Method and source data

Public counts collected 30–31 October 2026. The analysis window ends at 30 October 2026, 22:00 UTC. Window: 30 April–30 October 2026. 22 brands: Composio, Infisical, E2B, Browserbase, Firecrawl, Braintrust, WorkOS, turbopuffer, Daytona, Arcade.dev, Trigger.dev, Patronus AI, Fiddler AI, Convex, LlamaIndex, CrewAI, Vapi, Parallel, Clerk, MotherDuck, Doppler, ngrok. Supabase is excluded. Public owned-account posts only; founder accounts are outside scope.

Which posts are counted

7,808 normalized timeline records were collected through treg’s LinkedIn and X data routes. The analysis includes 3,939 eligible posts: 1,982 LinkedIn and 1,957 X. It excludes reposts of other authors, X replies, posts younger than seven days at the window end, posts outside the window, and records missing the platform’s primary public count. Underlying LinkedIn engagement IDs are deduplicated, preferring original posts over self-repost wrappers. Company-authored commentary on a quoted post is included; the quoted creative is identified in source metadata where available. Timeline pagination reached the time cutoff; Infisical’s X history required a provider fallback.

What “performance” means

The common outcome is combined public responses: reactions or likes, comments or replies, shares or reposts, plus quote posts where returned. It is a count of actions, not unique people or leads. X views are analyzed separately as reach; LinkedIn impressions are unavailable. Reactions and views remain visible on example cards, but views are never pooled with reactions in a common performance estimate.

Comparing like with like

1. For each category, fit a separate model within each company and platform. The outcome is log(1 + public responses); the predictor is membership in that category, with publication-month fixed effects. Only months containing both the category and other known categories are used. At least five category posts, five comparison posts and four residual degrees of freedom are required. HC3 standard errors reduce sensitivity to unequal variance.
2. Exponentiating the category coefficient gives an approximate multiplicative difference in the geometric mean of 1 + responses. A value of 2× means roughly twice the response; 1× means similar performance. This is not a median, an engagement rate, or an estimated causal effect.
3. Within a brand, LinkedIn and X estimates receive equal weight. Their combined variance uses an upper bound for unknown covariance plus observed platform dispersion. Across brands, a random-effects meta-analysis accounts for differences in category effects. Modified Hartung–Knapp intervals with a variance floor and t critical values reflect the small number of companies. At least three comparable brands are needed for an overall ranking or broad conclusion.
4. Higher/lower labels require a 95% interval that excludes 1× and a Benjamini–Hochberg adjusted q value at or below 0.10. Corrections are applied within each family of category comparisons. A point ranking is still shown when estimates overlap, but overlapping estimates are not evidence of an ordering.
5. Direct pair comparisons use only the two categories, within the same company/platform/months. These are preferable when asking whether one particular category beats another; independent category-versus-rest estimates can use different company sets.

When platforms are split out

For brands with both platform estimates, compare their log effects. The uncertainty uses a conservative bound for unknown cross-platform covariance. Show separate platform findings prominently only with at least three paired brands, an adjusted q value at or below 0.10, a 95% interval excluding no difference, and an estimated platform difference of at least 1.5×. Prominently split platform estimates use exactly the same companies on both sides. General platform details remain available for all comparable companies even when this threshold is not met. A lack of evidence for a difference is not proof of equivalence.

Ranking individual posts

Example rankings divide 1 + a post’s response count by 1 + the median response count of the same company, platform and publication month. When a month has fewer than five posts, the full company/platform median is used and the card states that fallback. This avoids ranking large accounts above small ones simply because of audience size. The cards show this relative score prominently, along with absolute counts. Equal top and bottom groups contain up to ten posts per company/platform; up to six examples at each end are displayed; small company samples show fewer. These illustrative rankings are separate from the all-post statistical models.

Limits that affect interpretation

This is an observational comparison of 22 selected brands, not a randomized experiment. Platform distribution, paid promotion, audience quality, post age within a month, campaign timing, topic and format combinations, and business outcomes are not fully observed. Topic and format models are separate; they do not isolate format from message. A viral outlier does not establish a repeatable play. A security-incident disclosure is retained in the inventory but excluded from performance models and recommended-example rankings. Uncertain labels are excluded in a sensitivity analysis, with fragile results flagged. Funding and recruiting contexts remain visible on relevant examples.

The old 208-post editorial sample is retained in source files for provenance; it no longer supplies the inventory or performance estimates in this report. The full corpus now drives both. Synthetic checks verify that the models recover a known within-company effect, reject audience-size-only confounding, and retain a null result when categories perform alike.