Key takeaways
- YouTube video A/B testing was announced at Made on YouTube on 23 September 2026. YouTube's own wording is that "coming soon" you will be able to test up to three video cuts to see which hook performs best. It is not available yet, and no eligibility rules have been published.
- The object being tested changes, and that is the whole story. Test & Compare swaps the packaging in front of a fixed video. Cut testing swaps the video itself while the packaging stays fixed.
- YouTube has framed the decision around the hook, not the whole edit — which is also the only part of a cut that a retention-decided test can judge quickly and cleanly.
- The cheap, sane way to build a second cut is to change only the first 45 seconds and keep every frame after it identical. That halves the work and removes the confounds.
- The real costs are operational: comments, timestamps, chapters, sponsor reads, captions, dubs and your retention graph all assume one canonical version of a video.
- Marques Brownlee's objection — that a viewer cannot tell they are in a test, and a commenter's timestamp may point at a cut you never saw — is the strongest argument against the feature, and it is an argument about disclosure, not about testing.
For twenty years, one thing on YouTube has been fixed: the video. Titles changed, thumbnails changed, descriptions were rewritten, tags came and went, and through all of it the file people watched was the file you uploaded. Every testing tool YouTube has built sat in front of that fixed object, tuning the promise made about it. The video was the thing the promise was about.
At Made on YouTube on 23 September 2026, YouTube said that assumption is going too. Among a long list of Studio features, its announcement post contains one sentence that matters here: "coming soon, you'll also be able to test up to three video cuts to see which hook performs best with the new video A/B testing feature." Three edits of the same video, served to real viewers, with the platform deciding which one wins.
Almost everything written about it since has been a rewrite of that sentence plus a reaction quote. What follows is the part that lands on your week: what a retention-decided test can and cannot measure, which edit decisions are worth two cuts and which are a waste of a render, what breaks the moment a video stops having one canonical version, and what to do before the feature arrives.
What YouTube has actually said about video A/B testing
Start by separating the announced from the assumed, because the gap is wide and most coverage does not mark it.
| Claim | Status |
|---|---|
| Creators will be able to test up to three cuts of one video | Stated in YouTube's own announcement |
| The stated purpose is finding which opening hook performs best | Stated in YouTube's own announcement |
| It is described as "coming soon" and is not live | Stated in YouTube's own announcement |
| It builds on title and thumbnail testing, which YouTube says has been used for more than 40 million experiments since 2024 | Stated in YouTube's own announcement |
| Cut testing for Shorts arrives from 2027 | Reported by TechCrunch, not confirmed by YouTube |
| The winner is decided on retention or watch time | Implied by "which hook performs best" and by how every existing YouTube test resolves; the exact metric has not been published |
| A similarity check will stop unrelated videos being entered as "cuts" | Reported by a single outlet; not in YouTube's materials |
| Eligibility, markets, test duration, minimum traffic | Not stated anywhere |
| What happens to comments, timestamps and chapters across cuts | Not stated anywhere |
| Whether viewers are told they are in a test | Not stated anywhere |
Two rows deserve emphasis. The first is "coming soon": there is no Help Center page for cut testing, which means anyone publishing exact eligibility thresholds or test windows for it is extrapolating. The second is the metric. On the packaging side, YouTube's A/B testing documentation is explicit that the variant with the highest watch time is the one shown to everyone afterwards — not the one with the best click-through rate. If cut testing inherits that logic, it is a well-behaved feature. If it resolves on something narrower, like 30-second retention alone, it becomes a machine for manufacturing openings that hold people for thirty seconds and then lose them.
Why "which hook performs best" is a narrower promise than it looks
Read YouTube's sentence again and notice what it does not say. It does not say test two edits. It says test cuts "to see which hook performs best" — and that framing is not shorthand, it is the only honest version of the feature. The reason is statistical.
A hook difference shows up early, in a part of the video every single viewer sees, and it produces a large effect. Two openings that differ — one that starts on the result, one that spends forty seconds setting up why the result matters — will separate on the retention curve inside the first minute, and they will separate for everybody who pressed play. A difference you make at eleven minutes is seen by whoever is still watching at eleven minutes, which on most channels is a minority of a minority. The same edit decision, moved later in the video, costs many times more traffic to evaluate, because the population that can register it keeps shrinking.
This is the same arithmetic that makes thumbnail tests slow, and we worked it through in the thumbnail A/B testing guide. One difference cuts the other way: a packaging test measures a binary decision taken by everybody shown the video, while a cut test measures a continuous outcome — minutes watched — among the far smaller group who actually started it. Fewer data points, noisier measurement per point. No power numbers have been published for cut tests, but the direction is safe: expect them to need more patience than thumbnail tests, and expect small edit tweaks to come back inconclusive.
The one-variable rule matters more here than anywhere
YouTube's own guidance on packaging tests notes that variants which are too similar to each other can make a test run longer, because there is not enough difference to decide a winner. Cut testing inverts the trap: it is easy to make two cuts that differ in ten ways at once — a different open, a different music bed, a tighter middle, the sponsor read moved — and then discover you have learned nothing transferable. If the two openings differ, change nothing else.
How a cut test differs from a packaging test, precisely
Superficially these are the same feature: three options, one video, YouTube serves them and names a winner. Underneath, almost every property is different.
| Test & Compare (titles and thumbnails) | Video A/B testing (cuts) | |
|---|---|---|
| What varies | The promise in front of the video | The video itself |
| What stays fixed | The video | The title, thumbnail and description |
| Who registers the difference | Everyone shown an impression | Only people who pressed play |
| Decision metric | Highest watch time, per YouTube's documentation | Not published; framed around hook performance |
| Availability | Live now, desktop Studio only, not available for Shorts | Announced, not released; Shorts reported for 2027 |
| Cost of a variant | One image or one line of text | An edit, a render, an upload, and a second review pass |
| Reversibility | Total — swap the image back | Partial — the losing cuts were watched, commented on and timestamped |
| What you learn | Which promise your audience responds to | Which opening structure holds your audience |
| Who can be harmed by a bad variant | Nobody; a weak thumbnail is ignored | Viewers, who spent real minutes on your worse edit |
That last row is the ethical centre of the feature and it gets almost no attention. A losing thumbnail costs a viewer nothing. A losing cut costs them the eight minutes they gave to the version you later decided was worse. Packaging tests spend impressions; cut tests spend other people's attention. That raises the bar for what is worth testing.
What is actually worth two cuts
The feature is only useful if you bring it a real disagreement. Here are the edit decisions that are big enough to separate on a retention curve, and what each one is really asking.
| Decision | Cut A | Cut B | The question underneath |
|---|---|---|---|
| Cold open vs context | Open on the most dramatic moment, explain later | Open by establishing why this matters, then deliver | Does your audience arrive curious or confused? |
| Payoff position | The answer at 0:20, then the method | The method, then the answer | Are you making a reference video or a story? |
| Branded intro | Keep the three-second animation | Cut it entirely | Is the identity worth the drop-off? |
| Hook length | A tight 15-second setup | A 60-second setup with stakes | Does your topic need patience bought up front? |
| Opening voice | On camera, talking to the viewer | Voiceover over the thing being discussed | Are people here for you or for the subject? |
| Promise restatement | Restate the title claim in the first ten seconds | Trust the title and start | How much of your traffic arrived from search, not the feed? |
Every one of those is a structural question with a different answer per channel, which is exactly what a test is for. What is not worth a cut: colour grades, transitions, a different music track, a joke swapped for another joke. Those matter cumulatively, but no realistic amount of traffic will separate them on retention.
If you want to know which of your own openings is weakest before you pick a question to test, the answer is already in Studio. The first thirty seconds of the retention curve is the most information-dense graph you have — we covered how to read it in the audience retention graph guide, and the structural fixes it tends to point at are in the piece on hooks and the first thirty seconds.
The production problem, and the trick that solves most of it
Three cuts of a ten-minute video sounds like three times the work. In practice it is worse than that, because the invisible costs multiply: a second export, a second upload, a second review pass, captions for both, and — if you localise — dubs for both.
The way out is to make the cuts share almost all of their frames. Build cut B so that it is identical to cut A from a fixed point onwards, and vary only what comes before. A different cold open and then a seam into the same body at 0:45 is a real test of a real question, and it costs one new opening rather than one new video. It also, conveniently, removes every confound: if everything after the seam is frame-identical, any difference in retention beyond that point is caused by the opening, by audience composition, or by chance — not by a second edit decision you forgot you made.
Three practical consequences follow from working this way, and they are worth building into how you edit rather than bolting on later.
- Shoot the alternative opening on the day. The expensive part of a second cut is rarely the timeline; it is discovering in the edit that you have no footage for a different open. Two takes of an intro cost ten minutes at the shoot and nothing afterwards.
- Keep the seam clean. Pick a point where the body of the video can start from either opening without a continuity break: a hard cut to a new location, a title card, the first step of the process. Design that moment deliberately, and your second cut becomes a swap rather than a re-edit.
- Decide in advance which cut is the canonical one. Whatever the test decides, one version is the one that goes in the description timestamps, the chapters and the clip exports. Not deciding leaves you reconciling two versions after the fact.
What a split video quietly breaks
This is the part the launch coverage skipped, and it is where most of the actual friction will live. Nearly everything around a YouTube video assumes there is one of it.
Comments and timestamps. A commenter writing "the bit at 4:20 killed me" is pointing at a moment in the cut they were served. Anybody watching a different cut sees either a different moment or nothing. Marques Brownlee put this better than anyone, asking on his own post about the announcement: "If I'm reading comments on a video, how do I know which version they watched?" There is no answer to that question yet.
Chapters and description timestamps. Chapters live in the description, which is shared across cuts. An opening that runs forty seconds longer shifts every later chapter by forty seconds, so one of your two audiences has chapters pointing at the wrong thing. That is a reason to keep your cuts close in total length.
Sponsor obligations. Brand deals specify placement: a read within the first ninety seconds, a set duration, approved wording. If the read lands at 0:30 for half your audience and 1:20 for the other half, you have shipped two deliverables against one contract. Keep the sponsor read out of the part that varies, and get both cuts approved when that is unavoidable.
Captions, dubs and localisation. Every cut needs its own caption track, and if you use multi-language audio, its own dubs. The cost of a second cut is not one render; it is one render multiplied by however many languages you publish in.
Your own analytics. A retention curve averaged across three cuts is a curve of a video nobody watched. Unless Studio reports per-cut retention — and nothing says it will — the graph you use to diagnose pacing becomes a blend for the length of the test, and so does average view duration.
Copyright and claims. If each cut is uploaded and scanned as its own file, as seems likely, one video can come back with two different claim outcomes — our guide to claims versus strikes covers why that distinction matters.
The transparency objection, taken seriously
The loudest creator response came from Marques Brownlee, who called it, as Dexerto reported, "their craziest, riskiest, and possibly worst idea yet". His case has two parts. The first is that a viewer has no way of knowing they are in an experiment. The second is that running several versions at once dissolves the idea that an audience watched the same thing — which is what makes a comment section a conversation rather than a pile of unrelated reactions.
The counter-argument is not nothing. Viewers already receive a personalised homepage, personalised recommendations and, under dynamic thumbnails, personalised packaging. Creators already cut the same material differently for different platforms, and nobody calls that deceptive. Each cut is still made by hand and entered deliberately.
Both things can be true, and the resolution is that this is a disclosure problem rather than a testing problem. Three rules keep you on the right side of it, and all three are available to you without waiting for YouTube to build anything:
- Test structure, never substance. Two orderings of the same material is a fair experiment. Two different claims, two different conclusions, or a disclosure present in one cut and absent in the other is not a test; it is two videos sharing a title.
- Say so. A line in the description or a pinned comment — this video was published in two openings while we worked out which one landed better — costs nothing and removes the entire objection about viewers not knowing.
- Keep anything a viewer might cite stable. Numbers, sources, corrections, sponsor disclosures and the timestamps of key moments should be identical across cuts, so that a comment about them is true for everybody.
Where this fits in the direction YouTube is travelling
Cut testing is not a one-off. It is the third step in a pattern that has been visible for two years. Step one was Test & Compare: you supply three thumbnails, YouTube finds the single best one and shows it to everyone. Step two was dynamic thumbnails, announced at the same 2026 event: you supply three images and YouTube stops looking for one winner, instead allocating different images to different segments of your audience. YouTube's discovery lead, Todd Beaupré, described that system to NetInfluencer as starting with some randomness in order to learn, then shifting traffic toward whatever works for each group — with segments like where a viewer is and how much of your channel they have already watched.
Step three extends the same logic past packaging to the artefact. In each step, the creator's job moves one notch away from deciding and one notch toward defining a range the system is allowed to choose from. The rest of what was announced that day, which we went through in the Made on YouTube 2026 round-up, points the same direction.
There is a reasonable worry buried in that trajectory. A system that reliably rewards whichever opening holds attention will, over thousands of iterations, select for openings that hold attention — regardless of whether that is the opening the video deserved. That pressure already exists; cut testing makes it measurable, which makes it both harder to ignore and easier to overdo.
What to do in the meantime
There is no date, so the useful posture is preparation. Everything here pays off whether or not the feature arrives on schedule.
- Audit the first thirty seconds of your last ten uploads. Write down the shape of each drop-off and what the opening was doing at that moment. You are looking for the one structural habit that costs you viewers on every video — that is your first test when the feature lands.
- Use the draft feedback tool that already exists. Ask Studio can take an unpublished video and comment on pacing and structure; YouTube's documentation says you can select draft, private or unlisted videos and get suggestions based on creative best practices and historical retention data. It is not a test with real viewers, but it is free and it runs before you publish, which is the one thing a cut test cannot do. We covered the wider toolkit in the Ask Studio guide.
- Probe hooks with Shorts. A Short is a hook with nothing attached, so publishing two framings of the same idea is a crude but real read on which one people stay for — minutes of work rather than renders.
- Shoot two openings as standard. Add it to your shoot checklist now, so that when cut testing arrives you already have a library of alternative opens rather than a list of regrets.
- Design a seam into every edit. One clean point, around the forty-five-second mark, where the body could begin from either opening.
- Keep running packaging tests. Test & Compare is live, free and needs advanced features rather than a subscriber threshold. It is also the cheaper experiment: the packaging decides whether anyone presses play at all, and no cut can rescue a video nobody opens. Our full guide to Test & Compare has the mechanics.
- Write down your disclosure policy before you need it. One line, decided calmly now, beats an improvised answer in a comment thread later.
The scepticism this deserves
Three caveats, because the distance between an announcement and a shipped feature is where most creator advice goes wrong.
The feature is announced, not documented. The version that ships may be limited to a subset of channels, may resolve on a metric nobody has guessed, and may arrive without per-cut reporting — which would be the most important thing in it, because a test you cannot read is a coin flip with extra steps.
The timing is unsettled. TechCrunch reported cut testing for Shorts from 2027, while YouTube's own post says only "coming soon" and does not split the feature by format. That is doubly odd given that YouTube's current A/B testing documentation states plainly that testing is not available for Shorts at all. Treat every date you read about this feature, including the ones above, as reporting rather than fact.
And the order of operations has not changed. Cut testing improves a video that people are already choosing to watch. It does not fix a premise nobody wants, a title that promises nothing, or a thumbnail that is illegible at feed size. Packaging decides whether the test gets any traffic to run on; the cut decides what happens next.
Frequently asked questions
Can you A/B test a whole YouTube video right now?
No. YouTube's testing tools currently cover titles and thumbnails only — up to three variants, resolved on watch time, in desktop Studio, and not available for Shorts. Cut testing was announced in September 2026 as coming soon, with no Help Center page yet.
How does YouTube decide which cut wins?
YouTube has not published the metric. Its framing — which hook performs best — and the precedent of every existing test resolving on watch time both point at a retention-based decision rather than clicks. That is the right design, because a cut that wins attention for ten seconds and loses it at two minutes should not be able to win a test.
Will viewers know they are watching a test version?
Nothing has been announced about labelling, and the current packaging tests are invisible to viewers. That is the substance of the criticism the feature has attracted, and until YouTube says otherwise, the only transparency available is the kind you add yourself in a description or a pinned comment.
Does video A/B testing work on Shorts?
Not yet, and the reporting conflicts with the current rules. TechCrunch tied Shorts cut testing to 2027; YouTube's existing documentation says A/B testing is not available for Shorts, and that a video converted to a Short loses access to any tests it had. Until there is a Help Center page, assume long-form first.
How do I make three cuts without tripling my edit time?
Vary only the opening. Build every cut to converge on the same timeline at a fixed seam — around forty-five seconds in — so a second cut is a new open plus a swap rather than a second edit. Shoot the alternative opening on the day, keep the total runtimes close so your chapters stay honest, and never change more than one thing between cuts.
Is this the same as dynamic thumbnails?
No, and the difference is instructive. Dynamic thumbnails never resolve: different viewers keep seeing different images, permanently. A cut test is meant to end with a winner, which means it gives you something to keep — a fact about your audience's patience that transfers to the next video.
The bottom line
Video A/B testing is the first YouTube feature to treat the video as a variable. That is a bigger change in principle than in practice: the useful version of it is narrow, aimed at the opening, decided by whether people stay, and most valuable to creators who already know which structural habit is costing them viewers.
The work that pays off before it ships is unglamorous and entirely within your control. Read your own retention curves honestly. Shoot a second opening. Put a seam in your edit. Keep testing the packaging, because the packaging is still what decides whether a cut test has any traffic to run on — and that part you can act on today with Test & Compare and the free A/B comparison tool, which puts two candidates side by side at the size viewers actually see.
If the obstacle is producing the options in the first place — three thumbnails for a test, a second framing for a video you are about to cut two ways — that is a production problem with a production answer. Thumblore generates thumbnails in your channel's style from a description of the video, which turns a second and third option into a few minutes of work rather than an evening. The habit of supplying a range, rather than committing to one answer, is the thing every feature YouTube announced in 2026 is quietly asking for.