You answered this in your own post...
The measurement is made in avg mode , if I check in pep mode this behaviour is not shown.
You did not describe the meter, so I can presume that the AVG on that meter looks at ALL power in which to calculate Average - so the carrier and the audio are SUMMED.
IN doing that - the carrier power and Audio being summed, then is averaged.
20W out of a RCI radio - pretty much tells me the ALC and or AMC are causing a contention. You see PEP not changing - so why AVG?
The meters way of processing summation is why.
The meter is not "intelligent" enough to know the differences between Carrier power and Audio Peak power (a wattmeter is just that - a wattmeter - not a graphic equalizer you can tune and adjust the audio bands) - just that audio is there and because the envelope is changing - that meters AVG circuitry is showing negative because of the change in power - both audio and carrier - is not staying the same - so it shows average - because of the circuits means to detect the PEP is not being used, the AVG is.
Most AVG meters are using a bias or weighting method to pass this RF power into a detector that uses a network of diode, resistor and capacitor - in such a way to show steady state RF - carrier power (because of the RF detector is small and the high-speed diode detectors in that assembly are looking at RF - not able to rectify the audio signal directly - just all the power passing thru the detector) and audio power in it - the detector circuit is only designed for RF - not look at the spectrum of Audio envelope within the RF signal directly - you can show forward power - but in one single tone - not a spectrum of audio - for the detectors are not able to see the carrier - because the envelope of audio around it - is confusing the detector - it's seeing the carrier signal and a "wrong signal" not in phase with it - it looks more like a negative element (envelope forward and backwards from the center carrier) and the detectors can't process audio - they only process RF.
A good example would be of a car on cruise control driven on flat terrain is not needing as much throttle as one that has to climb a lot of hills - so the needed power from the throttle and then the coasting - changes the way engine runs,
At one point its generating a lot of power to stay at that speed, then the next moment it's backing off to coast down the other side with less throttle to maintain that speed.
So, you're looking at Carrier (your 20W) - but the energy input is changing so you see AVG falling.
So back off that carrier power to produce a few less watts to make more room to get the audio power to swing forward - you review the AVG and reduce carrier and modulate until the meter doesn't swing forward or backwards - then you have tuned the AVG of the meter to show the radio is producing power, so the meters detection circuitry is not showing negative - or as much. Then the power envelope of audio or that out of phased audio envelope getting summed into the signal - then offsets this loss you see, when carrier power shows one thing then audio added in, AVG tends to fall.
Again, the detector is for RF - so yes, audio envelope is also RF - but 50% of audio power as envelope in that signal is also going thru the detectors as a negatively phased signal against the power it detects from the carrier - so the meter can't "see" the signal - power drops due to it.
You can also adjust ALC and AMC on the radio to offset the effect at higher power levels - but you won't like the results - you'll get to a point where the radio may be on AM mode - but people listening to you - find they can hear you better on FM mode in their radios.
Continue to adjust the wrong way, and all you have is distortion.
