Online research has experienced remarkable growth over the  past fifteen years. To keep up with demand,some companies have become quite creative. Rather  than continuing to rely exclusively on opt-in panelists, for example, they've developed new  methods to include non-panelists within online surveys. They've also figured out  how  to direct or route respondents who  do not qualify for one  survey  to another for which they  might. Despite these  important advances, the available supply of respondents to which any supplier might have access is insufficient at times for certain kinds  of studies,such as a tracking survey  of a rare population. To meet the requirements of such studies, researchers now  depend heavily on multiple samples sources  (e.g.,Panel A,B; River 1). Some evidence suggests,however, that the decision can have unintended consequences. In research carried out in 2008  evaluating seventeen different opt-in panels, for instance,the Advertising Research Foundation found "wide variance,particularly on attitudinal and/or opinion questions (purchase intent,concept reaction, and the  like)," even after holding constant socio­ demographic and other factors (Walker et al., 2009). Since that time,some researchers have mounted new research to understand how  to select multiple sample sources  for the  same survey  without increasing bias. Proponents of these  approaches cite  at least three benefits: (a) consistency (or interchangeability) of new respondent sources  with existing ones,(b) complementariness of new  respondent sources  with existing ones relative to an external standard, and (c) enhanced representativeness relative to the  US general population through calibration with non-online data sources. Although these  approaches have taken  a step in the right direction, we believe they have not  gone  far enough for three main  reasons: (a) they  restrict the  pool of potential respondents to those from sample  sources  vetted previously, thereby limiting supply, (b) they  seem to assume  that  the vetted sample sources  do not change over  time,and (c) they rely on benchmark data sets
that have either limited shelf lives or uncertain external validity. We therefore suspect that they  may not  produce the same levels of sample  representativeness and response accuracy as a new  methodology, which we refer  to as SmartSelect, that  selects  potential survey  respondents in real-time from  either a single sample source  or multiple sources  based on how  well  their characteristics match an appropriate, evolving standard with demonstrated evidence of external validity.