Bluesky feed engagement — image-count analysis

*** p < 0.001 ** p < 0.01 * p < 0.05 ns not significant · Cliff's d: <0.1 negligible · 0.1–0.3 small · 0.3–0.5 medium · >0.5 large

TL;DR

  1. 1 vs 2+ photos — essentially tied. Mann-Whitney p=0.13 ns, Cliff's d=−0.02 (negligible).
  2. 4-image posts underperform everything else (real & meaningful). Mean 28.6 likes vs 51.7 for 1-img; Cliff's d=+0.20 (small), p=4e-13 ***.
  3. 4-img drop is not a time-of-day artifact. Stouffer z=7.4 after stratifying by hour.
  4. The 4-img penalty is moderated by account size (small & large accounts feel it; medium doesn't).
  5. With typical-length captions (50–150 chars), the 4-img penalty vanishes (p=0.75 ns).
  6. Follower count is the dominant predictor (Spearman r=+0.47; regression IRR=2.0 per log unit).

Sample

Top-line: 1 photo vs 2+ photos

Means across engagement metrics

1 vs 2+ summary

Per-image count (1 vs 2 vs 3 vs 4)

Mean likes (95% bootstrap CI)

Median likes

Pairwise Mann-Whitney (Holm-corrected) on likes

Dedicated comparisons

1 vs 2 photos

1 vs 3 photos

1 vs 4 photos

Correlation r values (likes ~ predictor)

Time of day

Mean likes by hour of day (UTC)

Author "size" — two definitions, side by side

The left chart tiers authors by actual follower count (at analysis time). The right chart tiers authors by median likes per post — a proxy for effective reach. Interesting divergence: under the follower definition, medium-size accounts show no 4-img penalty; under median-engagement, the pattern is different.

By follower count

By median engagement

Tier-by-tier 1-vs-4 likes

Text features

Caption length by image count

Alt-text use by image count

1-vs-4 stratified by caption length

Additional analyses

Image orientation (all image posts)

Inequality: Gini by image count

Zero/low-engagement rate

Composition — per 100 likes

Multivariate regression

Negative binomial GLM: likes ~ image_count dummies + log(1+followers) + has_alt + caption_length + hour (sin/cos). This is the single most important result — it controls for everything at once.

Incident Rate Ratios (IRR) with 95% CI