How to read this
A claim is cheap. A falsifier costs something.
Anyone can say a tactic works. The question that separates a finding from an opinion is what result would have made you admit you were wrong — and whether you wrote it down before you looked.
So every claim below carries a counter-outcome: the direction of result that counts against it. If it happens, the claim is reported that way. Not reframed, not quietly dropped, not re-run until it behaves.
A counter-outcome is a direction, not a threshold. A small or uncertain difference settles nothing by itself, and some of these will return no verdict at all. The rules for reading each result are fixed before that comparison starts, and they are below.
From the first post on 26 October 2026 this document is closed. Any change after that is an amendment: dated, with its reason, published beside the wording it replaces.
The register
Nine claims
Slot codes mark which test period each claim runs in. B blocks are bench tests on Reels; P periods are feed tests.
1B4 · Hook length28 Dec – 17 Jan
ClaimFor matched opening content, topic and format, Reels whose on-screen opening runs to seven words or fewer record a lower skip rate than comparable Reels whose opening runs to eleven words or more.
Counter-outcomeOpenings of eleven words or more record a skip rate at or below openings of seven words or fewer.
Instrument. Skip rate, as the app defines it on the account checked. Its definition, availability and reading age are recorded before testing, and no other retention field is substituted later.
TF.E hypothesis. The seven-word cutoff is a house choice. The eleven-word cutoff is a comparison choice; any hook-collection benchmark needs an identified source and a recorded counting method before being cited.
2B1 · Problem framing26 Oct – 15 Nov
ClaimFor matched topic, format and publication window, problem-framed openings record a lower skip rate than openings presenting the corresponding desired outcome. Exact paired wording is fixed before publication.
Counter-outcomeOutcome-framed openings record a skip rate at or below problem-framed openings, read under the decision rules fixed beforehand.
Instrument. Skip rate. Save rate is also recorded, but is not the verdict.
Source adjacent — the peer-reviewed evidence measured news-headline click-through, not Instagram retention.
3B2 · B-roll vs text16 Nov – 6 Dec
ClaimFor matched message, topic and publication window, Reels using relevant B-roll footage earn more Viewers than comparable Reels using plain text-on-screen.
Counter-outcomePlain text-on-screen Reels earn Viewers at or above B-roll Reels.
Instrument. Viewers, with Views logged in parallel. Skip rate is secondary and cannot replace Viewers once results are seen. The comparison covers the footage and text treatments chosen, not the worth of a footage library in general.
Commonly advised — and it tests the premise of our own stock-footage library.
4B3 · AI-drafted hooks7 – 27 Dec
ClaimFor matched topic and post job, Reel hooks that are AI-drafted record a higher skip rate than hooks that are creator-drafted.
Counter-outcomeAI-drafted hooks record a skip rate at or below creator-drafted hooks.
Instrument. Skip rate. The model and version, the prompt, the editing allowance and both drafting processes are recorded before a hook is written. Topic, footage, typography, timing and payoff stay comparable, and neither arm is chosen after performance is known.
TF.E hypothesis — no source we found compares post performance by drafting method, in either direction.
5P3 · AI-drafted captions23 Nov – 6 Dec
ClaimFor matched topic and post job, captions that are AI-drafted earn a lower save rate than captions that are creator-drafted.
Counter-outcomeAI-drafted captions earn a save rate at or above creator-drafted captions.
Instrument. Saves ÷ Viewers, computed by hand and reported as saves per viewed post — the rule is below. Not Instagram's native Save rate, which the calibration carousel did not display. Both drafting processes are defined before the comparison begins.
TF.E hypothesis.
6P6 · Reposting your own work4 – 17 Jan
ClaimFor comparable topic and publication window, a substantially unchanged republication of our own previously published asset earns more Viewers than a new post on a comparable topic.
Counter-outcomeRepublications earn Viewers at or below comparable new posts, read under the decision rules fixed beforehand.
Definition. A repost is our own previously published asset republished substantially unchanged. An adaptation is not part of this test, and how originals are selected for reposting is fixed before the comparison starts.
⚠ This is not Instagram's “Repost”, which counts other accounts resharing your post — a different thing entirely, and the verdict must never be read from that field.
The contradiction this tests. Creators say repost your best work. Meta says that where it identifies matching content, it may recommend the original rather than the repost. That guidance is aimed at duplicates and aggregators, so the two may never collide — or they may collide squarely, if the system does not care whose work it was. Nobody has published a comparison, and we could not find one.
Commonly advised — repost your best work, the people who missed it will see it. We predict with the advice, against the platform's note. One of the two is less reliable than it sounds.
7P2 · Carousel cover framing9 – 22 Nov
ClaimFor matched carousel topic, length and publication window, covers that state the subject plainly earn a higher save rate than covers that open a curiosity gap without naming the subject.
Counter-outcomeCuriosity-gap covers earn a save rate at or above plain-subject covers.
Instrument. Saves ÷ Viewers, computed by hand on the same terms as claim 5, and reported as saves per viewed post.
TF.E hypothesis: plain-subject covers outperform curiosity-gap covers on the selected outcome. The general advice to improve covers does not establish this specific comparison.
8P4 · The freebie promise7 – 20 Dec
ClaimFor one product under one name on one landing page, a call to action framed as the reader's outcome earns more trigger comments per 100 Viewers than one framed as the author's competence.
Counter-outcomeThe outcome-framed promise earns trigger comments per 100 Viewers at or below the competence-framed promise.
Instrument. Eligible trigger-word comments per 100 Viewers, counted by hand — not the app's native Comment rate. Repeated comments, our own replies and invalid triggers are defined before testing. Link taps and opt-ins are secondary.
Adjacent contrast — the nearest published study tests negative versus neutral wording, not outcome versus competence.
9P5 · Direct ask vs comment-to-DM21 Dec – 3 Jan
ClaimConditional. Comment-to-DM earns more email leads per 1,000 Views than a link in bio.
Counter-outcomeA link in bio earns email leads per 1,000 Views at or above comment-to-DM.
It only activates if the reach is there. If the preceding fortnight records fewer than 15,000 Views in total, this claim does not fire, stays recorded as mechanism-only, and the reason is published alongside the reach actually reached.
Conditional comparison, recorded 27 September 2026. The reach gate is an operational choice; clearing it does not guarantee enough email leads for a verdict.
The rules, fixed in advance
How a verdict gets decided
A reach floor on the save comparisons. Claims 5 and 7 divide saves by Viewers, and a ratio built on a handful of people is not a rate. Both need at least 100 Viewers in each arm before a comparative verdict is considered; below that the result is published as no verdict — insufficient reach, with the counts actually reached. Claim 6 is counted in Viewers rather than as a ratio, so the same floor applies there per post.
The floor exists because a calibration carousel published in late September was seen by one person in 48 hours. Had they saved it, the save rate would read 100%; had they not, 0%. Both look like findings and neither is one. 100 is a practical minimum, not a statistical guarantee — clearing it does not promise enough information to separate two treatments.
The floor constrains reach only. No floor of any kind sits on saves, comments or any other numerator, so a null result can never disqualify itself. Nobody saving either version is a real finding.
What the denominator is. Viewers are summed across an arm's posts, so the count is viewer-impressions rather than unique people — someone who saw four of an arm's carousels counts four times. The figure is saves per viewed post, and every published statement of it says so.
Each comparison's protocol is written before its own first post. Dated, and covering the post count and pairing, how treatments are assigned, the exact wording or asset rules, the metric and the age at which every post is read, how the results are combined, the decision rule, and what gets logged. None of it is settled after a reading.
Secondary outcomes stay secondary. A comparison is not re-run until it wins, and the sample is not extended after the outcome is seen. Departures from the protocol are published with the result.