EDS software (as most of software goes) is dumb, if asked a wrong question it will give a wrong answer. To get better results you need to find a person capable of asking right questions (i.e. well trained operator). EDS suffers more in wrong hands than other analytical techniques.
In general, light elements are difficult to quantify by EDS due to limited sensitivity. Since oxygen is so common, most EDS software will quantify this by stoichiometry, which is typically more accurate, provided you know the valence state of the cations. Chances are, this is an option that can be switched on or off in your software and you may be able to change the stoichiometric ratio. If you are really interested in oxygen content and don't know the valence state (or it varies), I would suggest switching that off and doing a standard-based analysis. You'll need a reference standard(s), preferably of similar composition to your unknown sample.
Stoichiometry is good only for (surprise) stoichiometric compounds, works good for geological and some other specimens, but in many cases only standards can help. What's more, standardless quantification of oxygen containing specimens can lead to big errors, so even in case of stoichiometric calculations, use of standards should be encouraged.
It means you have 13% in weight and 48% atomic (# of atoms). This depends on the elements you have chosen and, as Vladimir Dusevich pointed out, EDS results rely on the experience of the operator and the questions that are being asked.
If you are using standardless quantification, I can affirm you are going to have incoherent and troublesome results. That is because the Standardless quantification tries to estimate the K-ratio and it uses ZAF correction (most common). This tends to overcompensate the k-ratio of Oxygen and you will end up with results that exhibits an excess of oxygen compared to the real composition...
There are 2 solutions:
if you sample is fully oxidized - use the stoichiometry function in the software - it will remove oxygen as a variable and will calculate all other elements, using the oxide ratio. As Vladimir Dusevich said, useful for geological samples, but also glasses and ceramics.
if you are not sure about the oxidation state or you sample is not fully oxidized - use standards - it can be a commercial oxide, another sample in your lab that you know the composition for sure, a sample you run a AAS or AES, commercial standard, etc. You just have to have (and know) a sample composition close to yours. It is a bit more work but this is the most accurate way to get reproducible EDS results and you can even get nice quantification for light elements (Oxygen and Nitrogen are common examples). From my experience you can have accuracy as small as 0.2% at when you use a standard. Note: you should use the standard every time to get concise results, since small variations in the beam parameters could lead to some differences in quant maths...
There are a few basic things which needed to be understood:
1. EDS is not a "magic tool". Sometimes people think so - just put a sample, press the button and collect representative results. It is totally wrong way.
2. It is necessary to understand the physics of the analysis: how does it work in general, which fundamental limitations it has, etc. Typical parameters like depth of beam penetration (well-known onion-like picture) and lines for most common elements (there are a lot of good tables from equipment manufacturers).
3. Just bear it in mind: EDX cannot replace normal chemical analysis for bulk samples. It operates only with pretty thin layer (about 1 um and less in SEM) and is very sensitive to any surface contamination. Not only contamination. E.g. oxidation is important too. Obviously, if the thin surface layer contains X% oxygen, it doesn't mean that the same concentration is in the whole volume.
4. An operator should know all software features - thresholds, marks, etc. The software is stupid itself, e.g. when there are not enough counts and the spectrum is pretty noisy, it can "recognize" any random "peak" in the noise as an element. Then the user asks: "Why does my sample contain plutonium?!!111111" :-)
5. It is necessary to have a statistically sufficient amount of counts. Don't try to finish the analysis ASAP. But sometimes (especially for thin films) beam induced damages can dramatically change the results and very long time or large current can give wrong results too.
6. The method itself is more suitable for qualitative studies, mapping and rough estimation. Precise quantitative studies are pretty tricky (proper calibration, quantitative estimation of possible systematic errors, etc.) and can be done not for all samples. The software can give results even in format "0.123456789", but it is necessary to understood that it is not real precision.
7. Conductive coating (if it was deposited) should be excluded with software.
in the time of black-boxes also magic tools are part of our job, unfortunately. First we have to consider what is possible with XRF. Second, we have to look our planet and after having studied both we can talk about detection of oxygen by XRF methods. When having get a qualitative result we can start thinking about quantification. After finish, we look back and think about the use of another method for solving the problem. To the first part: The X-ray fluorescence signals of oxygen are not able to cross large amount of material, they are absorbed. In some cases, only the uppermost atomic layers are analyzed, not more. Having in mind the abundance of oxygen in the earth-crust (nearly 50%) and the ubiquity of the element, we are possibly looking only to a contamination. Combining these, we have the analysis of the oxygen and we would like to consider if it is that of our sample? If yes, then we start with the quantification. There is more than one reason justifying the existence of analytical chemists and some of them may be replaced by a black-box or a magic tool… we have only to ask what we need, a reliable result or something else!
My understanding is that EDX is better utilized for elemental composition i.e. identification purposes and rather apply another technique like XRF or ICP to quantify the elements identified during EDX analysis.
looking the answers to this question, I become more and more desperate. It is not the chemistry it is simply physics making the restriction for a reliable result of oxygen analysis via XRF methods and as I wrote in my first answer, the high amount and ubiquity of this element in our planet. Please consult e.g. E. Bertin, the Bible of X-ray spectroscopy... I suspect that people are not reading the older literature containing valuable information...
under the action of the accelerated electron beam the X-ray yield of light elements is small which can be absorbed by the sample it self. that is why EDX technique is not so reliable in detection elements like oxygen.