"What percentage of YouTube videos get their thumbnail changed?" is a good question, it has a real answer, and you can find a dozen confident numbers for it online. Most of them have no stated sample, no stated window and no stated method, which means they cannot be checked and should not be repeated.
We are in a position to answer it properly, because answering it is a by-product of what this product does: it watches a set of channels continuously and records every packaging change with a timestamp. What we do not yet have is a sample large enough to publish from.
So this page is the method, written down before the result exists. When the corpus clears the bar below, the numbers get added here and this paragraph gets replaced with them.
Why pre-register the method
Because the alternative is what everyone else does: run the query, look at the output, then decide which definition of "changed" produces the most interesting number.
That is not fraud, it is just how analysis drifts when nobody has written the rules down first. Publishing the definitions in advance costs nothing and makes the eventual result checkable — including by us, later, when we have forgotten what we intended.
The bar for publishing
Two conditions, both of which must hold:
- At least 50 distinct channels in the observation set.
- At least 90 days of continuous observation.
Fifty channels because below that a single prolific channel dominates the denominator, and because packaging behaviour varies enormously by niche — a sample from three niches is a statement about those three niches. Ninety days because the interesting denominator is videos observed from their first hours, and a shorter window makes that population too young to have finished changing.
As of the date at the top of this page, the corpus does not clear both bars. There is therefore no percentage here. That is the entire reason the page exists in this form: the alternative was to publish a number from a sample we know to be too small, and this product's position is that a null with a reason beats a number without one.
The definitions, fixed in advance
A video enters the population when it is first observed at an age of six hours or less. Older videos are excluded, and this is the strictest and most important rule: a video first seen at three weeks old may have been retitled twice before we ever looked, and counting it as "unchanged" would deflate every rate we report. The population is therefore "videos we watched from the beginning", not "videos we know about".
A title change is a difference in the exact title string between two consecutive observations of the same video, excluding differences attributable to language localisation.
A thumbnail change is a difference between two consecutive observations whose perceptual hashes are more than a Hamming distance of 6 apart. That threshold is the same one used everywhere else in this system and is frozen, so this statistic and the product's own change record cannot disagree.
An A/B test is not a change. Variants served simultaneously to different observers are a separate event class, counted separately. A test that ends by settling on a variant that differs from the pre-test packaging counts once as a test and once as a change, and the report will say so explicitly rather than quietly picking one.
The windows are 7 days and 30 days from the video's first observation. The reported quantity for each is: the share of the population with at least one title change, the share with at least one thumbnail change, and the share with at least one of either.
Time-of-day is excluded entirely. It cannot be derived honestly for most uploads, for reasons set out in the timing guide. Any day-of-week breakdown, if published, will be restricted to the same watched-from-the-start population and will say so on the chart.
What the result will and will not support
When it exists, the number will support statements of the form:
Of N videos observed from within six hours of publication across M channels between date A and date B, X% had at least one thumbnail change within 7 days.
It will not support:
- "Creators should change their thumbnails." A rate is not a recommendation. If a large share of videos get changed, that tells you the practice is common, not that it works — we cannot see whether any individual change helped, and neither can anyone else outside the channel.
- A claim about YouTube as a whole. The sample is the channels this product tracks, which is not a random sample of the platform. It skews toward channels someone thought worth watching, which almost certainly means larger and more deliberate than average. That bias will be stated beside the number, every time.
- A trend over time, until the window is long enough that two comparable periods exist inside it.
What happens if the answer is boring
It might be. The result could turn out to be that most videos are never repackaged at all, and that changes cluster on a small number of highly deliberate channels. That would make a far less quotable headline than "X% of creators change their thumbnails".
It gets published either way, and this paragraph exists so that promise is on the record before the number is. The failure mode of published research is not usually a fabricated figure; it is a real figure from the one cut of the data that happened to be interesting, with the boring cuts left in a drawer. Fixing the population, the definitions and the windows in advance is what makes that harder to do accidentally.
The same applies to the negative results. If the 7-day and 30-day figures turn out to be nearly identical — that is, almost all repackaging happens in the first week — that is a genuinely useful finding and it will be reported with the same prominence as any other.
Why nobody else publishes this properly
Two reasons, and they are both structural rather than lazy.
First, the data has to be collected forward. There is no archive of previous thumbnails to query — no public API returns what a video used to look like — so a study of packaging changes can only be run by someone who was already watching before the changes happened. That rules out the usual approach of scraping a large sample once.
Second, the population rule is expensive. Restricting to videos observed from their first hours means continuously polling a set of channels, not sampling videos after the fact. It is exactly the cost that makes the number trustworthy, and exactly the cost most analyses skip — which is why so many published figures are computed over "all videos we could find", a population that silently under-counts every change that happened before the scrape.
The raw material, and what happens when the bar is cleared
When the corpus clears 50 channels and 90 days, three things will be added to this page: the numbers, the aggregate table they came from as a downloadable CSV, and the date range. The per-channel and per-video rows will not be published — the product's visibility model scopes observations to the person who chose to track the channel, and an aggregate is the only form of this that can be published without changing that.
Until then, the honest state of this question is: we are collecting, the method is fixed, and there is no number.
If you want the same record for your own niche while this fills up, the packaging archive is the same data, unaggregated, for the channels you choose to track — and what packaging means explains what to look for in it.