Apart from minimally adequate hardware components in consumer smartphones, the main source of perceived inaccuracy using satellite navigation is the same as almost every other flaw in the modern world.
Reality is failing to comply with the way we want things to be.
Early computer games notwithstanding, the world is not shaped like a donut or toilet roll (depending on whether you could move directly from south to north edge, or just east-west). Despite the lies-we-tell-to-children, it is not a sphere, or an egg, or a (aussie rules/rugby/american) football, or a pumpkin.
Initiates of the mysteries know that the true shape of the world... is like a potato.
As this isnt something that can be easily mathematically modeled, a number of best-fit approximations have been created, appropriate to different regions of the potato. A common trap for novices is using the model for some other part of the world. As we accumulate observations on the real shape of the world, these models are refined, and at intervals put into public use - Australia is about to move onto GDA 2020, (apart from our capital which is still running on AGD 66).
Even these best fit models need further kludges to get close to the shape of the potato, so there is what is called the N value or separation, which is a database of the local differences between the model and the potato, on an 1.8 km grid over all of Australia - (dont know how others handle it)
None of this was relevant before we started spinning satellites around, or tried measuring back from them to where we were standing.