Most ISPs use a five minute sampling and 95% utilization when calculating usage.
The 95th percentile is a widely used mathematical calculation to evaluate the regular and sustained utilization of a network connection. The 95th percentile method more closely reflects the needed capacity of the link in question than tracking by other methods such as mean or maximum rate. The bytes that make up the packets themselves do not actually cost money, but the link and the infrastructure on either end of the link cost money to set up and support. As a result, 95th percentile is commonly used among all major internet transit and peering networks, as well as datacenters and ISPs for both capacity planning and/or calculating usage.
Since most networks are oversubscribed, there is often some room for some bursting without advanced planning (hence "Burstable billing"). Ignoring the top 5% of the samples is a reasonable compromise in most cases (hence 95th percentile).
Many sites have the majority of their traffic on Mondays, so the Monday traffic determines the rate for the whole month. Some providers offer billing on the 90th percentile as an incentive to attract customers with irregular bandwidth patterns.
The 95th percentile is allows a customer to have a short (less than 36 hours total for the month) burst in traffic without overage charges. Basically the 95th percentile says that 95% of the time, the usage is at or below this amount. Conversely, 5% of the samples are bursting above this rate.
There are important factors to percentile calculation:
Sampling interval, or how often samples are taken (called also "data points").
Another option is “average usage billing”. While most datacenters and hosting providers use the "95th percentile" or “burstable billing”, others may use average usage billing. This is easier to implement since implementations can function similar to gas or electric utility meter readings (however, they are usually sampled more often to avoid losing data if systems lose power or reboot).
It could be argued that average use billing is more fair to customers, since the real cost to ISPs is somewhere between average use and 95th percentile due to oversubscription. Average use ignores the natural peaks and valleys that occur from day to day bandwidth usage. Here is an example:
Bandwidth is measured (or sampled) from the switch or router and recorded in a log file. In most cases this is done every 5 minutes. At the end of the month, the samples are sorted from highest to lowest, and the top 5% (which equal to approximately 36 hours of a 30-day billing cycle) of data is thrown away. The next highest measurement becomes the 'billable utilization' for the month.
Based on this model, the top 36 hours (top 5% of 720 hours) of peak traffic is not taken into account when billed for an entire month. Bandwidth could be utilized at a higher rate for up to 65 min a day with no financial penalty.