Think about the process of writing a paper: you do some keyword searches for recent articles on related subject. You then look at the bibliographies for those articles, pick out whatever looks relavent to your topic, look at the bibliographies of those articles, etc. What this means is that apart from your initial keyword search, the primary criteria for including an article in your research is: "has it been cited already?". Relevance is merely a secondary filter.
This paper pointed out the effects of this phenomena: the vast majority of published scientific papers are never cited again; a moderate number are cited only a few times, and the remaining few -- having reached a bibliographic critical mass -- are cited thousands of times. The authors of the paper made a strong case that this was not a good reflection of the quality of research. In many cases, the process of reaching bibliographic critical was simply based on the almost random chance of acquiring those first few citations. The authors provided several examples of important scientific ideas which had been lost for decades, arguably because they had not attracted a critical mass of citations in the years immediately after publication.
In other words, humans suck at pagerank.
Anyhow, it occurred to me that this is a problem which could be solved with technology. Imagine an online word processor which -- in a sidebar -- suggests potentially related articles from ArXiv and Google Scholar. This would be based not on crawling bibliographies, but rather on semantic analysis of the adjacent paragraphs.
I think this would create some real benefits. It would remove much of the problems with citation bias, ensuring that important ideas aren't lost, and also that prior research isn't unwittingly duplicated. Wish I had time to implement something like this!