How the research works

1. What each post is

A language model (Claude Sonnet by Anthropic) reads the text of each post and fills in 7 labels: the post type, the topic, how the first line opens, who it talks to, its tone, what it carries (an own number, a list, a quote) and what it asks of readers. Each label has a fixed list of options with a test for each. Another 16 are counted from the text and the post's data, such as length, line breaks, the first line, questions, emoji, hashtags, mentions, quote posts, the time of day and the weekday. On top, the analysis groups posts by the gap since the account's previous post. The model never sees a post's numbers.

2. Each post against its own account

Reach is the number of views X counts for a post, divided by the median views of the 30 posts the same account published closest in time. Replies are set against the mean of those same posts, bookmarks against the account's mean. A post with a reach of 2.0 did twice as well as its account usually does at that time, whether the account is small or large. Replies to other people's posts are left out.

3. A group against the rest

For every label value with at least 8 posts in an account, the group is compared with the account's other posts: medians for reach, means for bookmarks and replies. That gap is what a report shows, and on these pages it is the “median gap”: the middle of the gaps of all accounts that had the group.

4. The test against chance

The labels of the account's posts are shuffled 1,000 times, and each time the same gap is measured again. The share of shuffles with a gap at least as large is how often chance alone produces it; bookmarks and replies are tested on ranks, so a single viral post does not carry a group on its own. Because every account tests dozens of groups, the result is corrected for the number of tests (Benjamini-Hochberg). A gap passes when its corrected value is below 0.25, the threshold at which a report shows a finding; below 0.05 a report calls it strong. At that threshold up to one in four passing tests can still be chance, which is why one account alone says little and the pages rely on patterns that pass in several accounts.

5. From accounts to a finding

“Reliably higher in 9 of 13 accounts” counts the accounts where the gap passed the test and pointed up, out of the accounts that had enough posts in the group. A sentence on a research page states a direction only when more accounts passed that way than the other way and the median gap agrees; the pages are rebuilt from the data and fail to render when a sentence stops being true.

What the data cannot tell

• Whether the kind of post causes the gap. A group can do well because of the topics or the days it tends to come with.
• Anything about images, video or links: the reference set does not reliably tell a link from an attached image.
• The creator's time zone. Hours are read in UTC; a report on your own account uses yours.
• Accounts outside this niche or far larger than 3,735 followers.

Data and citation

research/data.csv holds one row per group and metric: posts, accounts, accounts tested, accounts reliably higher and lower, and the median gap. No handles, no post texts, no single account's numbers. The accounts stay anonymous. The data is licensed under CC BY 4.0: free to use, with a link to the source.

To cite: Own Edge Research, “Is there a best time to post on X (Twitter)?”, https://ownedge.app/research/best-time-to-post-on-x, data as of 29 Sept 2026. The same pattern works for every page.