Rarely, if ever, are we able to get responses from all of the people that we would like to survey.  Usually, we only get  information  from a subset, or sample, of the people or organizations that we are interested in.  One of the most important goals of survey research is ensuring that the sample we get information from is representative of the group (the population) that we wish to make statements about.  If the sample is large enough and if it represents the population well, then we can make accurate statements about the opinions, preferences or intentions of the group as a whole or of subgroups that interest us -- even when we are not able to get information from everyone. 

When the group that we are interested in is not unmanageably large and we are able to identify all members of the group, then we can approach sampling from the perspective of trying to get information from everyone in the group.  We refer to this as a census sample.  Often, surveying all of your employees, your business-to-business customers, your dealers, wholesalers, or other interested parties is manageable using Web-based surveying.

Random Subsets
At other times and depending on the survey goals, we may not need to or be able to survey everyone we are interested in.  In these cases, we draw a random subset or subsets of the population and invite them to participate in the survey.  Randomness is important in this context - all of the statistics that we use are based on the assumption that the sample we have obtained is a random subset of the population that we are interested in.

Sophisticated sampling procedures have been developed to help us sample from within subgroups of the population that we are interested in.  There are also procedures we can use to make adjustments in the data we receive to bring it into line with what we know about the characteristics of the original population we sampled from (sample weighting) and there are statistical techniques that we can use to estimate the accuracy of the findings we obtain.   These include computing margins of error or confidence intervals.