post published between January 1, 2025 and June 30, 2026 by founders in Y Combinator's W24 and S24 batches: 11,116 posts from 307 W24 accounts and 9,489 from 307 S24 accounts, all collected on July 16, 2026 so engagement had at least two weeks to settle. W24 was used for exploration and S24 held out for replication. Engagement is measured in reactions, the one public count with enough variation to compare.

Pooling mixes who posted with what was posted. Fitting each account separately removes the author: every founder with 18 or more posts was modeled as X_i ~ LN(μ_i, σ_i²), where e^μ is the account's typical post and e^σ its spread. The two parameters were uncorrelated (Spearman ρ = +0.05 in W24, −0.08 in S24).

2. How the benchmark applies the model

Four rules follow from the study. Each one changes a number you see on the dashboard.

Compare accounts by their typical post

An account is a person or a company page, never the two pooled. Its typical post is the geometric mean e^μ: average the logs, convert back. Unlike the arithmetic average, one viral post barely moves it.

Score a post only against its own author

A post's score is z = (ln reactions − μ) / σ for its author, shown as a multiple such as “2.3x their usual”. Ranking raw posts across accounts mostly ranks audience size.

Rank with uncertainty

The leaderboard orders accounts by typical post with a 95% interval, exp(μ ± 1.96·σ/√n). Where two intervals overlap, the order between those accounts is not settled, and the rank says so.

Measure post types within account

Category lift compares each account's posts of a type with that account's own typical post, then averages across accounts, so a category scores high only if it lifts the same people above their usual.

3. Other measures we report, and when

The typical post answers one question: how well an account's posts perform. Other questions need other statistics, and the dashboard uses them where they fit.

4. Data and limitations

• Sources. Public posts from tracked LinkedIn company pages and their employees' profiles, collected without logging in. Each post is classified once by type and format. Original posts only; reposts are shown separately.
• Maturity and minimums. A post counts toward a fit 14 days after publication. An account needs 18 or more such posts to be scored, the study's floor; below that it reads “not enough posts”.
• Time. Days and hours are in Pacific time. A competitor whose newest post is more than 45 days older than the freshest company's is left out of that comparison and named; everyone else counts through their own newest post.
• Coverage. Profile scraping stops at 500 posts per person, so all-time totals undercount prolific posters. Collection gaps happen; each page shows how far its data reaches.
• Exclusions. A short list of pages whose engagement comes mostly from one outside creator stays visible but is left out of rankings.
• What this is not. A single model over a long window does not capture trends over time, and LinkedIn changed its feed ranker during the study period. These are descriptive statistics, not causal claims.

{
  "mcpServers": {
    "imagine-benchmark": {
      "url": "https://benchmark.useimagine.ai/api/mcp",
      "headers": {
        "Authorization": "Bearer <your token>"
      }
    }
  }
}