Does this seem correct & fair?

"if Zestimate isn't more accurate!!"

Utah is one of the states where the sold prices are not part of the public record.  Zillow indicates on their analysis of the machine generated opinions of approximate value (Zestimate tolerance) that when they rate an area with one star, it means that the estimates are actually just the Tax assessor value, or that they don't provide an estimate.

But Zillow does provide numbers for 777.2k units out of 892.5k units on Zillow in Utah, meaning that only 12.9% don't have such "opinions".

But the estimates I look at on listed properties in Westpoint Utah are substantially higher than the tax assessed value, thus it appears they may be using the listing pricing for doing the computer modeling.  Anyway, they appear to be doing some kind of machine estimating, but they can't calculate the deviance of the closest prior estimate to a recent sale, as they can't get the "sold" price numbers.

If you don't like the numbers, try realquest.com express, as apparently that is an AVM that the banks will use.  They won't use Zillow's AVM as that is against the terms of use.

One that I chose randomly to test in Westpoint UT that is listed for sale, Zillow indicates $161k to $215k; approximately $169.5k; listed $179.9k.
Realquest indicates the same property not listed, estimated $178k, standard Deviation $13k.

"Please explain!"

1) The specifics of the method are "proprietary", and publishing it would make it available to the competition and
2) The "modeling" used factors in all the non-distressed ownership unit sales in the local area.  It is all done automatically, so no one looks at all the specifics for any individual case.

But a few "clues" as to what the process is: since the methods are similar, and the source of their data has been published, and they have been very forthright with what they don't include:

It is a "model" using county record data and owner/agent revised data only.

The recently sold data is collected for the area (by distance, not neighborhood; larger distance for rural, closer distance for dense city areas...), then modeled to the public data:
1) house sqft
2) lot sqft
3) # bedrooms
4) # bathrooms
5) # stories
6) year built
7) tax assessed value (adjusted for local rules)
8) last sold price extrapolated from sold date
9) covered parking (if in the county records).

They exclude all "distressed" property sales from the modeling.

Then the sold units are tested against the results, and if way out of line, those units are excluded, and a new "mathematical model" calculated.

Then, with those results, the county data for the properties are used to calculate an "estimate"; not a "quote", not a "CMA", not an "Appraisal", not a "market value", not a "sale price", not a "list price", not an "offer price", not a "loan qualification"...

I would use Multiple Linear Regression for the modeling.  The statisticians have indicated they don't use Linear Regression (Least Squares Curve Fit, look up the LINEST function in your Excel program, especially example #3), implying they are using a neural network approach, similar to that developed by the U.S. Postal Service for Optical Character Recognition.

Yes, if you did remodels and additions and got permits, those are factored in, both in changes to the # sqft, #bedrooms/bathroom, as well as the tax assessed value.  But the tax assessed value never goes up proportional to the "new value" when you make an improvement; only the new improvement part goes up.

And the trick in any such procedure is how to weight any of the given factors.

Again, you must consider the "purpose"; it is to automatically track the housing market and give an "approximation" starting point of what something "might" be worth compared to something else.  It is never intended to substitute for an appraisal nor anything else.

So, if two identical properties (which really can't be identical as they can't be in exactly the same position...) sold for exactly the same amount, but 6 months apart, and the market had been "declining"; the one that sold earlier would be assumed by the "model" to be "worth less" since the market declined since.  It doesn't mean it is worth less, it just means that was what the market trend implied.

Notice, that the only place that things like fine cabinetry and extra insulation are factored into the estimate are:

1) the last sold price/date
and
2) the county tax assessed value.