If non-nil, Gnus will gather all loose subtrees into one big tree and create a dummy root at the top. (Wait a minute. Root at the top? Yup.) Loose subtrees occur when the real root has expired, or you've read or killed the root in a previous session.

When there is no real root of a thread, Gnus will have to fudge something. This variable says what fudging method Gnus should use. There are four possible values:

Loose threads are gathered by comparing subjects of articles. If this variable is nil, Gnus requires an exact match between the subjects of the loose threads before gathering them into one big super-thread. This might be too strict a requirement, what with the presence of stupid newsreaders that chop off long subject lines. If you think so, set this variable to, say, 20 to require that only the first 20 characters of the subjects have to match. If you set this variable to a really low number, you'll find that Gnus will gather everything in sight into one thread, which isn't very helpful.

If you set this variable to the special value fuzzy, Gnus will use a fuzzy string comparison algorithm on the subjects (see Fuzzy Matching).

This can either be a regular expression or list of regular expressions that match strings that will be removed from subjects if fuzzy subject simplification is used.

If you set gnus-summary-gather-subject-limit to something as low as 10, you might consider setting this variable to something sensible:

(setq gnus-simplify-ignored-prefixes
      (concat
       "\\`\\[?\\("
       (mapconcat
        'identity
        '("looking"
          "wanted" "followup" "summary\\( of\\)?"
          "help" "query" "problem" "question"
          "answer" "reference" "announce"
          "How can I" "How to" "Comparison of"
          ;; ...
          )
        "\\|")
       "\\)\\s *\\("
       (mapconcat 'identity
                  '("for" "for reference" "with" "about")
                  "\\|")
       "\\)?\\]?:?[ \t]*"))

All words that match this regexp will be removed before comparing two subjects.

If non-nil, this variable overrides gnus-summary-gather-subject-limit. This variable should be a list of functions to apply to the Subject string iteratively to arrive at the simplified version of the string.

Useful functions to put in this list include:

You may also write your own functions, of course.

Since loose thread gathering is done on subjects only, that might lead to many false hits, especially with certain common subjects like `' and `(none)'. To make the situation slightly better, you can use the regexp gnus-summary-gather-exclude-subject to say what subjects should be excluded from the gathering process.
The default is `^ *$\\|^(none)$'.

Gnus gathers threads by looking at Subject headers. This means that totally unrelated articles may end up in the same ``thread'', which is confusing. An alternate approach is to look at all the Message-IDs in all the References headers to find matches. This will ensure that no gathered threads ever include unrelated articles, but it also means that people who have posted with broken newsreaders won't be gathered properly. The choice is yours---plague or cholera:

If you want to test gathering by References, you could say something like:

(setq gnus-summary-thread-gathering-function
      'gnus-gather-threads-by-references)