I would approach the problem this way:

c:\@Work\Perl>perl -wMstrict -le
"use Test::More 'no_plan';
 use Test::NoWarnings;
 ;;
 my $metadata = qr{ < [^>]* > }xms;
 ;;
 for my $ar_vector (
   [ 'somewordaa<metadata><moremetadata>r another wo<metadata>rd',
     'somewordaa<metadata><moremetadata>r another wo<metadata>rd',
     ],
   [ 'aword<metadata>aa<metadata> another wo<metadata>rd',
     'aword<metadata>a<metadata> another wo<metadata>rd',
     ],
   [ 'wor<metadata>dsandmor<metadata>ewordsaa<metadata>',
     'wor<metadata>dsandmor<metadata>ewordsa<metadata>',
     ],
   ) {
   my ($s, $expected) = @$ar_vector;
   $s =~ s{ aa (?! $metadata* r) }{a}xmsg;
   is $s, $expected, qq{'$expected'};
   }
"
ok 1 - 'somewordaa<metadata><moremetadata>r another wo<metadata>rd'
ok 2 - 'aword<metadata>a<metadata> another wo<metadata>rd'
ok 3 - 'wor<metadata>dsandmor<metadata>ewordsa<metadata>'
ok 4 - no warnings
1..4
[download]

Note that making the character class match 'atomic' or possessive may speed things up a bit for large data (for some definition of 'large') by suppressing backtracking on subsequent match failure:
    my $metadata = qr{ < (?> [^>]*) > }xms;
and for Perl version 5.10 and above:
    my $metadata = qr{ < [^>]*+ > }xms;
(both variations tested). Your data will determine if possessiveness actually makes a difference. The regex compiler may be smart enough to make this optimization on its own (I haven't checked), but giving the compiler a hint never hurts.

And of course, you will want to add many more test cases!