The new paper was published at Scientometrics by Li, Thelwall (still one of my dissertation advisors) and Giustini. They focused on the correlation between user count - the number of users who save a particular paper - and WoS and Google Scholar citations.
The researchers extracted from WoS all the Nature and Science research articles that were published in 2007 and their references. They ended up with 793 Nature and 820 Science articles, or 1,613 articles overall (not including references, of course). Then, they searched CiteULike for those articles' titles and number of citations, as well as for their user count in Mendeley. They also collected the same data from Google Scholar. It's important to note that Mendeley had 32.9 million articles indexed while CiteULike had only 3.5 at the time of the study.
Google Scholar's mean and median number of citations were higher than in WoS (not surprising; If you want better citation numbers, always use GS). They found that despite Mendeley being "younger" than CiteULike (launched in 2004 and 2008 respectively), CiteULike had only about two-thirds of the sample articles saved, while Mendeley had about 92%.
Spearman correlations between citations in GS and WoS were high in this research (0.957 for Nature and 0.931 for Science). The correlations between Mendeley's user count and the citations in GS and WoS were also rather good (0.559 and o.592 for WoS and GS respectively for Nature, 0.540 and 0.603 for Science). CiteULike had far weaker correlations: 0.366 with WoS and 0.396 with GS for Nature, 0.304 with WoS and 0.381 with GS for Science.
The authors remind us that correlation isn't causation, saying they can't conclude a casual relationship based on correlations between two data sources. Therefore, it can't be determined for sure whether there is a connection between a high user count and a high number of citations. Only Nature and Science were studied, so it can very well be that the results aren't true for other journals. Also, group-saved and single-user saved references were given the same weight. The number of saved references in Mendeley and CiteULike is much smaller than in the WoS counts and therefore the results might be less reliable.
The authors speculate that user count may represent a more accurate scientific impact of articles, and take note that one can measure the impact of all sorts of resources in online reference managers, unlike in the limited bibliographic indexes.
I think it could be reference managers don't always reflect readership: one could save a reference and forget about it all together later (so many articles, so little time...). On the other hand, citation counts might suffer from the same problem, as many scientists use a "rolling citation" from other articles citing an earlier article, without actually having read the article themselves.
Priem et al. also presented lately a study about web citations and WoS citations, based on data from the seven PLoS journals, but I think I'll wait for the journal article to cover it in the blog.
Li, X., Thelwall, M., & Giustini, D. (2011). Validating online reference managers for scholarly impact measurement Scientometrics DOI: 10.1007/s11192-011-0580-x