Design The internet was searched using an application programming interface to Google, a popular and freely available search engine.

Main outcome measures The proportion of reprints of journal articles published between 1994 and 2004 from within 13 subscription based and four open access journals that could be located online at non-journal websites.

Results The probability that an article could be found online at a non-journal website correlated with the journal impact factor and the time since initial publication. Papers from higher impact journals and more recent articles were more likely to be located. On average, for the high impact journal articles published in 2003, over a third could be located at non-journal websites. Similar trends were observed for the delayed or full open access publications.

Conclusions Decentralised sharing of scientific reprints through the internet creates a degree of de facto open access that, although highly incomplete in its coverage, is none the less biased towards publications of higher popular demand.