Marques Brownlee has spent fifteen years being careful about how he criticizes YouTube. He makes his living there. So when he calls something the platform’s “craziest, riskiest, and possibly worst idea yet,” it is worth stopping to ask what he actually saw.
The feature is video A/B testing. Starting in 2027, YouTube will let creators upload up to three different cuts of the same video and serve them to different viewers at the same time, then report back on which one held attention longest.
On paper it is a logical extension of something YouTube already does. In practice, Brownlee argues, it quietly breaks the one assumption that holds a comment section together: that everyone watching is watching the same thing.
The short version
- Up to three versions of a single video or Short can run simultaneously, each a genuinely different edit
- Announced at Made on YouTube in September 2026, arriving for select creators in 2027
- Viewers are not told which version they were served, and there is no public signal that a test is running
- Brownlee’s core objection is not the feature, it is the comment section. Timestamps and references may point at footage the next viewer never saw
- YouTube screens entries with a semantic analysis model that scores how similar the versions are. Too different and they do not qualify
- New uploads only. Creators cannot retroactively test an old video
- It builds on 2024’s title and thumbnail testing, which was uncontroversial because the video itself never changed
What YouTube actually announced
At its annual Made on YouTube event in September 2026, the company laid out a slate of creator tools: dynamic thumbnails, live dubbing, and the one everybody fixated on, A/B testing for full videos.
The mechanics are straightforward. A creator uploads two or three edits of the same piece. YouTube distributes them across the audience, measures watch time share on each, and produces a report naming the winner. The examples YouTube showed were not subtle tweaks either. The runtimes in its own demonstration differed by several minutes, which tells you this is meant for restructured videos, not a swapped opening line.
It works for long form and for Shorts. And it is, by any normal product standard, a reasonable thing to build. Creators have been guessing at hooks for two decades. Here is data instead of instinct.
Why 2024 was fine and 2027 is not
YouTube has run A/B tests since 2024, when it gave creators the ability to test multiple titles and thumbnails against each other. Almost nobody objected, and the reason is worth spelling out.
A thumbnail is packaging. Two people who click different thumbnails still land on the same video, see the same frames, and can argue about the same moment at 4:12. The shared object survives.
Change the video itself and that stops being true.
| What is tested | Launched | Viewers see | Comments still line up |
|---|---|---|---|
| Titles | 2024 | The same video | Yes |
| Thumbnails | 2024 | The same video | Yes |
| Full video cuts | 2027 | Up to 3 different edits | Not necessarily |
| Shorts cuts | 2027 | Up to 3 different edits | Not necessarily |
That last column is the whole argument. Everything else is implementation detail.
Brownlee’s actual objection
It helps to read what he said rather than the headline version of it, because he is not making a moral argument about manipulation. He is making a structural one about shared experience.
“As a viewer, if I’m watching a video, do I know if that video is part of an A/B test or not? Probably not,” he said. Then the follow up, which is the part that lands: “If I’m reading the comments on a video, how do I know which version of the upload they watched? And if someone’s comment has a timestamp, does that break if I’m watching the other version of that video?”
Anyone who has used YouTube seriously knows how much work the comment section does. Corrections live there. Timestamps live there. The joke about the thing that happened at 7:40 lives there. It is the closest thing the platform has to a shared room.
Brownlee put it plainly: “There’s this sort of unspoken rule of community on YouTube that, like, we’re all having the same experience together. We’re all watching the same video. So letting people put up two or three different versions of the same video at the same time just kind of fundamentally breaks that.”
The guardrail, and how much it is worth
YouTube is not blind to this. Brownlee says that after talking to the company’s engineers, he learned the plan involves a semantic analysis model that scans each uploaded version and scores how similar they are to one another. Clear the similarity bar and the versions are eligible for testing. Fail it and they are not.
That is a real constraint, and it rules out the nightmare scenario where a creator uploads two unrelated videos under one title. It does not solve the timestamp problem at all. Two cuts can be ninety five percent identical by any semantic measure and still be ninety seconds apart in length, which is more than enough to make every timestamp in the comments wrong for half the audience.
The other limit is that this applies only to new uploads. Back catalogs are safe.
Brownlee’s description of how the conversation with YouTube’s engineers went is the part creators keep quoting. By his account, the answers to his harder questions amounted to a shrug.
Worth keeping straight. This is not about AI generating different versions of your video. The creator still makes every cut by hand and chooses to enter them. The AI involvement is a similarity checker that decides whether the versions are close enough to count as the same video. Those two things are getting mixed up constantly in the reaction to this.
What creators actually gain
It is easy to read the criticism and miss that this is a genuinely valuable tool for the people it is aimed at.
- The first thirty seconds stop being a guess. Retention cliffs at the opening are the single biggest killer of a video’s reach, and until now creators have been diagnosing them after the fact
- Structural questions get answers. Does the sponsor read belong at the front or at the six minute mark? There has never been a clean way to test that on the same audience
- Length becomes testable. The eternal argument about whether a tighter edit beats a fuller one can finally be settled per channel rather than per opinion
- It levels the field slightly. Large channels already approximate this by posting variations across different accounts. Small channels never could
None of that is nothing. Brownlee himself was positive about most of YouTube’s other announcements, including its wider adoption of AI tooling. His complaint is specific.
The measurement problem underneath all of this
There is a pattern here that goes beyond one feature. YouTube keeps changing what it measures, and each change ripples outward in ways that are hard to see at announcement time.
The clearest recent example was the platform’s decision to redefine what counts as a view. We covered that in detail when YouTube changed how views are counted and every channel suddenly looked bigger, and the lesson there was that a metric definition is never just a metric definition. It changes what creators make.
Optimization pressure does the same thing. If the system tells you with hard numbers that the version with the louder hook wins, you will make the louder hook. Do that across millions of channels for a few years and you have reshaped the medium, one measured increment at a time. YouTube has done versions of this before, from monetization thresholds to the moment it rewrote its gameplay violence rules ahead of GTA 6 and changed what a whole category of channel could show.
The counterargument, which is fair, is that creators already optimize relentlessly. They just do it blind. There is something to be said for letting them see the dashboard rather than making them read tea leaves, and for what it is worth, the people who spend money chasing the algorithm rarely get what they paid for. Our look at what $118,000 of YouTube ads actually bought one channel is a useful reminder that the platform’s own tools beat brute force almost every time.
What would fix it
The objection has an obvious remedy, and it is strange that YouTube has not announced it already.
- Label the test. A small marker on the player saying this video has multiple versions costs nothing and ends the transparency complaint immediately
- Tag comments by version. If the platform knows which cut a commenter watched, it can show that, or at least group comments so timestamps stay meaningful
- Cap the length delta. Similarity scoring should include runtime, not just content. Two cuts within a few seconds of each other keep timestamps usable
- Publish the winner. Once a test concludes, consolidate everyone onto the winning cut so the video has a single canonical form going forward
None of that is technically hard. All of it depends on YouTube deciding that comment coherence is a feature worth protecting rather than an acceptable casualty.
The bottom line
Video A/B testing is a good tool with one badly thought through side effect, and Brownlee is right to name it before the feature ships rather than after. The creator benefit is real and measurable. The cost falls on viewers, who were not asked, will not be told, and will mostly never notice that the thing they are discussing is not quite the thing somebody else watched.
YouTube has until 2027 to add a label and fix the timestamps. If it does, this will be remembered as a useful upgrade. If it does not, the comment section will slowly get stranger, and almost nobody will be able to explain why.
Sources and further reading
- UNILAD Tech: Marques Brownlee slams YouTube’s latest feature as possibly their worst idea yet
- Dexerto: MKBHD calls YouTube’s new A/B testing feature its craziest, riskiest idea yet
- Dexerto: MKBHD raises concerns over YouTube’s new video A/B testing feature
- TechCrunch: YouTube adds new creator tools including video A/B testing, dynamic thumbnails and live dubbing
- 80.lv: YouTube adds A/B testing for video cuts, expanding Studio’s toolkit
- TechCrunch: YouTube creators can test multiple video thumbnails (2024)
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
