Building a confusion matrix takes time and this time should be planned in any map production project. The confusion matrix is obtained by counting the number of occurrences of the pairs "map class/real classes" for a set of (usually) randomly distributed coordinate points. You can extract the class of the map for each sampling point using basic GIS tools. However, the only way to extract a trustable reference dataset on Google Earth is by photointerpretation or, better but unfortunately rarely done in practice due to the cost, by going on the ground. Of course, even if the photointerpretation is not perfect and therefore can get some help from ancillary data if you are unsure (especially if you do not know the area that you are validating). However, at least you can avoid the bias of another classification. It is futhermore recommanded to use a higher resolution image for the validation than your actual product, and this also mean that your photointerpretation must take the different scales into account.
In practice, Google Earth and other services that provide very high resolution images are accepted for the validation of land cover maps. As you said, you could enter the Lat/long coordinates of each of your points and write theire photo-interpreted land cover to create you validation dataset. Idelally, i would seek a system to directly write the values using a drop-down menu to go faster. At least, import your spreadsheet to GE in order to make sure that you have the exact location of each point.
Alternatively you can compare your map with another map, but you will not know if the conflicts are due to errors in your map or in your reference map.
as a final remarks, some great initiatives are distributing their validation points. For example the data from "collect earth" (soon to be released) could be very usefull in your case.