This paper addresses epistemological issues raised by the use of elicited data in linguistic analyses. A common suspicion raised by experimental settings is this: participants in the lab do not replicate their everyday use of language, due to the artificiality of the tasks and of the contexts involved, so that elicited speech should not constitute a reliable source of data. This paper sets out the experimental settings and results of four empirical studies - two studies investigating the pragmatic value of prosodic focalization through the controversial use of elicited data, one study on dative alternations based on a corpus and on a rating task, and one study on the contextual determinants of intonational contours based on a production task - to dispel this methodological suspicion: the artificiality of elicitation protocols does not prevent participants from using language as they do in spontaneous interactions. Careful examination reveals that the biases observed in the first two studies arise because subjects are not provided with sufficient cues concerning the context. This paper borrows the Gibsonian notion of affordance to characterise the state in which a context provides optimal resources to enable the production of the targeted construction, and argues that elicited data are reliable only when contexts optimise affordances.