Engagement Rate Benchmarks for Short-Form Video: One Metric, Two Denominators

The same platform, two published engagement rates

Engagement rate benchmarks for short-form video are not comparable, because the same word is measured against different denominators: Instagram carousels read 6.9 percent engagement per person reached in one published study and 0.55 percent per follower in another. Compare an AI-generated cut against the wrong benchmark and the keep-or-kill decision is wrong before the data arrives. The fix is procedural, not statistical.

The cleanest large dataset available is Socialinsider's 2026 cross-platform study of 69 million short-form videos published between January 2025 and July 2026. It puts the average engagement rate at 2.60 percent on TikTok, 0.45 percent on Instagram Reels and 0.30 percent on YouTube Shorts, with TikTok down from 3.70 percent the year before and Shorts the only platform that rose. These are follower-relative numbers: interactions divided by audience size. Even the publisher of this dataset warns against reading the rate alone. The same study reports average comments per video of 50 on TikTok, 20 on Reels and 10 on Shorts, and notes that TikTok's higher rate is driven by a genuinely more conversational audience — not only by a different denominator.

Now put that beside a second published method. Buffer's analysis of more than 4 million posts measured Instagram carousels at 6.9 percent engagement per person reached, against 3.3 percent for Reels. Socialinsider's Instagram benchmark puts carousels at 0.55 percent and Reels at 0.52 percent. Same formats, same platform, same year — and the carousel number differs by more than a factor of twelve. Neither study is wrong. Buffer measures the quality of each impression a post receives; Socialinsider measures response relative to audience size. The error happens downstream, when both numbers land in the same comparison table and quietly become one benchmark that measures nothing at all.

There is a third split hiding inside the first. A 2.60 percent TikTok average is a platform-wide number, while brand accounts posting in the 15 to 30 second range benchmark at 6 percent in the same research. A benchmark can therefore be true of the platform and useless for your account at the same time. Anyone who has compared a small brand account against a platform-wide average has felt this: the average includes viral outliers, personal accounts and every possible length, while your cut competes inside one niche, one format and one call to action.

When definitions flip the verdict

The clearest demonstration is a question the industry keeps asking and the data keeps answering both ways. Buffer's 4-million-post sample found TikTok video at a 3.39 percent median engagement rate versus 1.92 percent for carousels — video ahead by 77 percent. Fanpage Karma's 2025 study of roughly 700,000 posts found the opposite on the same platform: carousels ahead by 81 percent. The benchmark collection that gathers both studies states the reason plainly: the samples, date ranges and engagement definitions differ, which is enough to flip the result.

For a team shipping AI-generated variants, the risk is not the average — it is the mismatch between the question being asked and the definition the study used. A motion-led generated cut and a frame-led cut answer different creative questions, and each can find a published benchmark that blesses it. Length already splits the same way, because {{link}} showed short clips winning engagement while longer ones won reach. Pick the verdict first, then search for the denominator that produces it, and the benchmark exercise becomes decoration for a decision already made.

The honest reading is that these studies are instruments, not verdicts. A per-follower rate tells a brand with a large, quiet audience whether its content moves that audience. A per-reach rate tells a performance team whether an impression was worth buying. Asking one number to settle which creative approach wins is like asking two thermometers calibrated in different units which room is warmer — the disagreement is the calibration, not the room.

Length already splits the same way, because the duration paradox in short-form video showed short clips winning engagement while longer ones won reach.

Cheap generation turns a niche error into an expensive one

Two 2026 datasets explain why this matters more now. Wistia's State of Video research analyzed more than 13 million videos and surveyed nearly 1,000 marketers; it found social engagement nearly doubled as a success metric in a single year, from 12 percent to 22 percent, and that 62 percent of teams already use or plan to use AI in their video workflows. Wyzowl's survey — its twelfth consecutive year, 266 respondents — found 67 percent of video marketers quantify ROI through views and 63 percent through engagement such as likes, shares and reposts.

Those are the numbers steering budget, and the pipeline behind them has changed shape. When a variant costs minutes instead of a shoot day, teams ship more of them, and every variant gets compared against something. The something is usually a benchmark number copied out of an article without its definition attached. Cheap generation widens {{link}} unless every comparison starts from a named denominator. The measurement discipline that used to be a quarterly reporting chore now sits inside the loop that decides which cut ships on Friday.

The sentiment data points the same direction. Wyzowl found the share of marketers reporting a good return from video fell from 93 percent to 82 percent even as more businesses adopt video than ever. Output is rising while confidence in judging output is falling. In that gap, published benchmarks act as a substitute for judgment — which makes their comparability an operational risk rather than a trivia question.

Cheap generation widens the proof gap between producing and demonstrating unless every comparison starts from a named denominator.

The hook number nobody can source

The same discipline applies to the most quoted number in short-form video: that 71 percent of viewers decide within three seconds, or that a good thumb-stop rate on Meta is 30 percent. The benchmark page that traces citations most carefully reports that every instance of these figures traced back to third-party blogs citing platform reports without a linkable original document, and that no primary study with a stated sample supports a specific percentage. The principle is real — the first seconds decide distribution — but the exact figures are folklore.

For AI video teams this is a filter, not a footnote. Generation produces numbers fast, and the temptation is to justify creative rules with whatever statistic is nearest. An unsourced hook percentage has a way of hardening into a template rule — front-load the reveal, cut before the third second — that then shapes every prompt in the library. When a number steers a decision, ask for the study, the sample and the denominator. If none exist, the rule can still stand, but it stands as a hypothesis to test on your own account, not as a benchmark.

A three-step rule for reading engagement rate benchmarks

The fix is procedural. First, name the denominator before the number: per follower, per view, per person reached, median or mean — a rate without its denominator is a sentence without a subject. Second, compare within a single source. Buffer and Socialinsider are both credible and mutually incomparable; a table that blends them has more in common with a collage than a benchmark. Third, sanity-check any rate against an absolute count such as comments per video, because absolute counts survive denominator swaps and rates do not.

Benchmark freshness has its own clock, and {{link}} is the reason an old average may describe a feed that no longer exists. The durable alternative is to shrink the distance between the benchmark and your own data. Wistia's numbers show why owned channels help: homepage videos see a 24 percent play rate, and almost 20 percent of viewers who see a lead generation form in a video fill it out — metrics with definitions you control and a denominator you can name. And when the comparison is done, benchmarks say whether a cut performed, while {{link}} says whether the feeling it produced was the intended one.

Nothing here requires new tooling or a re-organized team. It requires one habit: whenever an engagement rate enters a decision, the denominator enters with it. The teams that win with AI video will not be the ones that find the best benchmark — they will be the ones that stop needing one.

Benchmark freshness has its own clock, and the ten-day shelf life of a short-form post is the reason an old average may describe a feed that no longer exists.

And when the comparison is done, benchmarks say whether a cut performed, while emotion testing for AI-generated video says whether the feeling it produced was the intended one.