Do captions actually increase watch time?
Updated
Probably yes for feed video, and the numbers everybody quotes are weaker than they look. Almost every statistic in circulation traces back to a platform or a caption vendor, measured on their own inventory, and the most famous of them was a trade-press report rather than published research. The mechanism holds up better than the statistics do: a muted video autoplaying in a feed gives a scrolling viewer nothing to follow unless there are words on screen.
Where the numbers come from
If you search this question you will be told that 85 percent of feed video is watched without sound. That figure comes from a 2016 trade-press report of what a handful of publishers were seeing on Facebook. It was never a published platform statistic, the platform's own figures at the time differed, and it has been repeated for a decade without anybody re-measuring it.
The completion-rate claims have a similar shape. They are typically produced by companies that sell captioning, or by platforms encouraging people to use a feature they have just shipped. That does not make them false. It does mean nobody independent has checked them, the underlying data is not published, and the conditions they were measured under are not the conditions you are posting into.
This matters more than it sounds, because the advice built on those numbers is usually right for reasons the numbers themselves do not establish. The conclusion is worth believing on the strength of the mechanism instead.
The mechanism is the convincing part
Feed video autoplays muted. Every major feed has made that product decision, which means the default encounter with your video is silent and lasts about a second before the viewer decides.
In that second, an uncaptioned talking-head video is a person moving their mouth. There is no content yet. A captioned one has already said something, and the viewer can tell whether it is for them. That is the whole effect, and it does not need a study to be plausible.
Three smaller effects point the same way. People watch in places where sound is not an option. Viewers who are not native speakers follow written words more easily than spoken ones. And accents, fast speech and poor recording audio all stop being obstacles when the words are on screen.
Where captions do not help, and where they hurt
On video somebody deliberately chose to watch with sound on, the effect mostly disappears. A tutorial opened from a search result, a podcast episode playing in the background, a lecture recording: the viewer has committed and the audio is already on. Captions are still worth having for accessibility, but not as a retention tactic.
They actively hurt in two situations. Captions that are too long or move too fast are worse than none, because the viewer is now failing to read something rather than comfortably ignoring it. And captions placed over the part of the frame carrying the information cover the thing the video was made to show.
The third case is specific to platforms that generate their own captions. Burn captions into the picture, leave the platform's automatic ones switched on, and the viewer gets both sets at once, offset by a line. Pick one.
How to find out for your own account
Published averages are averages over somebody else's audience. The only figure that applies to you comes from your own posts, and it is cheap to get.
Post the same kind of content captioned and uncaptioned over a few weeks, roughly alternating, at similar times of day. Then compare retention in the first three seconds rather than total views, because that is where the mechanism operates, and total views are dominated by whether the feed picked the post up at all.
Be honest about the noise. A handful of posts tells you nothing: distribution varies more between two identical posts than captions are likely to move it. If you cannot run twenty or thirty, treat the result as a hint rather than an answer and fall back on the mechanism. Muted autoplay is real, and words on screen are the only way to speak into it.
What it costs to find out
Burned-in captions require re-encoding the video, which takes time and produces a new file. A separate .srt costs nothing to make but is not accepted by the short-form feeds at all, which take a video file and nothing alongside it.
That asymmetry is why the test above is worth running before captioning becomes part of every upload. If it turns out to do nothing for your audience, you have saved a step on every video from then on. If it does what the mechanism suggests, you will know by how much, rather than by how much a vendor says.