Every tool that prices a house hands you one number, and the number is the least useful part of what it computed. The work that decides whether you can act on it is underneath: which sales it chose, and what it added or subtracted to make each one look like your house.
That gap is why investors keep arguing about automated values in public. Ten distinct investor forum threads turned up in our search this week asking some version of the same question, and the ask is never for a better estimate. It is to see the comps and the adjustments so the investor can check the math themselves [3][4].
An estimate is weakest exactly when you need it
Zillow publishes the accuracy of its own model, and the split is the most useful number in the industry. For homes that are on the market, the Zestimate has a median error of 1.83 percent. For off market homes, the median error is 7.01 percent [1].
Read that again in the order that matters to an investor. The accurate case is a house that is already listed, where the model can see an asking price set by a human who walked the property. The four times less accurate case is a house nobody has listed, which is every house you look at before you buy.
The miss is not academic. A 7 percent error on a 400,000 dollar resale is 28,000 dollars, and the typical home flipped in the second quarter of 2026 produced a gross profit of 60,526 dollars at a margin of 21.5 percent, down from 27.6 percent a year earlier [2]. On those margins, one valuation error the size of the published median takes close to half the profit on the deal. Figures in the worked example below are an illustration, not a measurement.
So the question is never whether a model is good. It is whether this particular number, on this particular house, rests on sales you would have picked yourself.
What an adjustment is actually claiming
A comp adjustment is an argument in dollars. When a tool adds 12,000 dollars to a comp for square footage, it is saying: this house sold for X, it is 300 square feet smaller than yours, and in this submarket that difference is worth 12,000 dollars. Every line in an adjustment grid is one of those claims, and each one can be right, stale or made up.
The grid usually carries five or six kinds of line: size, bed and bath count, condition or finish level, sale date, distance, and lot. Three of them deserve your attention more than the rest.
Sale date. A comp that sold nine months ago is being asked to predict today. The adjustment that bridges those months is a guess about the local market's direction, and it is the line most likely to be wrong in a market that turned.
Condition. This is the line that decides a flip, and it is the line with the least data behind it. A sale record rarely says whether the kitchen was original. If the tool is adjusting for condition without a source for condition, it is estimating an estimate.
Distance. A quarter mile can cross a school boundary, a rail line or a flood zone. Distance is cheap for a model to measure and expensive for it to understand.
Four checks, ten minutes
Here is the sequence worth running on any automated value before you underwrite from it.
1. Count the comps and look at the spread. Three sales that agree within 3 percent is a number. Three sales spread across 20 percent is an average pretending to be a number. If the tool shows you a per comp adjusted value, read the highest and the lowest before you read the middle.
2. Find the largest single adjustment. Sort the lines by size and ask what the biggest one is paying for. As an illustration, take a subject at 1,800 square feet and three comps:
| Comp | Sold | Size | Raw price | Net adjustment | Adjusted |
|---|---|---|---|---|---|
| A | 2 months ago | 1,850 | 392,000 | -3,000 | 389,000 |
| B | 5 months ago | 2,150 | 448,000 | -54,000 | 394,000 |
| C | 8 months ago | 1,610 | 338,000 | +41,000 | 379,000 |
The three adjusted values sit within 4 percent, which looks reassuring. But comp B is carrying a 54,000 dollar deduction and comp C a 41,000 dollar addition, both driven mostly by size. Two of your three comps are the wrong house, corrected by arithmetic. The honest read of that grid is that comp A is the comp, and your ARV rests on one sale.
3. Check the direction of every line. A bigger comp should pull down, a more recent sale in a rising market should pull up, a worse condition should pull up. A line pointing the wrong way is a defect in the grid, not a nuance in the market, and it is the fastest way to catch a tool that is pattern matching rather than appraising.
4. Ask what was excluded. The sales a tool rejected tell you more than the ones it kept. A model that silently dropped the two closest sales because they were 200 square feet off has thrown away your best evidence to protect a tidy size band. If you cannot see the rejected set, you cannot see the ARV.
When you cannot read the grid, you do not have a number
If a tool shows a value with no comp list, treat the output as a lead filter and nothing more. It is fine for sorting a hundred addresses. It is not fine for an offer, and it will not survive a conversation with a lender or a cash buyer, who will order their own opinion of value and check the comps themselves.
That is the part investors learn late. The number is not the deliverable. The defensible comp set is the deliverable, because that is what a third party can audit.
Doing this on iComps
This is the reason our result page is built the way it is. Every ARV on iComps opens into a comp validation table that shows the sales used with their size, bed and bath count, distance and sale date, the disqualification criteria that excluded others, and the size, room and sale timing bands the search actually ran with. Each comp expands into its own value adjustments ledger, line by line, with the weighted ARV, the comp count and the average net adjustment next to it, so the largest adjustment is one click from the number it produced. When the comps are thin, the page says so instead of hiding it behind a confident figure. Run an address at icomps.io/calculate and go straight to the adjustments.
Do this next
Take the last deal you priced from an automated value and open its comp list. Find the single largest adjustment, and decide whether you would have paid that difference in cash for that feature. If the answer is no, you do not have an ARV. You have an average, and the market will charge you for it.
[1] Zillow, Zestimate accuracy, median error 1.83 percent on market and 7.01 percent off market, https://www.zillow.com/learn/influencing-your-zestimate/ (read 2026-10-04). [2] ATTOM, home flipping profits continue gradual two year decline, Q2 2026 typical gross profit 60,526 dollars at a 21.5 percent margin against 27.6 percent a year earlier, https://www.attomdata.com/news/market-trends/flipping/home-flipping-profits-continue-gradual-two-year-decline/ (read 2026-10-04). [3] https://www.biggerpockets.com/forums/67/topics/994956-apps-or-websites-for-finding-comps-to-determine-arv (2026-10-02). [4] https://www.biggerpockets.com/forums/80/topics/353042-zillow-zestimate (2026-10-02).
Photo by ErikaWittlieb on Pixabay.
