Full results table (views vs. the same accounts' videos without the feature)
All 1,272 videos pooled. "n" is videos with the feature in accounts that posted both with and without it (only those accounts contribute to the comparison). * = the 95% range excludes 1x.
TikTok saves and shares are rates per view against the account's median rate, compared the same way: videos with the feature against the same accounts' videos without it. Instagram does not expose saves or shares publicly, so they are TikTok-only. The full 59-row table, with TikTok and Instagram split out, is in the CSV.
By niche
Accounts in all four niches had a similar number of eligible recent videos (median 31 to 33), so the differences in best-video size are not a posting-volume effect.
Methodology
Collection. Public data was collected on Oct. 2, 2026 through a third-party scraping API (ScrapeCreators). TikTok accounts were found through 91 TikTok searches in four niches (relocation and immigration; local and professional services; software and apps; consumer brands), which returned 4,472 videos from 3,246 authors. A text pre-screen kept mostly brands and businesses, plus some niche creators. A vision check of four covers per account kept accounts that post their own footage with in-app text on at least three of four covers (632 of 1,162 screened). 140 TikTok accounts were then sampled by niche quota. Twenty Instagram brand and business accounts in the same niches were added by hand.
Who is in it. 1,120 TikToks from 140 accounts and 152 Reels from 20 accounts. By account type: 849 videos from business owners posting for their business, 136 from brand accounts, 287 from creators. Median TikTok account: about 28,000 followers, with a median video of about 3,000 views. 76% of the videos were posted in 2026; the oldest dates from January 2021. Median video length: 44 seconds.
Eligibility and sampling. A video was eligible if it was not an ad, was at least seven days old and was three minutes or shorter. Each account needed at least 10 eligible videos. For each account, every eligible video's views were divided by the median of all its eligible videos. From each account we analyzed the three best, the three worst and the videos at the one-third and two-thirds ranks: eight per account.
Labels. Labels were produced without access to view counts. Each video was sampled at eight frames (0.1 s, 1.5 s, 3 s, then evenly spaced), transcribed (OpenAI gpt-4o-mini-transcribe), and on-screen text and shot types were read from the frames (OpenAI gpt-4.1-mini). Hook type, format, ending and caption tags came from the text. Cuts per minute came from ffmpeg scene detection. A per-account vision pass marked whether each video shows the account's recurring presenter. Videos were downloaded transiently and deleted; only small frames were kept.
Statistics. For each feature, we compared the average log view-ratio of videos with and without it, using only accounts that had both, and report the result as a multiplier. 95% ranges come from 2,000 bootstrap resamples of accounts (fixed seed), so one prolific account cannot drive a result. "Share of top-3 / bottom-3" in the CSV is a second, design-robust view: how often a feature appears among accounts' three best versus three worst videos.
Limits.
- Because the best and worst videos were deliberately sampled, multipliers are larger than they would be across all videos. Trust the direction and ranking more than the exact size.
- This is observational. A feature that tracks with higher views may be a marker of something else (a better idea, a better week), not the cause.
- The accounts were screened for a particular style (own footage, in-app text). Results describe that population, not all of TikTok or Instagram, and not entertainment or celebrity accounts.
- Instagram results come from 20 accounts and are directional. TikTok drives most findings.
- Several striking results rest on small groups: credential hooks (30 videos), question captions (23), giant title cards (27).
- Labels come from AI models and were not hand-checked video by video; individual labels can be wrong.
- Many features were tested (59), so a few "significant" results could be chance.
Reproduce. The aggregate results for all 59 features can be downloaded as a CSV in the data section above, and every step of the method is described in this section.
The methods every Postdom study shares are on the research methodology page.