The X Post Format Benchmark 2026

1. 01The format finding in one answer
2. 02Why normalization changes the answer
3. 03All four formats, pooled
4. 04The link penalty, measured within accounts
5. 05What accounts actually post
6. 06The full 17-account table
7. 07The wider market, sourced
8. 08Method and dataset (CSV download)
9. 09Adjacent reading
10. 10Frequently asked questions

The format finding, in one answer

This is a measured association on a 168-post sample, not a claim about how any ranking system works internally. Format is not randomly assigned, so an account that saves video for its biggest news is confounding format with newsworthiness. The paired test narrows that risk without eliminating it, and the limits are listed in full in the method section.

Why normalizing by audience changes the answer

Almost every social benchmark you can find reports raw medians. Raw medians are the wrong instrument for a format question, because they measure which accounts happened to be in the sample at least as much as they measure the thing under study. This dataset shows the problem cleanly, and the disagreement between the two readings is more useful than either reading alone.

Rank the four formats by raw median likes and plain text comes joint second at 241, level with image. Rank them by likes per 100,000 followers and plain text falls to last at 59.5, well behind image at 108.0. Nothing about plain text changed between those two sentences. What changed is that the second one stops crediting the format for the fact that two of the largest accounts in the sample, OpenAI at 5,126,187 followers and AnthropicAI at 1,610,472, happen to post plain text often. An operator acting on the raw column would copy a format that works for a five-million-follower account and wonder why it did nothing for theirs.

All four formats, both readings side by side

Every one of the 168 original posts was classified into exactly one format: an outbound link, a video, an image, or plain text with no attachment and no link. The table reports the raw medians most benchmarks stop at, and the audience-normalized column beside them. Video leads on both. Plain text trails on the normalized read despite a respectable raw figure, which is the disagreement worth understanding.

FORKOFF X Post Format Benchmark, n=168 original posts, 17 accounts, 2026-07-14 to 2026-08-26, pulled 2026-08-26. Per-follower figures are computed per post then aggregated as a median. Medians of different formats are not summable, so the total row reports n/a rather than a meaningless number.

Measured inside each account, where audience size cannot lie

The pooled comparison can always be attacked on sample mix. This test cannot. It takes only the accounts that posted at least three link posts AND at least three non-link posts in the window, 8 of the 17, and compares each account's own link median against its own non-link median. Follower count, topic, brand strength, and posting habit are all held constant inside a row, because a row is one account measured against itself. A ratio below 1.0 means that account's link posts underperformed its own baseline.

The two accounts at parity are the interesting rows, not the exceptions to be waved away. @vercel posted 15 of its 18 original posts as links, the highest link share in the sample, and paid no measurable penalty at 1.022. @supabase came in similarly at 1.053. Whatever the penalty is, it is clearly not a flat tax applied to every link. The honest conclusion from a sample this size is that the cost varies with how the link is presented and with what the account has trained its audience to expect. The steepest row in the table, @PlanetScale at 0.377, gave up 62 percent of its own typical likes on a link post while @vercel gave up nothing. A brand posting links by default without ever measuring its own ratio is guessing about a variable it could simply check.

What these accounts actually post, by vertical

The behavioural half of the finding is the part a strategy can act on immediately. Across the whole sample, 48 posts carried an outbound link and 48 were video, an identical 29 percent and 29 percent of the sample. The weakest format on the per-follower read is being posted exactly as often as the strongest. Break it down by vertical and the habit is concentrated rather than universal.

Developer tools and infrastructure is the vertical with the habit. It posts 52 percent of its content as outbound links, far ahead of every other vertical, which makes sense for teams shipping changelogs and docs and makes it the group with the most to gain from measuring its own ratio. AI and ML products sit at the other end at 7 percent links and 53 percent video, leaning on the format the pooled data rewards most.

All 17 contributing accounts, every field published

The complete per-account table follows, carrying 13,907,614 combined followers with a median account size of 376,393. Combined follower counts overlap heavily across these audiences, so that total is a reach ceiling rather than a unique-person count and is reported as such. Format columns sum exactly to the post count on every row, which is the arithmetic check that the pooled totals are rebuilt from these rows rather than asserted alongside them.

FORKOFF X Post Format Benchmark, 17 accounts, n=168 original posts, pulled 2026-08-26. Reposts and replies were dropped and the author was verified against the requested handle before a post was counted. Account medians are not summable across accounts, so the total row reports n/a rather than a fabricated aggregate.

The platform this benchmark sits inside, sourced

The benchmark above is first-party and measured on this sample. The figures below are named third-party benchmarks that frame the platform, each with its source and year. They are directional context and none of them is pooled into the first-party numbers.

The Socialinsider figure is the one that most directly justifies this study's method. X carries the lowest per-follower engagement floor of the major platforms, which means a raw like count says more about how many followers an account has than about whether a post worked. That is precisely why the format comparison here is normalized by audience and then repeated within accounts, and why the two readings are published side by side rather than the flattering one alone.

How the benchmark was measured

The sample frame is 20 named public X accounts across four verticals that map to the markets FORKOFF sells into. The frame was fixed before collection and no account was substituted after the fact, which matters more than it sounds: swapping an unresolvable account for a convenient one is how a sample quietly becomes a selection.

The CSV carries the per-account table, the pooled format comparison in both raw and normalized form, the within-account paired ratios, and the per-vertical format mix. Attribution is FORKOFF (2026), X Post Format Benchmark 2026.

Adjacent reading

• Twitter and X Marketing , the managed service this benchmark was built to inform, run as a pipeline channel rather than a vanity feed.
• Reddit Community Access Index 2026 , the sibling study on the other community platform, measuring what 24 subreddits permit and how much each one actually discusses.
• Views-to-Likes Benchmark 2026 , the companion X study on reach authenticity, calibrating the point where a view count runs ahead of real engagement.
• State of Launch Videos 2026 , what happens to these formats at launch scale across 30 tracked product launches on X.
• The full research catalog , every FORKOFF first-party study with its sample size and published methodology.

X post format, common questions