In specific reply to Bator.
For analysis of ORE, you are using it not to identify actual composition of ore, but to provide qualitative assessments of ore, based upon quantitative traditional fire assay. So as long as you have the library that says ore prepared to this particle size, with this pretreatment method, that analysis with XRF says is ABC is known by fire assay to have X PPM Au. That is, it is NOT a quantitative analysis of the ore. Sample preparation is extremely important, as is presentation and analysis of data. It is NOT replacing traditional assay by fire assay and quantitative chemistry or ICP.
For analysis of auto catalyst...exact same as ore. It's all about sample presentation and assay library. A world class analyst was using his extensive knowledge energy dispersive XRF to analyze automotive catalyst and only coming within 20% of actual PM content. His problem wasn't the XRF, but the library of samples he used to do the curve fitting. He wasn't a chemist...he was an XRF geek...a darn fine one. And keep in mind, even at 20% you are still going to be pretty close to the limitations of the machine on what is really a low grade material.
For analysis of prepared melts. Best method is pin sample from well mixed melt, rolled flat and polished with pumice. Build a library of known, independently assayed samples.
As far as just trusting the numbers on the screen...I wouldn't for anything other than prepared karat melts. I've seen variations as high as 10% on Pd in well a prepared sample. That was 10% higher than what was read on desktop energy dispersive as well as independent assay.
Niton has improved their library immensely from when it first came out, but initially it was missing key elements, like zinc in the jewelry assay. It is entirely limited by the knowledge you prepare yourself with, and the library of samples you have to compare the spectra to.
GIGO-Garbage in, garbage out. Put your time in to sample prep, and understand the limitations of the assay, and it will work well for you....expect it to replace a knowledgeable analyst, and you will not like it.
The most advantageous thing you can do is to understand the spectral resolution of the sensor, so that you can understand the overlaps, and what COULD be there. Knowledge of what IS LIKELY there is still key in interpreting data.