My Exact System for Finding a House's ARV

TLDR

ChatGPT can help you comp a house, but it will miss flood zones, lean on Zestimates, and ignore a stale listing that tells you the real ceiling. The system that works is three good comps that match on features, date sold, and neighborhood, and a red-flag check for deal killers.

The Three Methods an Actual Appraiser Uses
You Do Not Care What the Zestimate Says
The 70% Rule Is Your Real Price
What Makes a Comp a Comp
The Seven Dead on Arrivals
How I Use AI Anyway

The Three Methods an Actual Appraiser Uses

I have appraised hundreds of potential deals in 14 years of buying real estate. A human appraiser uses three methods.

The income approach. Value based on the income the property brings in. Think of the 1% rule. Rent minus expenses gives you your net operating income. A property can be valued off NOI. Mostly used on multifamily or bigger apartment complexes, not single family.

The cost approach. Cost to build new plus land minus depreciation.

The sales comparison. You pull comps. Same as your after repair value. Same as a broker price opinion. This is the method that works on a single family home.

When I tested ChatGPT out of the box, it did not use any of those three. It used a fourth method, the AVM approach. The automated valuation model. That is the Zestimate. The Redfin estimate. ChatGPT averaged the estimates from those sites and then added what it guessed the improvements might be worth. It told me: “If the property is in average condition after full rehab, kitchen, bath, floors, windows, it might climb to the top of the market for that area.” That is a guess.

AI is guessing at a number real appraisers would not even be allowed to use.

You Do Not Care What the Zestimate Says

Zillow has a lot of data. It uses that data to compile an estimate. I am not saying the Zestimate is accurate. I am saying ChatGPT has no clue what the underlying data is. So when it averages those estimates, it is taking a guess on top of a guess.

You should not care what the Zestimate is. You should not care what today’s value is. You care what the house will be worth after you fix it up, and what you need to pay for it to hit your profit.

Common Mistake

Using today’s Zestimate to price your offer. That number has nothing to do with what the house will be worth after rehab, and nothing to do with what you should pay.

The 70% Rule Is Your Real Price

Here is the only equation you really need for flippers or BRRRR. Same equation either way.

Target purchase price equals 70% of ARV minus rehab cost.

Quick example. Say a house sells for $300,000 after rehab. 70% of $300,000 is $210,000. You think the renovation is $60,000. $210,000 minus $60,000 is $150,000. That is your target purchase price. That is what the house is worth to you today.

You do not give a crap about the Zestimate. $150,000 is your number. For the full breakdown of when to adjust that 70 and when to stick to it, see The 70% Rule Is Just a Ballpark.

What Makes a Comp a Comp

Three things make something a real comp.

Features. The house has to be similar. Reasonable square footage match. Bed-bath count matches. Garage, carport, basement status matches. Style matches: a bungalow is not the same as a split level. Age of build is in the same era: not a 1900s house against a 2000s house. Lot size is similar. Condition is the condition you plan to hit, not the condition you are buying in. Basements and ADUs do not count as equivalent square footage.

Date sold. Sold, not for sale. This is where people mess up constantly. “The neighbor is about to sell for $275.” That tells you nothing until someone writes a check. I had a house listed at $275 that nobody would buy. The best offer I had was a $200,000 lowball. If I took that offer, the “comp” everyone was pointing to was a lie. Stick to sold in the last 6 months. If you have to stretch, that is the first thing you fudge.

Neighborhood. This is the biggest one. Boundaries are major roads, highways, railroad tracks. Look at a map and see what a neighborhood actually is. The feel matters too. You can drive a block and tell when you are in a different neighborhood. The key I did not know when I started is the census tract. Not zip code. Zip codes are way too big. A zip code has multiple neighborhoods inside it. Pull a census tract map and use that.

Pro Tip

If you cannot find three to five clean comps, fudge date sold first. Never fudge neighborhood. Never.

The Seven Dead on Arrivals

Before a comp counts, I check for deal killers. I call them DOAs. Dead on arrival. These kill the value of a house and most of the time kill the deal.

#	DOA	Why It Kills Value
1	Flood zone	Unless every house around it is in the flood zone, stands out and scares buyers
2	Major [[zoning	easements]]
3	Zoning conflicts	Duplex in an R1 zone can get forced back to single family on permit pull
4	Built into the setbacks	House sits closer to a boundary than city code allows
5	Ceilings under 7 and a half feet	Never worth as much as regular ceilings
6	Houses on rural land with nothing to compare	No comp set means no confidence on ARV
7	Odd birds that do not match the neighborhood	You cannot predict a sale price for something that has no twins

I buy a duplex once thinking I could build another house next to it. Cleared the trees. Looked up and there were power lines running right across the middle of the lot. Utility easement. I had no clue it was there. Thought I was getting a great deal. That is how a DOA works. You do not see it, then you are stuck with it.

On the AI test I ran, ChatGPT tried to comp a house on the edge of a flood zone and totally missed it. When I asked it separately how much a 100-year flood zone devalues a house, it got the answer right: roughly $30,000 off at today’s interest rates plus the stigma. So it knew. It just did not apply that knowledge to my actual address. That is the problem.

Deal Killer

A DOA that the AI missed does not become less of a DOA. If you are using AI to comp, you still have to check flood zones, easements, and zoning yourself.

How I Use AI Anyway

ChatGPT is a great sidekick. It saves time on the heavy lifting if you know what to look for. I trained a custom GPT with three rules: identify three to five comps, confirm none are DOAs, break the comps down to price per square foot, and then look for currently listed or pending houses to prove or disprove the comp set.

Even trained, it still missed the flood zone. And it still missed the stale listing down the street that should have capped its ARV number. Those are two things I cannot let it miss. So I give it the research and I verify before I trust.

This is the whole theme of the solo house flipper model. You do not give blind trust to a machine, a piece of software, or a subcontractor. You know what you expect. You have the rules. You lead. You hold accountability.

AI does the legwork. You still make the call.

FAQ

How many comps do I really need?

Three to five clean ones. If you cannot find three in the right census tract that sold in the last 6 months and match on features, that is a signal the deal is in a weird spot.

What if nothing has sold in the neighborhood recently?

Then either stretch date sold out carefully or pass on the deal. Do not force comps from a different neighborhood. That is how people buy bad deals.

How do I find the census tract for an address?

Search “census tract map” for your county. Most counties have a GIS viewer. Drop the address in and you can see what tract you are in and what borders it.

Can I comp using just for-sale listings?

No. Use sold comps for your number. Use for-sale listings as a sanity check on market conditions, especially stale ones that are not moving. A stale listing tells you the real ceiling.

I am just starting out. Should I even use AI for this?

Use it to draft a comp list and save time. Then verify every comp yourself. Drive the neighborhood if you can. The goal is to build the gut for comping, not to outsource it.