An operator analysis of home valuation, pricing, and the AI tools reshaping them. Named tools, covered fairly, with the figures sourced and the blind spots stated, including our own.
Form 1004 Expires in November
On 2 November 2026 the 1004 stops being accepted. What replaces it is not another form but a dynamic report where free-form commentary becomes structured fields. The appraisal is becoming a database record while the CMA stays a document.
Read more →A Single Figure Is the Hedge
The category treats the point estimate as the product and the range as a qualifier. The range carries the information. The single figure is a reduction that discards the dispersion it resolved, without reporting how much resolving was required.
Read more →Why price per square foot misleads
Price per square foot is the number everyone reaches for and the one that misprices homes most reliably. The relationship is not linear, the denominator is measured inconsistently, and finish and location swamp raw area. Here is why, and what to use instead.
Read more →Comp selection is the whole game
Every valuation method is a comparable-sales method underneath. The number is downstream of one decision: which sales count. Here are the three dials every method turns, and the test a rule cannot run.
Read more →What confidence should mean in a valuation tool
Most tools give you a number to the dollar and hide how much they know. An honest valuation does the opposite: a visible range that widens where the data is thin. Here is the standard.
Read more →AVM, CMA, or reasoned valuation: what each delivers
Four ways to put a number on a home, the AVM, the CMA, the reasoned valuation, and the appraisal, each answering a different question, each with a built-in blind spot. Here is what each one is for.
Read more →How AVMs work, and where they break
An automated valuation model prices a home from public records and comparable sales. It looks accurate on listed homes and quietly fails on off-market ones, condition, luxury, and thin-data markets. Here is how it works and where it breaks.
Read more →The Condition Record Deletes Itself
Interior photos come down after closing, so the evidence of what a home looked like at sale disappears within weeks of that sale becoming a comparable. Condition is the largest adjustment in most valuations and the only one with no durable record behind it.
Read more →The Handoff Nobody Owns
Most agents run comp selection in one system and the client-facing presentation in another, and move the numbers across by hand. The error enters at the retype, and no vendor owns the seam because it sits between two products.
Read more →The MLS Is the Last Non-Conflicted Dataset in Real Estate
Every automated valuation and instant offer reads a source that is partial, lagged, or working for the party that built it. The MLS is the one dataset that is complete, timely, and neutral at once, and that is why it stays the foundation any trustworthy number is built on.
Read more →What the public record gets wrong
Every automated valuation reads the public record as if it were ground truth. It often is not. Square footage is off, renovations never posted, bedroom counts are stale, and a prior sale was not a sale at all. When the input is wrong, the confident output is wrong. Here is where the record fails.
Read more →What a Zestimate is doing under the hood, and the inputs it cannot get
A Zestimate is a statistical model reading public records and comparable sales. It is good at what it can measure and blind to what it cannot. Here is what goes into the number, and the inputs no algorithm can reach.
Read more →Why every home-value tool shows a different number for the same house
Type one address into Zillow, Redfin, and your county site and you get three different numbers. The gap is not a glitch. It comes from different data, different comparable sales, and different incentives. Here is what drives the spread.
Read more →The Rule That Already Applies
Six federal agencies wrote quality control standards for automated valuation models and the rule took effect on 1 October 2025. It reaches lenders making credit decisions. It stops before the estimate a consumer reads, which is the one that sets the expectation before any agent arrives.
Read more →The Missing Failure State
Feed any valuation tool a thin comparable set and a figure comes back anyway, to the dollar, with a confidence label beside it. A confidence band describes the output. It does not decide whether the output should exist. Almost no product has a state in which it declines.
Read more →Where the Model Is Reliable
The category has documented every failure of automated valuation and never specified the boundary. Error is a function of how much the unobserved variables differ across the comparable set, which makes the limit testable rather than rhetorical.
Read more →The Training Data Is Being Withdrawn
As a growing share of homes sell off the public market, days on market and price-cut history leave the record that every automated valuation reads. The models get thinner inputs. The outputs keep the same certain voice.
Read more →When the AI invents a comparable
A general AI asked for comparable sales will sometimes produce them: an address, a price, a date, stated with total confidence, for a sale that never happened. This is a different failure from stale data. Here is why it happens, and the one habit that catches it.
Read more →Why the AI market read is months old
Ask a general AI about your housing market and the answer arrives in confident present tense. The data behind it is not from this week. Here is where the model gets its market read, what moves faster than it can, and the question that exposes the timestamp.
Read more →When a client brings you a ChatGPT valuation: the questions that expose its blind spots
More sellers are arriving with a number from an AI chatbot. It sounds authoritative and is often built on nothing local. Here are the questions that show a client, without an argument, what the estimate could not see.
Read more →How to judge an AI valuation tool
Judge an AI valuation tool by whether it shows its reasoning, shows an honest range, and admits what it cannot see, not by how confident the number looks. A black-box number is the one to trust least. Here are the tests.
Read more →What AI changes in pricing, and what it relabels
AI has widened the data a valuation can use and sped up the work. The core method is the same one AVMs have used for years, and much of what is sold as new is the old model with a new label. Here is the line between signal and noise.
Read more →Twenty Buyers
Twenty MLSs, four percent of the national total, serve half of all subscribers in the United States and generate roughly half the sector's revenue. Most software an agent opens was not bought by that agent, which means the specification is written by a group small enough to fit in one room.
Read more →The Wrong Opponent
An automated estimate sets the expectation. It cannot sign a listing agreement. The industry has spent years building arguments against the half of the problem that has no commercial incentive and cannot be negotiated with.
Read more →Non-Financeable Is a Segment, Not a Defect
A property that fails the government-backed financing floor is not distressed and not cheap. It is a separate inventory class with its own buyer pool, its own price behaviour, and no one counting it.
Read more →When the Model Cannot See the Comps: Instant Offers in a Non-Disclosure Market
Texas does not disclose sale prices, so every automated valuation behind an instant offer runs on comps it has to estimate. In Dallas-Fort Worth the agent with MLS access holds the sale data the algorithm structurally cannot reach.
Read more →The Anchor Arrives First: Instant Offers in Phoenix
Phoenix is the market where the instant offer usually reaches the seller before any agent does, and arriving first is most of its power. Here is why the agent inherits an anchor, and how to replace it rather than argue with it.
Read more →What an instant-offer price really is
An instant offer is a real convenience and a specific kind of number. It is the price a company will pay to buy your home, carry it, and resell it, with a service fee and estimated repairs taken out. The headline and the net can diverge. Here is how the number is built.
Read more →What the algorithm can price, and what only the block knows
Two homes with identical records can sell far apart, because price is set by the block, not the parcel. Road noise, a school boundary, a drainage history, the view line, none of it reaches a public record. Here is the inventory of what only the block knows.
Read more →The anchor problem: why the first number shapes the whole deal
The first number attached to a home shapes every number that follows. It used to be the list price. Now it usually arrives before the agent does, from a portal or a chatbot. Here is the mechanism, and how to replace an anchor instead of arguing with it.
Read more →Non-disclosure states and the data gap
In roughly a dozen states, the price a home sold for never becomes public record. The valuation models that run on recorded sales lose their ground truth there, and their errors widen. Here is the mechanism, and who holds the real numbers.
Read more →