Attention: Answers.ros.org is deprecated as of August the 11th, 2023. Please visit robotics.stackexchange.com to ask a new question. This site will remain online in read-only mode during the transition and into the foreseeable future. Selected questions and answers have been migrated, and redirects have been put in place to direct users to the corresponding questions on Robotics Stack Exchange. Additional details are available here.
If you have a good map and not-so-challenging environment, I think it is ok if your odometry is not so good, because you can always tune AMCL to perform localization intensively.
If you have a good laser, you can consider using laser odometry to replace wheel odom completely. Hector SLAM (http://wiki.ros.org/hector_slam) and Laser Scan Matcher (http://wiki.ros.org/laser_scan_matcher) could be some good options.
I personally think a good odometry should have error not more than 5%, most of the time should be less than 2%. A good option is to perform sensor fusion between your wheel odom and other sensor (gyro, IMU, laser odom...) using UKF or EKF which is part of robot_localization package itself.