Postdom research

Research methodology: how Postdom measures short-form video

This research methodology page sets out the methods every Postdom study shares: how we sample public posts, the account-median baseline we compare against, the group sizes we publish, and the rules that keep a finding honest. Each study adds its own details on its page.

Postdom compares each post with its own account's median views, so a few large accounts cannot drown out the rest, and marks any study group under 30 posts as directional.

Definition. The account-median baseline is a post's views divided by the median views of all eligible posts on the same account, so 1.00x is that account's typical post.

Last updated 10 October 2026. Next review: 5 January 2027. Data lead: Ana Ibáñez Lario.

Research methodology: the account-median baseline

Big accounts get more views whatever they post. To see what a feature does, we compare each post with its own account's typical post rather than with other accounts.

For each account, we take the median views of all its eligible posts (not ads, at least seven days old, three minutes or shorter in the video studies).
Each post's views are divided by that median. 1.00x is typical; 2.00x is twice the account's usual views.
To measure a feature, we compare posts with it and without it within the same accounts, using only accounts that have both, and report the result as a multiplier.
The 95% ranges come from 2,000 bootstrap resamples of accounts, so one prolific account cannot drive a result.

Sampling, and when we oversample

Public posts only. We study public posts, collected through a third-party data API. Accounts are found by searches in the niches each study names, then screened.
Oversampled studies. The video studies published on 5 October 2026 take, from each account, the three best posts, the three worst and the posts at the one-third and two-thirds ranks by views: eight per account. This over-represents the extremes and makes every multiplier larger than it would be across all posts. Read those studies for direction and ranking, not exact size. Every such page says so.
Top-of-search samples. The TikTok Trends Report reads the top of TikTok's search results, so it shows what worked, not typical outcomes.
Random samples. Claims about all posts need a random sample. The Short-Form Benchmarks sample now being collected is a fixed frame of business accounts per business kind, with each account's most recent posts. The first re-test on it is the "comment below" CTA finding, due 7 November 2026.

Minimum group sizes and the directional rule

Every page publishes the size of every group.
In studies, a group under 30 posts is marked directional and does not lead a page or a headline.
The TikTok Trends Report uses its own published threshold: a format with fewer than 10 posts is directional, because the report describes top posts rather than estimating an effect.
Where we estimate an effect we give its 95% range, and say when the range includes 1.00x.

Went with, not caused

Our studies are observational. A post that ends with "let me know below" may also differ in its idea, its hook or its week. So we write that a feature went with more or fewer views, never that it caused them. Where a claim needs a cause, we say what test would show it.

How AI models label posts

In the video studies, AI models label each post (hook, format, ending, tone, captions) from sampled frames, the transcript, on-screen text and the caption, without seeing its views or engagement.
Labels were not hand-checked post by post, and individual labels can be wrong. With many features tested, a few clear results could be chance; each page says how many were tested.
Where a platform shows its own label on a post, such as an AI-generated content label, we name the metric exactly ("labelled as AI"), because a label undercounts real use.

Accounts, privacy and what we publish

Published files are aggregates: group-level figures with no account handles and no links to posts.
Accounts with fewer than 10,000 followers are never named or linked. Larger accounts may be named where a page cites a public example.
Post-level files contain handles, so they are shared with journalists on request, not published.

Rules for customer data

No Postdom research page uses customer data yet. When one does, these rules apply:

Aggregates only, never one customer's figures. The terms of service say we may publish anonymized aggregates, and that a published figure never identifies a customer or workspace.
At least 10 workspaces and 30 posts in every published group.
Internal workspaces are excluded.
No names, handles or domains.

Corrections and updates

Every update and correction is listed on its page and in the research changelog, newest first, with its date. When a re-test shrinks or reverses a finding, we publish that too. Every page shows its next update date, and the press page lists the release calendar.

All Postdom research