Zillow is a regression model with a time series component to pedict real estate prices, which is much more useful on the aggregate level than on the individual decision level.
Let's say their formula averages the insanely high R- Squared of .5 that means only 50% of the variances in home prices can be explained by the factors they are considering. Even with this high R- Squared value, the standard error is off the charts (meaning that we are perhaps confident that 95% of observations fall within +/- 75% of the median predicted interval)
Y (Predicted Value)= a(size of house)+ b(number of bathrooms)+ c(number of bedrooms) + d
Where values a, b, c, and d are estimated at the "neighborhood level" over the past 4 years.
As you can see, there is no real estimate for the quality of the home itself, or other measurable attributes such as the presence of an HOA, whether you rent the lot, specific proximity to resources (that should be taken care of in a, b, c and d, but often isn't), etc.
Zillow is very useful for specing out neighborhoods, and to gain trend knowledge, but its not useful for picking individual winners/losers. Well, I suppose you can use the picture function for that, but thats not the prediction portion.