EndNote Web mostly imports data from major fields correctly, with sporadic errors. Problems often occur with minor fields. Especially for CSA Illumina it was hard to find a working import filter. Import from PubMed and Scopus worked remarkably well (except for a special character error in Scopus).

Basically, import of references should work via both direct export (from Ovid and Web of Knowledge) and importing from a file after selecting the appropriate import filter. In practice, however, there were problems with references from most retrieval systems. Even important fields like author, journal, pages, or year contained errors in some references. Encoding errors also occurred in some references.

None of the tested retrieval systems offers BibTeX as an output format. However, BibSonomy imports only BibTeX data. Another possibility is to let BibSonomy directly import references from website text via browser bookmark buttons, but references imported this way are often fragmentary.
Thus, for this test we converted data from various retrieval systems into BibTeX using either the online converter BibTranslator or the MPG SFX button. The latter option is particularly cumbersome since only one reference can be converted at the same time.
Most imported references showed severe errors, e.g. in the authors, editors, title, issue, abstract, and pages fields. There were also character encoding errors in some references.

Zotero can automatically capture references from retrieval system websites, as well as import references from files in several formats. Automatic Capture worked well with Scopus, PubMed, and vLib, not so well with CSA Illumina, and not at all with the other systems. Ovid, Web of Knowledge, and SciFinder do not seem to offer export formats that Zotero can import either. (The ISI format from Web of Knowledge can be translated into a more common format with BibTranslator and then imported into Zotero.)
The test references imported from Scopus were the most complete, showing only minor errors (missing addresses, ISSN, DOI, etc.). References from all other retrieval systems often contained errors in important fields (journal, volume, issue, pages, year, etc.). Special characters were imported correctly.

JabRef is able to import references in various formats, and often automatically selects the correct import filter.
With 3 of the 7 tested retrieval systems, the imported references contained major errors, e.g. missing author, year, or journal data. References from the other retrieval systems were not completely correct either, with errors in metadata fields like addresses, keywords, reference type, etc. Character encoding errors occurred with data from CSA Illumina and Scopus, and one of the test references from Scopus couldn't be imported at all.

To be imported into Connotea, references need to have a URI, DOI, ASIN or PMID. Therefore, many references are lost when importing them from a retrieval system into Connotea. Connotea accepted most common metadata formats like RIS, BibTeX, EndNote, MODS, etc. However, sometimes valid data in these formats was not imported at all. References from CSA Illumina and SciFinder couldn't be imported at all. Those references that Connotea managed to import were usually missing data, because Connotea only supports the most important data fields. Sometimes even this basic data contained errors. There were no character encoding errors.
There is also browser bookmark button for importing single references.

CiteULike is able to import BibTeX as well as RIS data. Thus, references can be imported from many sources, but not from some retrieval systems like CSA Illumina, Web of Knowledge, or OvidSP.
Even when the import of references is possible, there are still often errors or missing data. The only tested retrieval system CiteULike works well with is Scopus.
The browser bookmark button for importing single references only works on certain supported websites and retrieval systems.