Amherst, Mass. (CBS CONNECTICUT) — Streaming online viewers have no tolerance for videos that take longer than just a few seconds to load onto their screens.

Video-streaming services – such as YouTube or NetFlix – win and lose millions of viewers and customers in just a matter of seconds. And according to a new study from the University of Massachusetts, about half of the people who use a high-speed, fiber-optic connection believe that five seconds is too long to wait, National Public Radio reports.

YouTube alone averages 4 billion hours of streaming video each month, and that is only a fraction of the overall online-streaming community. And for a business that serves an average base of 800 million people a month, every single second counts, and the more people who click away, the larger the problem quickly becomes.

Many users won’t wait even a couple of seconds before navigating away to another web source.

“What we found was that people are pretty patient for up to two seconds,” Ramesh Sitaraman, science professor at the University of Massachusetts, told NPR News. “If you start out with, say, 100 users — if the video hasn’t started in five seconds, about one-quarter of those viewers are gone, and if the video doesn’t start in 10 seconds, almost half of those viewers are gone.”

Sitaraman and his colleagues’ research revealed viewers begin to abandon video if it does not start up within two seconds. Each additional second of delay resulted in a 5.8 percent increase in the abandonment rate.

The research also found that viewers are more forgiving when slow starts are associated with longer content like movies than they are with short videos like news clips. Viewers with better connectivity are less patient and abandon slow-starting video faster than viewers using mobile devices.

The study looked at nearly 6.7 million viewers who watched almost 23 million videos play – about 216 million minutes of video playing time.

To conduct the research, Akamai provided the researchers access to millions of anonymous traces containing information regarding the way users watch videos and the quality they experienced.

Viewers were then assigned to two groups – one that received a good-quality video viewing experience and the other that experienced poor-quality service, with problems like failure to start and repeated freezes. A member of the first group was paired with the second in a way that matched as closely as possible other characteristics, such as geography and connection type, so that the differences between the two centered on quality of experience.

“This is really the first large-scale study of its kind that tries to relate video-streaming quality to viewer behavior,” Sitaraman told NPR.

Sitaraman worked together with S. Shunmuga Krishnan at the Association for Computing Machinery (ACM) Internet Measurement Conference in a talk titled, “Video Stream Quality Impacts Viewer Behavior: Inferring Causality using Quasi-Experimental Designs.”