Three set point are needed for average curve fitting and base line correction. You can use standard of same or parallel group for comparison with near about similar molecular weight. 

I assume the question focusses on a calibration curve for quantitation. - I do not think there is a general rule. In my lab we usually do about 7 different concentrations with at least double injections. If there is a reasonable number of samples we would run this set, than e.g. 50 samles in double injections than another calibration set etc etc.

It is generally agreed that you should cover your whole dymanic range. If you find concentrations below or above the calibrated range this is outside the proven linear range and thus not really a valid result.

For each point of the calibration curve, generally three replicates are needed. Results of the three analyses are averaged giving one point of the curve. Better to obtain more than three set points in which your sample concentration should come in the middle.

You need minimum five points to cover your calibration range.

Dear Muhammad, Dear Wentai, I am aware several regulations make suggestions. (Codex Alimentarius, diverse Quality Assurance systems for drinking water determination etc. Though I have been around for some years I do not think I have ever seen a general rule that is valid for all situations so if you could give a reference that would be helpfull.

At least 3 of the 5 best

I also have not seen a rule. I  think it depends on how good (and how stable) your correlation factor is. We use automated SPME GC-MS/MS for trace analysis of organic contaminants from drinking water. We run a calibration curve every day (on every fibre-every 50 injections) and a verification sample every 10 injections. We only do one injection for each calibration point and have very stable corrrelation factors over 90 % for all analytes (11). Running triplicates for each calibration point would decrease our productivity very much. The intra-day RSD is 2-3 % for the lowest calibration point. I realy do not see what improvement running the calibration curve points in triplicate would bring. In fact, the variability due to matrix is much higher and we try to address this by using stable isotope internal standards for each analyte.

Now its more clear to me.Thank you all for your help I appreciate

Unless you have problems with repeatability you should not need more than a single point at each level. If you construct a calibration from the means of replicates at each level then you must use the mean of the same number of replicates form samples to calculate results.

I agree with Kai Bester. There is no general rule. In our lab, we make 4 different concentrations and inject every standard 3 times. If the relative standard deviation values for area are good then no problem. If the deviation is too much then we repeat the calibration process.

No "standard" method that I know of requires the use of replicate injections for standards. You invariably measure your method precision elsewhere in the analysis (matrix spikes, blank spikes, 2nd source standards, etc). Since you almost always need to use internal standards for GC-MS analyses you are already controlling for variability in the injection.

There is no general rule for the number of points and replicates needed.  However, as a general rule for calibration, the actual number of calibration points should be based upon the width of the working range and the shape of the calibration curve, and should insure the accuracy of the determination. Non-linear, or quadratic curves require a minimum of five calibration standards to fully characterize the curve.  Also if you are running standard methods, it is likely that the methods or the agency will have requirement for the number of points for calibration. 

The following EPA documents have some statement about calibration:

http://www.epa.gov/fem/pdfs/calibration-guide-ref-final-oct2010.pdf

http://dnr.wi.gov/regulations/labcert/documents/guidance/-LODguide.pdf

Dear Marcelo I. Guzman , is this a general rule, to not include the averaged values? Usually everyone averages replicate values before calibrating. But, the results are different.... So what is the "right" way?

Thanks a lot.

Benedikt Engel If the calibration is constructed using means of replicates, then the results must be calculated using means of the same numbers of replicates of each sample.

I agree with Guzman, run at least 5 points, check the SD. And you are good , there is no need for the replicated runs, except were you have reproducibility issues

Thank you for your answers. I think replicates are important - not measurement replicates, but preparing each standard of the calibration e.g. three times...

The results of the calibrations are different if calculated with means or single values. So I wondered if there is a guidline or similar for that...

I am looking for the same answer. I haven't found a guideline yet. I think as long as the RSDs of each point are less that 3%, we can average them.