Jev makes content research
feel like cheating.

1. Collection snapshot: 21 September 2026, 11:12 UTC. The 500 collected posts span 1 April to 21 September 2026. These are saved figures, not current Instagram counts and not the entire account.
2. Coverage: 447 completed classifications, 45 no-audio posts, seven with too little speech, and one failed media item. Four of the completed posts were already marked as duplicate scripts. Removing those leaves 443. Removing 12 posts less than seven days old at collection leaves the 431-post comparison sample. All 431 have positive plays and known likes and comments.
3. Plays are the available metric. The views field is absent in this sample, so we do not rename plays as unique reach or views. One person can generate more than one play. The baseline is the median of the 431 eligible posts: 358,306 plays, 17.64 likes per 1,000 plays and 0.267 comments per 1,000 plays.
4. For each label, we report the median post, not the average and not the pooled engagement rate. Likes per 1,000 = likes / plays × 1,000, calculated per post before taking the median. Multiples compare a category median with the overall sample median, which includes that category; they are descriptive, not incremental lift.
5. All eight dimensions use the stored Jev labels: topic, hook mechanism, opening form, structure, evidence, emotion, specificity and spoken CTA. Confidence of at least 0.70 is a sensitivity filter, not proof the label is correct. Missing confidence fails that filter. We reviewed the complete stored transcripts of the six featured examples and additional candidate and weaker examples. We did not independently verify every line against video or reclassify all 431 scripts.
6. Groups with fewer than 15 posts are shown but marked as small samples. Fifteen is an editorial screening rule, not a statistical guarantee. No significance tests, causal estimates or predictive validation were performed. We checked selected patterns by topic, publication period, label confidence and repeated-script sensitivity.
7. A second duplicate check found five later posts with substantial script overlap beyond the existing duplicate flags. It compared unique five-word sequences, required at least 35 such sequences in each script and flagged overlap of at least 75% of the smaller set. Removing the later candidates gives 426 posts; steps retain a 627,228 median across 30. These are sensitivity results, not a new primary sample or five manually confirmed duplicates.
8. The personal reveal DZu7FiJhmNU was labeled steps, illustrating category noise. Removing only that post gives 584K median plays across the other 30 steps posts. Lists of advice and sequential instructions also share the steps label. Read the examples before treating it as a precise creative format.
9. This is an observational sample from one large established account. Posts differ in age, topic, duration, speakers, format and likely distribution. Seven days is a minimum-age filter, not equal observation time. We do not have retention, watch time, saves, shares, follower growth attributable to each post, paid distribution, leads or sales. The results describe associations, not why a Reel succeeded or what will work on another account.
10. Example quotations are short excerpts from machine transcripts and may contain transcription errors. Open the original before quoting or adapting. The templates are newly written suggestions, not Hormozi quotations. Public CSV exports contain source links, snapshot metrics, labels, confidence and short openings; they omit credentials, signed media URLs, raw provider responses and full transcripts.

Source: saved @hormozi Creator Lab run, collected 21 September 2026. Analysis prepared 24 September 2026. No fresh collection or paid classification was needed for this report.