As Jessica mentioned, it may be the individual Goldens we have in the reference panel (12 dogs) don’t represent all lineages of Golden out there. For the new breed panel (v2) that I am constructing and testing right now, we have a different selection of 12 Golden Retrievers (who had deeper sequencing data), so the percentage calls on Bee might shift again when we re-analyze Bee’s DNA with v2 of the panel.
In addition to a wider panel of breeds, Embark may have more than 12 individual dogs / breed represented in their reference panel. For every new purebred Golden Retriever that genotypes through Embark, they might further expand the panel, thereby representing more variation across Goldens.
We also use a different type of machine learning algorithm than Embark, which may introduce some differences. Our method is very similar to the approach that the human genetics test 23andMe uses. In the long term, I’d like to do some comparisons between ancestry calling methods and see what works better on mixed breed dogs.