Theoretic question about filters

Suppose we have an RF power amplifier that the manufacturer has stated that in order for the amplifier to provide max power , it has got to "see" a load characterized by some impedance vs freq.

We also know that the amplifier in NOT linear (something like class C for example). We want to use the amplfier in a narrow band system, so we try to match the amplifier only in this narrow band of frequncies in order to obtain sufficient power tranfer to the load. We must also not be disturbing other bands, so we have to employ filterring. Suppose we decide to design a reflective type filter, that reflects the energy in the stopband back to the source.
We take our LINEAR simulator and design such a matching network and filter, and indeed we have an almost perfect match at the frequency of our band, and great missmatch losses at the stopband-->which means the stopband is efficiently filtered.
IF the amplifier was LINEAR,and if we examined the plot of S21 vs freq and found that at xxxMHz(in stopband) it is -30dB, we could confidently say that this frequency is attenuated by 30dB. Is that correct?
However, our amplifier is NOT a linear device, so a possible mismatch of 30dB (calculated using LINEAR theory) doesnt mean that the amplifier will deliver 30dB less power than that it would deliver if it had an optimum load connected. What I want to say is that (if for example the optimum load values were derived using load pull analysis) it could be possible for the amplifier to give 20dBm to a load of 20+j30, 19.5dBm to a load of 10+j5, 19dBm to 1-j20 (and so on), which is totally uncorrelated with the linear results...
So , the plot that we derived using the linear simulator is not at all accurate concerning the filtering characteristics. It maybe just shows the tendency of the filter - if the amplifier does not "see" the optimum load, by definition the output power will be less, otherwise the load wouldnt be optimum. But we dont know the amount of reduction in the output power.

Is all that true, or am I seriously missing something?

Look at how the manufacturer determines the optimum load. Some of them will give you the test circuit. Others use a triple stub tuner which has who knows what impedance at harmonics of the operating frequency. In each case the numbers they measure are representations of the physical world.

Thanks for the answer! So the procedure is: For every frequency I obtain from the manufacturer the power that the amplifier delivers to every possible load (eg in the form of load pull lines on a Smith chart).
For the network I have designed I know its input impedance (when the ouptup is well matched) for every frequency.
So if I combine these two infos I can find what is the power that the amplifier injects into my network at every frequency. If the network has only ideal reactors then the power injected into the network will be equal to the power that is deliverd to my final load! (If finite unloaded Q components are considered , my network will introduce some insertion loss at each frequeny, so the power deliverd to the load equals the power injected minus the insertion loss).

On the other hand tunning such a network (setup) , becomes an optimization problem , where we try to find (from the load pull data for example) what are the optimum values of the reactors that make my filter show an impedance (at the frequencies in the stopband) to the amplifier for which the amplifier injects minimal energy, and of course in the passband the filter shows the impedance that the amplifier wants to see for maximum power transfer.